Financial Markets Tick by Tick [PDF]

Zitiervorschau

Other Titlesin the Wiley Trading Advantage Series New Market Timing Techniques: Innovative Studies in Market Rhythm and Price Exhaustion Gaming the Market: ApplyingCame Theory to Create Winning Trading Strategies Trading on Expectations: Strategies to Pinpoint Trading Ranges, Trends, and Reversals F ~ d a m e n t aAnalysis l Technical Analysis Managed Trading, Myths and Truths M c ~ i l l a non Options a ~ r e n c eG. illan

Pricing Convertible Bonds

si

ci

JOHN WILEY 8r SONS

Chichester * New York*~ e i ~ h* Brisbane e i ~ * Singapore *Toronto

Copyright0 1999by John Wiley & Sons Ltd, Baffins Lane,Chichester, West SussexP019 lUD, England Copyright 0 Chapter 3 1997 Elsevier Science ~ ~ t i o ~ u 01243 l 779777 ~~ter~u (+44) t i o1243 ~ ~ 779777 ~ e-mail(for orders andcustomer service enquiries):[email protected] Visit our HomePage on http://www.wiley.co.uk or http://www.wiley.com

Reprinted April 1999 All rights reserved. No part of this publication may be reproduced, stored in a retrieval system, or transmitted, in any formor by anymeans, electronic, mechanical,photocopying, recording, scanning or otherwise, except under the terms of the Copyright, Designs and Patents Act 1988or under the termsof a licence issued by theCopyright Licensing Agency, 90 Tottenham CourtRoad, London,UK W1P 9HE,without the permission in writing of John Wiley and Sons Ltd, Baffins Lane,Chichester, West Sussex,UK P019 1UD.

John Wiley & Sons, Inc., 605Third Avenue, New York, NY 10158-0012, USA WILEY-VCH Verlag GmbH, Pappelallee 3, D-69469 Weinheim, Germany Jacaranda Wiley Ltd, 33 Park Road, Milton, Queensland 4064, Australia John Wiley & Sons (Asia) Pte Ltd,2 Clementi Loop #02-01, Jin Xing Distripark, Singapore 129809

John Wiley & Sons (Canada)Ltd, 22 Worcester Road, Rexdale, Ontario M9W 1L1, Canada

Library of Congress CataZo~ng-in-Pu&Zication Data Financial markets tick bytick/edited by Pierre Lequeux. p. cm. Includes index. ISBN 0-471-98160-5 (cloth :alk. paper) 1.Money market-Mathematics. 2. Capital market-~at~ematics. I. Lequeux, Pierre. HG226.F56 1999 332-dc21

98-35143 CIP

tis^ L i b r a ~ C a t a Z o ~ini nPubZication g Data A catalogue record for this book is available from the British Library ISBN 0-471-98160-5 Typeset in 10/12pt Palatino by Laser Words,Madras, India Printed and bound Great in Britain by Bookcraft Ltd, Midsomer Norton, Somerset This book is printed onacid-free paper responsibly manufactured from sustainable forestry, for which atleast two trees are planted for each one usedfor paper p r ~ u c ~ o n .

To Final6 and Carn6lia

This Page Intentionally Left Blank

S

~orew#~d

ix

Ac~nowle~g~ents

xi

C#ntributinga ~ t k o ~ s

xiii

~ntrod~ction

xxi eries, ~ o l a t ian ~i~

1

1.

3

2. Skinn-~ukLin, ~ # ~ n~ i gnkand t Dr.S ~ e p ~ e n € 1Satckell 1~00~

3.

65

Stepken 7. Taylor and Xi~zkongXu 4. 5.

ency F i ~ a ~ c iTime a l ~erie§

mar^ Lundin, Mickel M. Dacorog~aand ~ l r i c A. k M~ller ighs and

Lows: Ti

the in ay

~~rrency

€mman~elAcar and ~ o b e rTt ~ ~ e l art I1 tati is tical Fe~ture§uency of Financial

~eries arket ~ ~ r i afor ~ l ~ §

Owain ap Gwilym, ~

i

~

27

e and Stepken ~ ~Tkomas c ~

l

e

91

127

149 151

viii Contents

7.

arket Tnte~rationin Euro Allison ~ o l l a n d

8.

9.

Em~iricalStudy of TFFE ont tracts Pierre Lequeu~ Mickael Gav~idis,~ a ~ k aN.e ~l a r ~ e l land o s Terence C. Mills

10. ~ ~ ~ a nAcar u eand l Pierre Lequeu~ igh Fre~uencyFinancial Series an ers ~ ~ ~ l i c a t i o n s

Paul MacGregor

190

207 227

255

303 305

uantitative ~ u r r e n ~ y Hai~oH. J.D i j ~ s ~ Marcel ra, A. L, Vernooy and Dr T j a Tjin ~ ~

324

E Futures Positions ~ a r r e n ~ o u l s o n , S a ~ i n e and T o ~Alison l s o n Sinclair

347

376 ndex

389

The working of markets forms the centrepieceof economics. Economists are professionally fascinated by them. Directly or indirectly, markets help to determine our opportunities, incomes and lives. Yetwe h o w far less about their nature and characteristics than ideally we should. Amongstthemostimportantmarketsarethoseforfinancialassets, foreign exchange, equities, bonds, short-dated financial assets (such as bills, commodities) and for derivativesof these underlying spot markets, e.g. forwards, futures and options. In these markets too there are many anomalies, whereby for example, asset prices exhibit more large jumps, and vary more than can be easily explained on the basis of publicly- own 'news' about economic fundamentals. Some of thesemarketsarecontinuous(apartfromweek-endsand holidays), for example the foreign exchange market. Others open and close each day at pre-determined times, e.g. most main equity markets, but are continuous whilst open. For some purposes occasional snap-shots of prices,e.g.attheclose of eachday, of tradingvolumes,sayina two years, may suffice. But if one really wants particular week once every to get inside the skin of these markets, to get a feel for their actual working, one needs continuous tick-by-tick data of price quotes and trades (number and volume),all time-stamped, to provide adequate analysis. It is only recently, with the dramatic development of electronic information technology, that it has become possible to collect, store, manipulate and analyse the millionsof events, of bytes, that capture the continuous operation of key markets. In the second paper of this book 'Modelling Intra-Day Equity Prices and Volatility using Information Arrivals', by Lin, Knight and Satchell, the short recent historyof applied research in high frequency data is dated as starting with the First International Conference on High Frequency Data in Finance, sponsored by Olsen & Associates, March 29-31, 1995. I find that attribution personally pleasing because I played a role in helping to organise that Conference, (at which severalof the papers in this book appeared in their initial form). As an external memberof the Monetary Policy Committee, I have been madeactivelyaware of theperception of parts of thegeneralpublic that economic academics live in a theoretical ivory tower far removed from ordinary life. That separation was once true of the financial arena also, where matters such as trading, portfolio management and risk control

were viewed as best handled by the experienced practitioner, better unsullied by academic theory. Again what greatly pleases me, is how closely practitioners and academics are now working together, throu of the discipline of financial economics, aware of the contributions each c m make to help the other advance. editor, The Pierre Lequeux, in h~self ivesevidence of thebenefits of linkingtheoryandappliedpractice, academic and practical work; and the roll call of authors shows how this subject unites academic disciplines (e.g. economics, maths, finance and physics) and practical experience in the pursuit of underst~ding.This book gives witnessto such combined efforts. But you cannot study markets unless there is an underlying database to study. The move towards the useof electronic systemsin the conduct has made the studyof the high-frequency operationof trading feasible. For example, the paper by ap Gwilym, Buckle and Thomas makes use of detailed LIFFE data, where the driving force behind their provision (as MacGregor describes in his paper on ”The Sources, Preparation and Use of High Frequency Datain the Derivatives Industry”) came from”a particularly technically advanced section of the trading communi^,. .. , made up of hedge funds, unit trusts and pension funds; an ‘alternative i n v e s ~ e n t sindustry’ ”. Fortunately, the combination of increased use of electronic systems in trading together with pressure from the practitioner and academic communities to have access (subject to appropriate confidentiality conditions) to the resulting data series has led to a huge step-forward in the availability of high frequency data sets. That process of expanding dataavaila~ilityis-I am happy to report-ongoing. The resulting data series have a number of particular features, millions of consecutive data points where the events are irregularly spaced in time, a combination of deterministic intra-day seasonal effects together with marked stochastic volatility, effects of differing periodicity, large kurtosis and price jumps, etc, etc. All this leads to the development and use of a weird and wonderful collection of new statistical and mathematical methods and techniques, such as truncated L6vy flights, discrete wavelet transforms etc, etc. Whether or not you are young, and keen enough up on these new techniques, there is much fascinating new information on the workingof these key markets here for allof us. Progress in understanding the working of financial markets is now being made, and quite rapidly. Thisbook provides another milestone in this advance.

High frequency modelling in the financial markets is a vast subjectthat encompass numerous techniques. It would not have been feasible to give the reader a thorough review of the topic without calling upon the knowledge of leading authorities the in field. For making this book possible by c o n t r i b u ~ their g timeand expertiseI wish firstand foremost toaddress my warmest thanks to each of the authors, namely:E m a n u e l Acar, Mike Buckle, Dr Michel Dacorogna, Haijo Dijkstra, Michael Gavridis, Owain ap Gwilym, Allison Holland, Pr. John Knight, Shim-Juh Lin, Dr Mark Lundin,PaulMacGregor,RaphaelMarkellos,PhilippeMichelotti,Pr. Terence Mills, Ulrich Muller, Dr. Stephen Satchell, Dr. Thomas Schneeweis, Alison Sinclair, Dr. Richard Spurgin, Pr. Stephen Taylor, Pr. Steve Thomas, Dr Tjark Tjin, Robert Toffel, Dr.Darren Toulson, Sabine Toulson, Marcel Vernooy and Xinzhong(Gary) Xu. Ialso want tothankPr.Charles Goodhart for accepting to write the preface to "Financial Markets Tick by Tick" at such short notice despite his busy agenda. Thanks goes also to Robert Amzallag, whom whilst General Manager ofBNP London, shared with me his keen interest for financial modelling and gave me the opportunity to develop my career along this path. Finally, many thanks to the LIFFE who gave mesupport and allowed this book to become reality by sponsoring it. Pierre Lequeux

This Page Intentionally Left Blank

Emmanuel Acar is a proprietary trader at Dresdner Kleinwort Benson. His current duties include elaborating and managing futures and foreign exchange portfolios for the bank account. He has experiencein qu~titative strategies,as an actuaryandhavingdonehisPhDonthestochastic properties of trading rules.He has been trading and researching financial markets for the past eight yearsBZW at and Banque Nationale de Paris in London. ike Buckle is a Lecturerin Finance at the European Business Management Schoolat the Universityof Wales, Swansea, UK. His PhD involved the developmentof financial forecasting models. His main research interests lie in the areaof market microstructure with reference to derivative markets. He has published several papers on aspects of derivative markets and is co-author of “The UK Financial System”, (MUP, 1995) and ”The OfficialTrainingManualfortheInvestmentManagementCertificate’’ (Institute of Investment Management and Research, 1997), acoro~naafter completing his undergraduate and graduate studiesin Physics at the University of Geneva, accepted a postdoctoral position at the Universityof California at Berkeley. At both universities, he concentrated on solid-state physics, assumed extensive teaching duties and assisted in the evaluation of various computer systems. His main research interest is the application of computer science and numerical analysis to dynamic systems in various fields in order to gain insight into the behaviour of such systems. In 1986, Dr. Dacorogna joined Olsen & Associates (O&A), then a fledgling research institute, to become of one its founding members. In April 1997, he became head of the Research and Development department at 0&A. He has devoted the past few yearsto an extensive research and development project involving a real-time, valueadded information system in the fieldof applied economics. In addition to his research duties at O&A, Dr. Dacorogna has assumed a leadership role in organizing the first international conference on high frequency data in finance. Throughout his career, Dr. Dacorogna has gained considerable experience working with sophisticated computer environments, ranging from Grays toS W workstations. Dr. Dacorogna has remained close to his academic background, continuing to offer internal and external seminars

xiv ~ o n t r i ~ ~A ~u i~n~go r s

andpublishingtheresults of hisresearchinawiderange of scientific journals and internal documents. Dr. Dacorogna’s mother tongueis French; he is fluent in English and also speaks German. Married with two daughters, Dr. Dacorogna favours spending his leisure time withhis family. ijkstra, with 14 years of banking experience, has worked in Corporate and Investment Banking, and was active as a money-market and currency options trader for five years. For the last three years Haijo Dijkstra was Headof Funding and Structures, a successful team at Rabobank International involved in funding activities on behalf of the Rabobank organisation, investment-activities and tax-driven capital market transactions. Recently, Haijo Dijkstra was involved in a currency overlay project at Rabobank International. is joinedBNP’sGlobal M etsResearchTeamfrom Bank in 1996 as aSeniorantitativeAnalyst.He is responsible for the evaluation and application of new modelling techniques to the forecastingof financial timeseries ranging from traditional fundamental economic models to more advancedtime series econometric modelling. He holds aPh.D. in Financial Economics from Brunel University and an M.Sc, in Project Analysis, Finance tk Investment from the University of York.

is a Lecturerin Finance at the Departmentof Management at the Universityof Southampton, UK. Prior to taking this position he was a Research Assistant at the University of Wales at Swansea working on a research project on intra-day empirical regularities in LIFFE f u ~ r e s and options. He has a PhD in Finance from Swansea which studied the index options market at LIFFE, and has published several papers in international journals including the Journal of Futures Markets, Journal of Derivatives and Journalof Fixed Income. o l l a works ~ ~ as policy advisor on secondary market issues at the newly formed UK Debt Management Office. Prior to this she spent a number of years at the Bank of England, which she joined in 1993. Her first two years at the Bank were spent on macro-economic analysis and forecasting of theandtheother G7 economies.Thiswasfollowed by a period of empirical researchinto market microstruc~reissues with the Bank’s Market and Trading Systems Division. She holds a MSc in EconometricsfromtheLondonSchool of Economicsanda BA (Joint Hons) in Economics and Mathematics from University College Cork. She has published papers in the CEPR’s Discussion Paper series, the Bank’s

C o n t r i ~ u ~Authors in~ xv

Working Paper series and in the Economic Journal. She also has a number of articles in the Bank‘s uarterly Bulletin and Financial Stability Review. t is a Professor in the Department of Economics at the University of WesternOntario.Hisrecentresearchinterestsare in financial econome~ics.He has an ongoing interest in theoretical econometrics,an area in whichhe has published extensively. eux is Assistant Vice President at Banque Nationale de Paris ich he joined in 1987. Pierre is a graduate in international trade and holds a diploma of the Forex Association. He joined the l3NP ~uantitativeResearch and Trading desk in 1991 as a dealer after gaining experience on the treasury and corporate desk.He is primarily active in the research and development of trading models and portfolio management techniques. His approachis principally based on statistical models developed by him. Pierre is a member of the Forex association and the Alternative Investment Management Association. He is a frequent contributor to academic investment conferences and publications and is a member of the editorial boardof ”Derivatives Uses Trading Cllr: Regulation” as well as the editor of the AIMA newsletter’s currency section. He is chairman of theAIMAbenchmarkcommitteewhichhasrecentlyproducedan acclaimed “Reviewof ~ethodologyand Utilisationof Alternative Investment Benchmarks” and also a memberof the AIMA Currency Advisory Group. After heading the ~uantitative Research and Trading desk of BNP London where he developed a new active management curre mark (FXDX), he is now focusing his attention on develop Exchange Business for BNP.

Lin completeda 1Ph.D. inEconomicsattheUniversity of Western Ontario. He now teaches at the University of Technology, Sydney. The main theme in his dissertation is modelling and examining timeseries properties of information flow in financial asset markets. Heis also interested in doing research on financial market structures and the impact of information dissemination on the distribution of financial asset returns. in completed a doctoral degree in High Energy Particle Physics with Universite Louis Pasteur, Strasbourg, France, in 1995. He also holds a Bachelor of Science degree in mathematics and computer science from the Universityof Illinois. His thesis research was performed both at Fermi National Accelerator Laboratoryin the US and the European Centre for Particle Physics Research (CERN) in Switzerland. Soon after finishing his doctoral degree, hejoined Qlsen and Associates in Zurich as a Research Scientist.Theemphasis of his workinthefield of financehasbeen

xvi Con~ri~uting Aut~ors

in the areas of dynamic currency hedging, portfolio management and the fundamental processes inherent to multivariate analysis of financial time series. Dr. Lundin has acted as an invited speaker for the Association for Investment Management and Research (AIMR), participating in their continuing education series on managing currency risk as well as participatin~in Olsen and Associates research workshops on high frequency data and model building.He has recently joined the Research and Strategy Group of Fimagen Asset Management in Belgium, a member of the Generale Bank Group.

regor joined LIFFE Market Data Services in 1994 as Statistical Manager, and extendedhis responsibilities to Marketing Manager for all aspects of LIFFE Market Data in 1997. He has been instrumental in the launch of LIFFE Tick Data (1995), all aspects of market data on the LIFFEnet website (1995), and more recently the launch of LIFFE’s new historical data product range “LIFFEdata” in 1998. Duringhis time at LIFFEhe has worked closely with the alternative investments industry ~orldwideto develop and continuously improve LIFFE’s market data products. Paul isagraduateinEconomics,andholdsadiplomainMarketingfrom the London School, of .Economics. He has previously worked at British Petroleum and the British Plastics Federation.

~rkellosis reading towards his Ph.D. at the Department of Economics, Loughborough University, UK where he holds a Junior Fellowship from the Royal Economic Society. For the past three years he has consulted in industry on quantitative financial analysis and econometrics. His research interests and publications are mostly concerned w non-linear models and trading systems.

i c ~ e ~ o moved tti to the UK after graduatingin business studies in 1989. He then focusedhis interest on the energy market, primarily the electricity futures pricing. In 1994he joined Phibro of Salomon Brothers, as an energy trader to contribute to the expansion of the electricity and naturalgastradingdesk. In 1996,PhilippeMichelottidecidedto join Creditanstalt Global Futures Investment Management Ltd. (now BankAustria/Creditanstalt Futures Investment Ltd.) to expand his horizon in the financial markets and derivatives funds management. As part of a small team his responsibilities have ranged from risk management analysis to structured products research. He has been instrumental in ~romotingand increasing the reputation of the BA/GA Futures Investment in Europe, the Middle East and Asia. Today, Philippe Michelotti, as Head of the Business Development, concentrates his efforts in advising his customers on portfolio diversification. He promotes BA/CA Futures Investment’s

C~ntribufin~ Authors xvii

expertise by offering the bank‘s futures fundsbyor proposing tailor made products at the client’s request. ills is Professorof Economics at Loughborough~niversity, usly held professorial appointments at the Universityof and City University Business School and has worked for the Mon~tary PolicyGroup at theBank of England.Heisauthor of ”Time ~ e r i e s Techniques for Economists” and ”The Econometric Modellingof Financial Time Series”, both published by Cambridge University Press, and over 100 articles in journals and books. His research interests are in the of area time series econometrics, with particular interests in finance, macroeconomics and forecasting. Ulric er studiedphysicsandgraduated at theSwissFederalInstitu of Te gy,Zurich.Hisprize-winning PhD. thesisledtoapatente thermoacousticheat-pump.After his studies,heworkedat an e neeringcompanyandafterwards as aself-employedconsultant. workencompassedseveralfieldssuchassemiconductoroptics, dynamics and industrial risk analysis.In August 1985, he was one of the founding membersof Olsen and Associates, Research Institute for Appli Economics, in Zurich. His pioneer work helped to create the 01s unique and Associates high-frequency database and the O&A financial i n f o r ~ a tion system for banks and other financial institutions.his With colleagues, he conducted some extensive fundamental research of financial data from the foreign exchange market and other financial markets. He is (co-)author of most of O&A’sscientific publications.An important pieceof research is the stochastic process HARCH which reproduces the behaviour of empirical market prices by explaining their volatility as generated by traders with different time horizons. Ulrich Muller also developed forecasting and trading models. Ulrich Muller is supported by the younger researchersof O&A’s research group but likes to be personally involved in the design and programming of research and development projects. He has given invited talks at several conferences and seminars in academic institutio~s. Ulrich Miiller likes books and music of different kinds, traditional Chinese painting, bicycling and developing some sortsof computer graphics. sings in a choir and spends a large part of his free time with his two children.

~ t e ~ ~h~ tec ~ h eisl la lecturer in Economics at the University of Cambridge, and a fellow of Trinity College. He has a keen empirical and theoretical interest in most areasof finance andis particularly intriguedby all issues of asset management, risk management and measurement.He advises a number of city companies and has published extensively in both academic and practitioners’ outlets.

xviii Con~ributingAut~ors

is Professor of Finance at the School of ManagementattheUniversity of Massachuse~sinAmherst,Massachusetts andDirector of theCentreforInternationalSecurityandDerivative Markets (CISDM) at the School of Management. He obtained his PhD. in 1977 fromtheUniversity of Iowa. He isco-author of ”Financial Futures: F~damentals,Strategies,andApplications”(RichardIrwin) andtheauthor of ”Benefits of ManagedFutures,”published by the lternative Investment Management Association (AIMA). He is on the oard of Directors of the Managed Funds Association and editor of the Journal of Alternative Investments. He has published over 50 articles in academic finance and management journals, such as the Journal of Futures Markets, Journal of Finance and ~uantitativeAnalysis, Journal of Portfolio Management, Journal of Finance, Journalof Futures Markets, Journal of Derivatives, Derivatives Quarterly, and Financial Analysts Journal. He has also published widely in financial practitioner magazines such as the AIMA and Barclay Newsletters. He has been a Ful~rightResearch Fellow in France, taught at ESSEC in France, and is Visiting Professor of Corporate Finance at Institute of Economic Research, Lund University, Sweden. inclair is a consultant at Intelligent Financial Systems Limited, specializing in database management and forecasting. With a 13.S~.and M.Phi1. in Economics, she previously worked as a researcher and lecturer in the UK and Germany. in is Assistant Professor of Finance at the Graduate School of Management at ClarkUniversityandAssociateDirector of the Centre for International Security and Derivative Markets (CISDM) atthe university of Massachusetts. He holds a Bachelor’sdegreein mathematics from Dartmouth College and received a PhD. in Finance from the Universityof Massachusetts in 1995. He has published research in academic journals such as the Journal of Derivatives and the Journal of Futures Markets as well as in practitioner journals as such the Derivatives Quarterly. He has contributed toa number of edited books in the areas of high frequency data and alternative investment strategiesa and member is of the editorial board of the Journal of Alternative Investments. Before joining the faculty at Clark University, Dr. Spurgin was Director of Fixed Income Research for Thomson Financial in Boston.

en Tayloris a Professor of Finance at Lancaster University, England From 1995 to 1998, he washead of theDepartment of ~ c c o ~ t i n g andFinance, a departmentratedinthemostprestigiouscategoryin the 1996 UK research ratings exercise. His numerous pub~icationsover 20 years of research include “Modelling Financial Time Series” (Wiley),

Contri~u~ing A ~ t ~ o rxix s

in which he presented the first description of stochastic volatility models and pioneering a analysis of GARC continues e actively to research a wide-range of volatility issues.He has taughthis own advanced financial econometrics coursein England, Aus~ia, ~elgium, Hong Kong andAustralia.ProfessorTaylorobtained his M.A.andPh.D.degrees from Lancaster U~versity,following his ree in Mathemati~sat Ca~bridge University. S is Professor of Financial Markets at the D e p a r ~ e n of t M ~ a g e m e n tattheUniversity of Southampto Professor at the ISMA Centre, Universityof Readi extensivelyin inte~ational journals,includingth 1 of International Money arid Finance, Journal of Bankin and Finance, Journal of Futures Markets, Journal of Derivatives, Journal me and the Economic Journal, and is co-author of the "Investm ment Certificate Official Training Manual" for the IIMR. He is also consultant editor of FT Credit Ratings ~ n t e ~ a t i o nand a l a quantitative finance consultant with Charterhouse Tilney Securities.

is recognized as a successful researcher with a Ph.D. in Theoretical Particle Physics. Hehas seven years of professional working experience at severalfirstrateinstitutionssuchastheUniversity of M~chen and Shell Research. TjarkTjin has been involved in a currency overlay project from September 1996-September 1997 at Rabobank International.HenowworksattheTrading CO pany,one of thelargest stock-option brokersin ~msterdamas Directorof Research. has just completed his ~ndergraduateDegree in Mathematics at Imperial College, London where he o b t a ~ e da First Class. He has previously been a trader at Titan Capital Management implementing proprietary trading models in the FX markets. His current focus is one of system developmentof numerous tradin~strategies and money management techniques for application across a wide range of markets. o~lson is a directorof Intellige~tFinancial Systems Limited, a company developing financial forecasting and trading systems. He holds a B.Sc. in ath he ma tics and Physics and a Ph.D. in Neural Networks and Time Series Analysis from King's College London. He is the author of a number of papers on neural networks applied to time series analysis, h a g e processing and financial forecasting. abine To~lsonis a director of Intelligent Financial Systems Limited. She read Economics and Mathematicsat University College London and holds an M.Sc. with Distinction in Neural Networks from King's College

xx ~ o n t r i b u t i Authors n~

London. Shehas published several papers on exchange rate analysis us neural nets and portfolio management.

ooy is an experienced project manager who spent a large p of his professional career, five and a half years, with the RobecoCroup. e started as a quantitative researcher at INS, the ~uantitati~e research lns~tuteof Robeco, and was primarily responsible for the development of forecasting models for equity markets, bond markets and asset-mix portfolios. At the same time, he acted as a consultant on these matters towards the investment departments and the foreign offices of the Robeco Group. Marcel Vernooy has been involved in a currency overlay project at Rabobank International.

U is employed by the Bank of England. Until 1998, he was Senior LecturerinAccountingandFinance at theUniversity of Manchester, England, wherehe taught investment analysis and capital market theory Hisresearchinterestsareinvolatilitymodellingandforecasting,the efficiency of financial futures and options markets, and empirical tests of asset pricing models. 'Three innovative papers were published from his l3h.D. thesis on exchange rate volatility, including important contributi aboutthetermstructureandsmileproperties of impliedvolatilities. He obtainedhis BSc. fromPekingUniversity, his M.B.AfromAston ~niversityand his PhD. from Lancaster University.

Financial markets are being swept by important changes affecting the waymarketparticipantsinter-actandoperate. As arecentexample, bothoverthecounterandexchangebasedmarketpractitionershave considered and appliedin some sort of way computerized trading asan alternative to open out-cry. These developments have contributed toward a more efficient market and potentially a better service to the end of user financial products. The phenomenal progressin information technology made over the last decade has been a catalyst to these changes. Maybe a less apparent but not the least important contributor to these changes isthewideruse of highfrequencytimepriceseriesinthedecisionmaking process. The advent of second generation microprocessors has made it possible to process large amount of data within a realistic time frametomakepracticaluse of it.Undoubtedlytheanalysis of high frequency data brings us a wealth of information about the behavior of financial prices and new perspectives in the field of risk management and forecasting which were previously inconceivable.It permits market practitioners to test hypotheses and new trading strategies. It provides new resources to model and generate correlation and volatility estimates to input into pricing and risk models. Acknowledging the demand, statistical departments of financial exchanges and other data suppliers have started to release ”clean” high frequency price data in convenient format. This has translated into a steady flow of research papers on high frequency modelling producedby both academics and market practitioners. This book intends to give the reader a broad view on the uses and avenues of research presently investigated. It regroups researches from leading academics and market practitioners in the field of high frequency data. It is structured around three sections addressing practical issues that are paramount to the financial community. The first section of the book is dedicatedtopricevolatilityandriskestimators,thesecond section concentrates on statistical features and forecasting issues. Finally, the last section illustrates how ”tick data” affect the way that market practitioners operatein the financial markets by giving practical examples of applications.

Estimating risk is probably one of the most arduous and discussed issues in finance. It affects a wide area of financial activities from pricing options to managing portfolios of assets or evaluating the day to day risk of a dealing room. The interest for new measures of risk has been epitomized by the recent arrivalof risk mana ment techniques involving co-variance matrices of financial ins efficiency of these models being a function of therobustneseestimatesused.Keepingthisin &d, thefirstpart of ”FinanciaetsTick by Tick”givesareview of how high frequency data contributes to a better ~ d e r s t a n d i and n ~ modelling

Spurgin and Thomas ntra-day volatilityby introducing new historic volatility estimators relying on the trading These estimators are shown to provide a more accurate forecast than the traditional closing price and Parkinson estimators. Shin-Juh Lin, Professor John Knight and Dr. Stephen Satchel1 investigate the choiceof proxies to measure the flow of informa

both the number of trades and the n u b e r of price changes are much better explanatory variables than trading volume itself. They find that their model reduces the persistence in volatility and consequently leads

..

quotationsandinimpliedvolatility.Theyfindexchangeratequotation information to be ore informative than options information. Their conclusion is supported by an out of sample comparison of forecast of

ime ~ e r i e by ~ ”Mark Lundin, ses the problems associated y. They use a co-volatility or non existent data. The stability of correlation over me and the exponential memory of financial time series return corre tions is investigated. They find an inverse relationship between the rate of correlation a ~ e ~ u a t i oand n the level of activity in the instruments involved and also that returns short term are

likely tobe uncorrelated evenif highly correlated over the long term. They do not find a "best" time interval for measuring correlation since this is dependent on the pertinence of time horizon for specific applications. Emmanuel Acar an stylized facts of extreme clustering for the GME currency contract. They show that contrarily to hetereoscedacity, skew and kurtosis, the impact of drift has little effect on the timing of intra-day extremes. These evidencesof clustering could be of great help to a liquidity or directional trader.

Non-linear methods have provided financial market practitioners with new forecasting tools. High frequency data and faster computers have undoubtedly been the main factors contributing towards the development and useof these methodologies. The second section of "Financial Markets Tick by Tick"acknowledges thisby providing the reader with insights on the statistical featuresof high frequency financial time price series. Owain ap Gwilym, Mike Buckle and Stephen Thomas investigate the properties of time series and interaction of market variablesin "The i e h a v ~ oof ~ rkey ~eriv~tivesff. use They a large high frequency data-set to examine the intra-day behaviour of return, volatility, trading volume, bid-ask spread and price reversal across a range of financial contracts traded on the LIFFE. They compare their findings with previous researches on other markets and find that market structure plays a vital role in determining the behaviour of these variables and this particular1 government bonds and their associated future contracts. She finds that price discovery occurs in the future markets with the spot market following with a lag. Wlhereas arbitrage activity appears tobe limited by the presence of market friction, the importance of spread trading is highlighted in the caseof dually traded futures contracts. These are highly integrated g very active arbitrage between the markets. frequency relationship between the LIFFE Gilt and FTSE contracts, Pierre Lequeuxuses10years of 15-minutedataforbothGiltsandFTSEto

investigate how the intra-day relationship between the two instruments evolved over time. igh-fre~uencyrandomwalks?” MichaelGavridis,Raphael S and Terence Mills discuss the departures of high frequency pricesfromthesimplerandomwalkmodelandtheimplicationsfor short term risk management and trading. Their arguments are supporte by the analysis of 30-minute prices for 13 currency pairs. Their results suggest that although fat tails characterize the distributions of returns, variances are finite, This impliesthat short term investors face finite but non-Gaussian risks. They also find intra-day seasonalities in systematic risks and long-run dynamics. nuelAcarandPierreLequeuxin ”Tradingrules rlying time series prope~ies”provide a better un of forecasting strategiesby using stochastic modelling. They derive tests of random walk and market efficiency from the stochastic properties of trading rules returns. Their propositions are then applied to a data-set of foreign exchange rates.

IGH FRE~UENCY FINANCIAL ~RACTI~IO~ER§ ~PPLICATION§

uilding on the previous sections, examples of practical applications and arguments for the use of high frequency data are presented in the four l chapters of this book. Paul MacGreor in ”The sources, preparation and ouse atives ma~kets~’ gives us a concise de aregeneratedforbothoverthecounter(OTC)and financial exchanges. Collectionof data for both open outcry and computerizedtradingaredetailed.Theinnovativeapproach of theLIFFE in supplying clean historical data whilst meeting end users’ requirements is emo on st rated in a ”guided tour”of their LIFFEstyk product. design of a ~uantitativecurrency overlay program” by ,Marcel Vernooy and Tjark Tjin raises practicalinissues desi acurrencyoverlayprogram.Theyfirstprovidethereaderwiththe essentialanalyticstodesignacurrencyoverlayprogram.Thenthey illustrate their methodology withan example using high frequency data models as a core indicator for the decision making in the currency overlay. ctingamanagedportfolio of high fre~uencyL~FFE ns” Darren Toulson, Sabine Toulson and Alison Sinclair demonstrate how discrete wavelet transform, neural networks and high

ln~ro~uctionMCV

frequency data canbe used to design a risk-managed portfolioof LIFFE futures contracts. After detailing all the practical issues they assess the performance of the portfolioby testing their strategy out of sample. Pierre Lequeux and Philippe Michelotti, in ”Is short tern better? An insight through managed futures ~erfornances”,quantify the economic value of short term traders and consequently the use of high frequency data. They analyse the performance of a universe of trading managers operating over three distinct time horizons: Long term, medium term and short term. They find that, generally, short term traders are less volatile and offer a better risk adjusted return than traders operating on longer time frames. They also find that the minimization of risk is better obtained through a portfolioof short term traders than long term traders. The topic of high frequency data in the financial markets is very broad and the implications for market practitioners are numerous. We hope that of the 14 chapters of this book will contribute toward a finer knowledge this very specialized field as well as giving some orientation in termsof future research. Pierre Lequeux

This Page Intentionally Left Blank

This Page Intentionally Left Blank

.~ ~ u r and ~ i Thomas n ~chnee~eis ~ s s i s ~ ~rofessor an~ of €i~ancef ~ l a ~niversi~yf r ~ ~rofessor of €inancef ~ n i v e r sof i~ a s s a ~ ~ ~ s e ~ ~ s

This chapter introduces new methods of estimating the historic volatility of asecurityfrom its trading range.' Parkinson (1980) showedthat the range of a security contains considerably more information2 about the re~rn-generatingprocess than does the period-to-period return. A number of papers have been published on this topic, all of which have two things in common. First, the authors assume security prices follow geometric brownian motion (GBM). Secondly, eachof the several existing range-based variance estimators is based on squared trading ranges. Estimators derived from the second sample momentof observed ranges are highlyefficient.However,theyareshowninthischapterto be more sensitive to misspecification of the underlying process than estimators derived from lower sample moments. A new classof variance estimators is proposed in this chapter. Some members of this class are shown to provide more accurate variance estimates than existing range estimators and the close-to-close estimator as well. There are a numberof potential applications of this research. First,by correcting the biases known to exist in range-based variance estimators3 and improving their efficiency,it may be possible to accurately estimate historic volatility over short time frames. Improved estimation would have a direct application to option and other derivative pricing. More accurate variance estimation would also allow for more efficient estimation of a security's beta, more accuracy in event studies, lead to more accurate ~ i ~ a ~ ~c i a l r Tick ~ by~Tick s Edited by Pierre Lequeux. 0 1999John Wiley & Sons Ltd

4 Ric~ardB. Spurgin and Thomas Schneeweis

models of the the-varying properties of return volatility, and generally improve any statistic that relies on an estimate of the variance or covariance of a security. eview of Previous Research Trading Range Studies

The distributionof the trading rangeof a security that follows geometric Brownianmotionhasbeenextensivelystudied.Parkinson(1980)first considered this problem. Using a distribution first derived by Feller (1951), Parkinson found a variance estimator for a security whose log follows azero-meandiffusion.Hisestimatorisaboutfivetimesasefficient4 as theconventionalclose-to-closeestimator.GarmanandKlass (1980) extend Parkinson’s approach, incorporating the open and close prices and the trading hours of the security. Ball and Torous (1984) find the ~ n h u m - v a r i a n c range-based e estimatorby solving for the MLEof the joint distribution of high, low, and closing prices. Kunitomo (1992) and Rogers and Satchell (1991) develop estimators that allow for drift a term in the Parkinson and Garman and Klass estimators, respectively. Despite theoretical results suggesting range-based estimators are sev times more efficient than classical ones, empirical tests to demonstrate fail their superiority. Garman and mass (1980) use simulation to show that range-basedestimatorsaresensitivetodiscretenessinpricechanges, producing downward biased estimates. Beckers (1983) reaches a similar conclusionwithactualdata,andshowstheefficiency of range-based estimates is only slightly better than the classical approach. Wiggins (1 1992) studies the properties of several estimators for a wide range of securitiesandfindstheperformance of range-basedestimators is not significantly better and often worse than traditional estimators. Rogers, Satchell and Yoon (1994) find that the Rogers and Satchell (1991) estimator is accurate when tested with simulated data, but is considerably biased when actual stock data are used. The problems may be due in part to microstructure issues. Because prices are reported in discrete increments, the true high and low of a security that followsGBM are unobservable. The rounded values distort both the trading range and the price, lastbut the influence on the range w be more pronounced. Marsh and Rosenfeld (1986) employ a model that shows a discretely observed range be will smaller than the true range, and thus estimators based on the observed range will give downward biased estimates of the true variance. Rogers and Satchell (1991) derive a model in a continuous-time framework that corrects this bias when the time interval between price changes is known. Ball (1988) describes the bias when estimating variance with discrete end-of-period prices, and shows

E~cientEstimation o ~ ~ n t r aVolatili~y -~a~ 5

that variance estimates using observed returns will be downward biased. Gho and Frees (1988) reach the same conclusion using a first-passage time approach. Another possible explanation for the poor performance of range-based estimators is misspecification of the underlying return-generating process. Range-based variance estimators are more sensitive to misspecification of the underlying process than classical estimators (Heynen and Kat, 1993). If the assumptionsof geometric Brownian motion are violated, estimators derived from the trading rangebewill adversely affected to a larger degree than close-to-close estimators.

ISTORIC VOLATILITYUSING THE T

R A ~ IRANGE ~G

Methods of estimating historic volatility using the trading range of a security are described in this section. The approach makes use of the distribution of the range of a binomial random walk (Weiss and Rubin, 1983), the rangeof a diffusion process (Feller, 1951), and the distribution of thenumber of distinctsitesvisited by abinomialrandomwalk. Prior research has focused on quadratic estimation techniques. However, This section estimators caneasily be derived from other sample moments. also identifies the reasons why previous tests of range-based variance estimators have been biased. i s t r i ~ ~ t i oofnthe Rangeof a ~ i n o m i a ~ The distribution of the range of a binomial random walk was derived by Weiss and Rubin (1983). Although this function has been known in the scientific literature for some time, it has not been explored in finance literature. The probabilityW that the rangeis exactly l is: 1

+

h is the difference operator, h a f ( a ) =f ( a 1) - f ( a ) . Q n is the probability that the walkis contained withinan interval (-Q, b) given that the starting point is 0 and there aren steps in the walk,

6 ~ichardB. Spur~inand ~ h o ~Sc~lneeweis as

Unis the probability that the walk is at priceY at stepn, given it is restricted to (-a, b), and began at0, 00

U&; -a, b(0)=

+ + 22(a + 6))).

(Pn@+ 22(a + 6)) -Pn(r 2a

3

l=-00

P, is the unrestricted probability that the walk is at location Y at step n. If the process follows a zero-mean binomial expansion5, this probability i transformed Binomial distribution, 4

The transformation maps the range of the random walk [--n,n] to the binomialrange [0, n]. Thereareothermethods of derivingtherange distribution of a binomial variable. A function that generates range probabilities for an n-step walk given then -1 step probabilities is described in the Appendix. istribution of the Rangeof a ~ontinuousRandom

The distributionof the rangeof a security following a zero-mean diffusion process was first solved by Feller (1951). This function and the moment generating function are:

n=l

6

where the2's are observed ranges, N is the sample size, erfc(z) is one minus the integral of the Gaussian distribution, and ((x) is the Remann zeta function: The distribution has one unobservable parameter, the diffusion constant a2. ution of the Numberof Sites Visitedby a

Therange of acontinuousprocesshasbeenstudied by anumber of authors,buttheseresultstranslatepoorlytodiscreteprocessesfora

Eficient Esti~ationof Intra-day ~ o ~ a t i l ~7f y

simple reason: Feller defined the range differently than Parkinson. Feller in defines the range as the numberof sites visitedby a random walk, while finance the range is the difference between the high and low trades. Thus the continuous range density5 is the limiting distributionof the number of the range of sites visitedby a random walk, not the limiting distribution of a random walk. Although these two quantities are equivalent in the limit, in discrete time the numberof sites visited (in finance, the number of distinct prices traded during the period) is generally one trade-size larger than the range.6 This is the principal sourceof discreteness bias in Parkinson’s estimator. The distribution of the number of sites visited by a binomial random of the P distribution employed walk (V), is created with a transformation by Parkinson. Assuming that the number of sites visited is always one in this model, then the probability that D(2 1) = greater7 than the range V(2) = P(1).

+

oments of the Range ~istributions

Figure 1 graphs the moments of the three functions: The density employed by Parkinson (P), the discrete range density from Weiss and Rubin (D) and the PDF of the number of sites visited by a discrete random walk 12 c

0 C

10

l?

W-

0

0

0

1 1.5 2 2.5 3 3.5 Distribution Moment, Discrete Walk Length= 20, Unit Variance

0.5

4

Figure 1. Comparison of Moment Functions: Parkinson, Discrete Range, and Sites Visited.

. Varianceisnormalized,andthediscretewalksfurtherassumea lk-length of 20. According to the uni~ueness theorem: if the V function and the P function have a common moment-generating function, then the dis~ibutionsare the same. M i l e there is no closed solution to the of V n t - g e n e r a ~ gfunction of V, Figure 1 shows that the moments are nearly identical over the first4 moments (including fractional moments) and that the moments of are significantly different9 WhileFigure1demonstratesthesimilaritybetween V and P is an approximation of V, are not the same function. Since givenmoment, V convergestoa sli tlydifferentnumberthanthat predicted by P. This difference is small and easily correctedin any case. A numerical correction of this difference, termed the difference constant (Cm), is described in Appendix 2.

The varianceof a diffusion process is given by a2t.It is generally assumed that time is observable, and hence the unknown parameter is a2.Assuming a w i t t h einterval, Parkinson (1980) derived an estimator of the Feller distribution from the second moment of 6,

other estimator proposedby Par son is based on the first moment of the distributionlo: 8

The variance of a zero-mean binomial random walk is S2n, where n is number of steps taken and S is the step-size or m ~ trade. mThe ~ step-size is assumed to be observable and the true source of uncertain^ is assumed tobe the numberof price changes per period. Other researchers studyingthisquestionhavemadetheopposite a s s ~ p t i o n - t h a t the number of steps per period is known and the variance per step is unknow wi arm an and Klass (1980), Torney (1986), Wiggins (1991), Rogers and Satchel1 (1991)). Assuming an unknown step-size is reasonable if prices are assumed to follow a continuous process. However, a s s u ~ the g walklength is unknown seems more consistent with observed market beha During periods of increased volatility, it is generally the transaction rate that increases, not the size of each pricechange."

E ~ c i e n t E ~ t i ~ aoft i~onnt r a - ~ aVoZatiZi~y y 9

The V function and the moment correction discussed in the previous section lead to a natural estimatorof the variance. For example, the first moment of the V function canbe expressed as an adjusted functionof the first moment of the P distribution (which is calculated from Equation6) and the difference constant (Appendix 2):

(

+

E[r,] = E(S2n) n ,4995)

1/2

.

9

This relationship is rearranged to form a variance estimator based on theaverage of observedranges.Thisestimatorrequirestransforming observed ranges into number of sites visited, averaging observed ranges, and reducing the totalby about .499 times the tick-size of the security:

,where vi = li

+S

10

An estimator of this type can be created from any of the moments of V and P. In general,if vi = Ei S and C , is the difference constant, then

+

11

ATA

The estimators are tested with daily and intra-day data on two financial futures contracts, the GME SP500 contract and the CBT Treasury Bond contract, The tick-by-tick data, which were supplied by Tick Data, Inc., cover the period October, 1989 through June, 1994. Each trading day is brokenintofoursegments.The day session of T-Bondtradingis400 ~ u t e long, s so each ~ t r a - d a yperiod is 100 minutes. Night trading for GBT Bonds was excluded from the analysis, as was T-Bond futures trading on foreign exchanges (Tokyo and LIFFE). TheS&P500 trading session is 405 minutes long.In this analysis, the first three periods are 100 minutes, and the final period is 105 minutes (from 2:30-4:15 Eastern Time). In each intraday period the range is calculated as the difference between the high and low recorded trades. Returns are based on the nearby contract, rolling to the next contracton the first day of the delivery month. The daily range is calculated as the natural logarithmof the high trade less the log of the low trade. The return is the log of the closing quote less

N

U

N

zs

10 R i c ~B.~Spurgin a and Thomas Schneeweis

N

os"

N

d cn

l

Ejicient Esti~ationof Intra-day Vo~atility 11

the log of the open.No attempt is made to account for overnight activity, asdataontheovernightrangearenotreliable. As aresult,variance estimates will underestimate the true variance per 24-hour day. Table l reportsdescriptivestatistics of thedata.Thefirstpanel is S&P500data.Resultsarereportedcross-sectionally by timeperiodas well as in total. There is considerable differencein variance between the time periods. The first and last periods account for 70 percent of the daytime variance and the middle two periods only 30 percent. Returns are negatively skewed, extremelyso in the fourth period, where skewness is -4.54. Returns are also fat-tailed. The coefficients of kurtosis range from a to 75.41 in the fourth. low of 7.09 in the third period The extreme readings in the fourth period are entirely due to the events of one day-the "mini-crash" on October 13,1989. On that day the stock market fell nearly 9 percent in the final hourof trading. The secondpanel of Table 1B repeats the analysis in Table 1A excluding October13 and 16, 1989. After excluding those two days, return volatility in the fourth period declines by 25 percentto .l83 and overall intra-dayvolatility declines by 15 percent to .131. Skewness and kurtosis measures also decline sharply. Declines of similar magnitude are recorded in the volatilityof the intraday trading ranges. However, the mean intra-day range declines by only .003, or about $ percent. Thus, estimators based on the average trading range will be only marginally influencedby the outlier, while estimatorsbased on higher moments will be highly influenced. Table 1C reports descriptive statistics of Treasury Bond futures data. As with thestock data, the mean returnis negligible. Returns are slightly negatively skewed and fat-tailed, though as not severely as the stock data. Consistent with prior research on the empirical distribution of returns, the higher momentsof the daily return and range distribution are considerably smaller than for intra-day data1*.Skew and kurtosis for daily bond returns, for example, are -.21 and 3.77, not far from normality.

EFFICIENCY OF VARIANCE FORECASTS

h this section, the reliability of variance forecasts produced by each model is tested. Variance estimates are studied two ways. First, simulated data are used to understand the propertiesof the estimators, in particular the bias and efficiency of each estimator. The second test involves historical security data. Data are first segmented into non-overlapping blocks. A variance estimate is derived from the observations in each block. This estimate is used as a forecastof the nextperiod's variance. If an estimator

12 ~ i c h ~B. r dSpurgin and ~ h o ~ a s S c h ~ e e ~ e i s

isareliableforecastingtool,thenthedifferencebetweenthecurrent period's estimate and the next period's estimate will be small. Forecast errors are squared and summed over the entire sample. The best foreca willminimizemean-squareerror. In orderto be useful, an estimator must be more accurate than simply assuming that variance is constant. Usingtheclose-to-closeestimator,Figlewski (1994) demonstratesthat forecasts become more accurateas the amount of historical data used in the estimation is increased. This implies that estimating volatility over short time intervals is not productive. This notion is tested by comparing the mean square error (MSE) of consecutive forecasts with the variance of all forecasts. e§cri~tion of Variance E§timator§

Tests are conducted on eight variations of the Binomial model: Method of moments estimators are calculated using Equation 11 starting with the $ sample range moment and continuing to the second moment in increments,l3 The formulas for each of the eight binomial estimators are given in Appendix 3. Equation 11 requires knowledgeof the tick-size, S. ForTreasuryBonddata,theassumedtick-sizeis of apercent.For S&P500 the tick-size usedis .05 percent. The closing price estimator,14 12 and the Parkinson estimator 7 are calculated for comparison.

Table 2 reports the simulation results. 50-step binomial walks were simulated.Thehigh,lowandclosingvalues of eachwalkwererecorded. Twenty such walks were simulated and the data used to calculate each of theestimators.Thisprocesswasrepeated1,000timesinorder to generate information about the sampling distributionof each estimator. Results indicate that binomial estimators derived from the different moments are approximately unbiased,15 while the Parkinson estimato a downward bias of 16 percent. The CLOSE estimator is also unbiased, but its variance,at 259, is more than five times as large as the varianceof the range-based estimators. Thereis little difference among the different binomial estimators. Estimators based on low moments ou~erformedthe higher moment estimatorsby about 10 percentin terms of efficiency. The Mean Square Error of each estimatoris about twice the estimator variance,

r i

0 0 0

€ ~ c i e n t€ s t i ~ ~ t i oofnIntra-day VoZa~ilify 13

8 m m m o d ooco

14 Ric~ardB. Spurgin and ~

h Sckneeweis ~ ~

s

an unsurprising result. MSE is the variance of the difference between consecutive estimates.If consecutive estimates are independent, the MSE will be twice the varianceof a single estimate. er~ormanceof Variance EstimatorsUsing

This sectionreportstheperformance of thesame 10 estimatorsused in the previous section, this time using contiguous blocks of observed data to calculatetheestimators.Usingactualdataseverelylimitsthe number of estimators that canbe calculated, as overlapping observations areknowntointroducebias.16Twentydays of data are used in each estimate, so the sample size fordaily results is 20. With four periods per day, each intra-day sample contains 80 observations. The 1,196 days of observations result in59 realizations of each estimator (the final 16 days are ignored). Estimators are evaluated using the same methodology as the simulated random walks. The only difference is that the true varianceof the simulated random walksis known. With S&P500 and Treasury bond data this parameter is estimated from the full sample, and hence itself is subject to estimation error. All results are divided by the square of the step-size. Thus the quantity estimatedis not the variance, per se, but the rate at which price changes occur. iscuss~onof S&P500 Results

Boththeintra-dayS&P500results(Table3A)andthedailyresults (Table 3B) show the range-based estimators be negatively to biased relative to the full sample estimate andthe CLOSE estimator tobe mbiased. For example, the Binomial(1) estimator is downward biased by 8.32 percent usingintra-daydataand16.46percentusingdailydata.This bias is entirely dueto the influence of the " m i n i - ~ r a s ~ The ' . ~ ~median estimates for CLOSE and the low-moment binomial estimators are approximately equal for both daily and intraday data.18 The variance of estimates is much larger with actual data than with theiidbinomialsimulation.Thevarianceincreaseisrelativelymore pronounced for binomial estimators than for the CLOSE estimator. The normalize^ variance of CLOSE increases from ,102 to 1.18, a factor of 11.6, while the Binomial(1) estimator increases from ,019 to .357, a factor of 19.3. The efficiencyof Binomial( 1) declines to 3.98 from 5.53. There are a n of possible explanations for the reduced efficiency. First, if vola~lity is timevarying, a given sample is likely to have a greater concen~ationof high or low volatility periods than the population as a whole, resulting in great

E ~ c i e Esti~ation n~ of Intra-day Volatility 15

rii

E ~ c i e n t E s t i ~oftri n o ~t r ~ - Vol~tiZity ~~y 17

dispersion of estimates. Secondly, the distribution of intraday returns is know to have longer tails than the normal distribution (for example see Guillaume et d., 1994, and Granger and Ding, 1994). The efficiencyof all the variance estimates decreases as the tail thickness increases, and the decrease in efficiency is more pronounced for range-based estimators than for the closing price estimator. Forecast efficiency for Binomial estimatorsis considerably better than for the CLOSE estimator. For example, the intraday Binomial(1) estimator has normalized variance of ,357,but the (normalized) forecast MSE is .157. Thus for this estimator, knowledge of the previous estimate is a better predictor of the next period’s estimate than the mean of all forecasts.19 The Variance/FMSE ratioof the CLOSE estimator is 1.03, suggesting the most recent CLOSE forecast is no more informative than the average of all forecasts.20 The efficiencyof low moment binomial forecasts is much higher than any of the quadratic estimators. For example, Binomial has Forecast MSE efficiency of 8.26, a considerable improvement over Binomialt~)at 3.45 and Parkinson at 2.80. A comparison of dailyandintra-dayresultsshowsthatintra-day estimation with range estimators provides a reasonable proxy for daily results (for example, the daily mean estimate for Binomial(1) of 209.99 is almost exactly four tirnes the 51.82 mean intra-day estimate). This is not true of the CLOSE estimator, however, which measures about 10 percent more intra-day volatility than daily volatility. The normalized variance and Forecast MSEof all the estimators is lower for intra-day results than daily results, generally10-25 percent lower. This suggests intra-day data contains useful information about rthe e t u r n - ~ e n e r a process ~g that daily data do not contain.

(i)

is~ussionof T r e ~ s u ~

~uturesResults

Results using Treasury Bond data are largely similar to the S&P500 data. Using daily data (Table 4B), all the range estimators are biased lower. With intra-day data thebias is mixed-low moment estimators are downward biased (Binomial($,-11.8 percent), thosein the middle are about unbiased (Binomial(l),-l.6 percent), and larger moment estimators are positively biased (~inomial(2), 14.4 percent). The efficiency of binomial estimators is lower than with stock data or in simulation (for example, the intra-day efficiency of Binomial(1) is 1.73). The range is capturing more information thanthereturns,butthedifferenceisconsiderablylessthanimplied by theory. Binomial range estimators have a higher forecast efficiency than either the Parkinson estimatoror the CLOSE estimator, The VarianceIForecast

18 ~ i c h a B. r ~Spurgin and Thomas Schnemeis

E ~ c i e n Esti~atio?z t o~lntrff-da~ Volatili~ 19

20

chard B. Spurgin and ~ h o ~ a s

Schn~~~is

MSE ratio for the binomial estimators is about 1.5, indicating the most recent estimate is abetter forecast of the next period's variance than the average forecast. As with the S&P500 data, the CLOSE estimator has a ~ariance/Forecast MSE ratio of 1.0, suggesting the most recent estimate is no more informative than the long-run average. The intra-day binomial estimators outperform their daily counte~arts on the order of 20 percent. For example, the ~inomial(1)estimator has .l52 withdaily normalizedvariance of .l26 usingintra-daydataand data. Forecast MSE using intra-day results is .081 compared to the daily result of .109.However,theintra-dayvarianceandForecastMSE of the CLOSE estimator are actually higher using intra-day data than with daily data, suggesting that for this estimator, the noise introduced by utilizing intra-day data outweighs the benefits of more frequent observations. As withthestockdata,binomialestimatorsappear to scalebetter thantheCLOSEestimator.Theaverage ~inomial(1)estimate of 235.0 transactions per day is almost exactly four times the average estimate of 58.61 transactions perl O O - ~ u t period. e The CLOSE estimator scales poorly, with the daily estimate of 265.26 about12 percent higher than four times the intra-day estimate of 59.29, The bias is about the same ma~itude as with stock data, but the sign is reversed. With stock data, intra-day estimates overes~matedthe daily figure.

The variance estimators proposed in this paper outperform the CLOSE and Parkinson benchmarks. Binomial estimators have lower variance th either estimator and produce more accurate forecasts..h o n g binomial estimators, those based on the lower moments of the range distribution have properties that are well suited to estimating the v o l a t of ~~ intraday financial time series. These estimators place less emphasis on extrem values than quadratic estimators, so they are less sensitive to the long tails generally observed in intraday return series. They also appear to scale q accurately. On average, the s m of intra-day variances was quite close to the daily estimate. '"his was not true of the CLOSE esti~ator.While binomialestimatorsareclearlypreferabletotheCLOSEor Par~son estimators, the estimators proposed here did not perform as well with

E~cient Estima~ionof Intra-day ~olatili~y 21

real data as with simulated data. More importantly, they did not perform well enough to allow for accurate short-term volatility forecasts. Despite theoretical results that suggest highly accurate variance estimates should be possible with small sample sizes, none of the estimators tested here performed close to the theoretical benchmarks. More research in this area is needed. The binomial model seems to provide more accurate variance estimates than the diffusion model, but is itself very limited. Range-based estimators derived from return models thatallow time-va~ingvolatilityshouldprovemorerobustthanthe estimators described here. The use of low sample moments to estimate volatility should be studied further. Tests using different securities, time horizons, and different sample sizes are needed to see if the encouraging results reported here are generally applicable to financial time series or specific to the data analyzed.

E R I ~ ~ T I OOF N RANGE OF A

This appendix describes a generating function that provides exact range probabilities using a recursive approach. It can also be used to rapidly generate a complete set of range probabilities for any walk-length. The approachisessentiallyabrute-forcemethod, as each of thepossible outcomes is accounted for. In this model two attributes are sufficient to describe eachof the 2" outcomes-the range of the outcome and distance from the location of the walk to the lowest location traveled in the walk. All outcomes that share these attributes form an equivalence class. Thus there are at most n2 different classes for each walk-length, which is a much more manageable calculation than 2n for large n. To calculate the range density of an n-step walk, fill a t~ee-dimensionalarray according to the following rules a(n,j, 0) = a(n -1 , j-1 , O ) +a(n -l,j, 1)

a(n,j,~)=a(n-l,j,~-l)+a(n-l,j,~+l),~,ltoj-l a(n,j,j)=a(n"-1,j-1,j-l)+a(n-l,j,j-l) and, a(0, 0,O) = 1 The first column is the length of the walk, the second is the trading range, and the thirdis the location relative to the minimum. Range probabilities

22 ~ i c ~3.r Spurgin d and Thomas Schneeweis

are calculated by summing across the third column and dividing by the total numberof outcomes, 2n. ~ P P E N ~ 2: I XC O ~ P U T A T I O NOF THE ~ I F F E R E N C ECONSTANT

Althoughthe moment~generatingfunctions of theFeller (P) andthe distri~utionof the number of sites visited by a binomial random walk (V) are nearly identical, thereis a small discrepancy which canbe corrected numerically. The ratioof the expected valuesof any moment multip~ed by the walk length will differ from the walk length by a small margin. This margin is essentially independentof the walk length,so the ratio of the expected values converges to 1 as the walk length tends to infinity. For shorter lengths it may make sense to correct for this difference. For example, the relative second moments at three selected walk-lengths are:

El61 n 10.384, error = .0384 For n = 10, E[$

m21

=

For n = 50, 2 , 50.396, error = .0079 E[$]

The2ndmoment

of the V functionexceedsthe2ndmoment

of the

P function by asm 1 constant(approximately0.39).Estimationerror induced by thisdifference is about4percentwhen n is 10 butonly 0.4 percent whenn is 100. W i l e this difference is dominated in practice

by the variance of the estimator, the discrepancy is easily corrected by A1 shows the adding the difference constant to the sample moment. Table difference constant calculated at $ moment increments for the first five moments of the distribution. Figures not listed can be interpolated. Exhibit Al: Difference constant calculated for various sample moments. Sample Dif m o ~ constant ~ t ~___

0.2 0.4 0.6 0.8 1.0

0.5624 0.5424 0.5268 0.5081 0.4986

Sample

mom^^

Dif Dif Sample D@ Sample constant ~ o m cons~unt ~ t omen^ constant

~-

1.2 1.4 1.6 1.8 2.0

0.4790 0.4590 0.4374 0.4153 0.3925

2.2 2.4 2.6 2.8 3.0

0.3688 0.3446 0.3195 0.2808 0.2675

3.2 3.4 3.6 3.8 4.0

0.2401 0.2133 0.1854 0.1570 0.1283

Sumple Dif m o ~ e n t cons~ant

4.2 4.4 4.6 4.8

5.0

0.0989 0.0693 0.0394 0.0009 -0.0211

EjSCicient Esti~ati@n of Intra-dayV@latility 23

INOMIAL ESTIMATORS

The estimators usedin empirical tests are derived from sample moments of observed ranges. The li are observed ranges, S is the step-size, and the variance subscript denotes the sample moment used to generate the 11 estimate. Other estimatorsof this type can be computed with equation and Table 1A. The results reported in Tables 3 and 4 are divided by the square of the tick-size.

( = 0.40956 ( = 0.41812 (

6114~ = 0.41812

li1I4/N) -0.5574s

&p2

li112/N)4-0.5364s

$314~

67.514

= 0.36846

(

l;I4/N)

'l3 -0.5128s

li7I4/N)'l7 -0,4208s I,'/N) -0.3925s

NOTES The trading range is the difference between the recorded high and low price for a security over some time interval. Informative in the statistical sense. Estimatorsderived from the trading rangewill have lower variance than estimatorsderived from returns. For example, see Garman and Klass (1980), Beckers (1983), Wiggins (1991, 1992), and Rogers, Satchell, and Yoon(1984). Parkinson and subsequent authors define the efficiency of a range estimator as the ratio of the variance of the closing price estimator to the variance of the range-based estimator being studied.

24 R i ~ B.~Spurgin ~ d and ~ h o m a S~hneeweis s 5.

6.

7.

8. 9.

10. 11.

Assuming a zero-mean process is not a requirement of the model. However, Figlewski (1994) has shown that assurning a mean of zero yields the most efficientforecasts. For example, consider a common stock that trades in Q point increments. If the high trade is at 7f and the low at 7, then the range is f (or 4 times the trade size) but the number of prices traded is5. Similar results are obtained by Rogers and Satchel1 (1991) in continuous h e . They calculate the amountby which theactual high (low) of a diffusion process exceeds the discretely observed high (low).They find the expected discrepancy to be about 0.90 (as opposed to 1.0 in the discrete case) and the expected squared difference to be 0.28 (as opposed to0.25). A proof of this theoremcan be foundin Freund (1992). The particular example plotted is for a walk-length of20. At smaller walk length, the difference between V and P will be more pronounced.As the walk-length tends to infinity, the moment-generating functions of V and P approach convergence, but this does not take place until n is very large (several thousand). For example, the first moment (mean) of a 100-step random walk evaluated with the P function is nearly 10 percent less than the V function. The reason for the discrepancy is that most realizations of a 100-step random walk will have a range of less than 10. Since the V function adds 1 to observed ranges, small-range random walks areincremented by a large percentage. The mean range of a 100-step random walk slightly is greater than 10, so adding 1step to each range results ina difference between the two means of nearly 10 percent. This estimator is not in wide circulation because it is slightly biased (Garman andKlass, 1980). The size of price changes will certainly increase in periods of extreme volatility. However, for Treasury bond futures, in excess of99 percent of all recorded price changes are in increments. Assuming a fixed walk-length and a variable trade-size would seem an unlikely approach to modeling thissecurity. See, forexample, Guillaume et al. (1994), Bailleand Bollerslev (1990). Moments higher than two were excluded because preliminary tests showed estimators based onthese moments are highly susceptible to misspecification of the tails and hence unlikely to provideuseful forecasts. The CLOSE version employed is the MLE, or populationversion of the estimator.This estimator is more efficient than the standard variance estimator, though it is slightly biased. Both versions were tested. The MLE version was found to less be biased as well as more efficient in the MSE sense, though thedifference was very slight. There is a slight bias in binomial estimators,For an analytical solution to this bias, and a correction formula, see Spurgin (1994). See Figlewski(1994) for a discussion of overlapping intervals. The 20-dayperiod containing the mini-crash produces the highest estimate for both the CLOSE and the Binomial estimators. For CLOSE, this figure is 450.7 (see the Maximum column onTable 3A). Thissingle period accounts for 7.6 percent of the mean of 56.90. However, the maximum Binomial(1)estimate of 157.59 accounts for only 2.7 percent of the meanestimate of 51.82. This differenceaccounts for the estimator bias. This is a potentially useful result for practitioners. Estimates of historic volatility often skyrocket after an "outlier" day and then p l u m e t again 20 or 50 days later when the data point ages out of the sample. Since low moment range estimators place less emphasis onoutliers, this problemis largely eliminated. The mean forecast is only known ex post. In a forecasting model, the mean would be replaced by an estimate from a very long sample.

&

12. 13.

14.

15.

16. 17.

18.

19.

20. This doesnotmean the mostrecent CLOSE estimate contains no ~ o r ~ a t i o If n. consecutive estimates were independent,then the expected value of a Vari~ce/FMSE ratio would be 2.00. There is some i ~ o r m a t i o nin CLOSE, simply not enou it a better forecast than the mean.

Baille, R,T, andBollerslev,T. (1990), "Intra-day and terma ark et Volatilityin Foreign ExchangeRates", Review of Financial Studies,58,565-585. Ball, C. (1988), "Estimation Bias Induced by Discrete Security Prices", ~ u ~ r nofa l ~inance,43,841 -866. Ball, C. and Torous, W. (1984), "The M a x ~ u m Likelihood Estimationof Security Price Volatility: Theory, Evidence, and Applicationto Option Pricing", ~ o ~ r ~ a l of ~ ~ s i n e s97-112. s, Deckers, G.E. (1983), "Variances of Security Price Returns Based on Hi ~siness, and Closing Prices",~ o ~ ~ n a Z o f ~ 56,97-112. Cho, D. Chinhyung and Frees, E. (1988), "Estimating the Volatility of Discrete Stock Prices", ~o~rnaZ of~inance, 43,451 -466. Feller, W. (1951), "The Asymptotic Distributionof the Rangeof Sums of Independent Random Variables", An~als OfMat~e~aticaZ Sta~is~ics/ 22,427-432. Figlewski, S. (1994), "Forecasting Volatility UsingHistorical Data", ~ o r Paper, ~ ~ ~ g Stern Schoolof Business. Freund, J.(1992), M a t ~ e ~ a ~ iStatis~ics, caZ Prentice Hall. Garman, M. and KIass, M. (1980), O ''n the Es~mation of Security Price Volatilities From HistoricalData", ~ o ~ ~ n a Z o f ~53,67-78. ~siness, Granger, C.W.J. and Zhuanxin, D. (1994), Stylized Facts on the Temporal Distributional Properties of Daily Data from Speculative Markets, ~ o ~ ~Pap~r. i n g Guillaume, D., Dacorogna,M., Dave, R., Mtiller, U., Qlsen. R and Pictet, 0. (1994), "From the Bird's Eye to the Microscope: A Surveyof Ne Stylized Facts of the Intra-daily ForeignExchangeMarkets", ~ o r ~ i Paper, ng A Research Group. Heynen, R. and Kat, H. Volatility,in Advanced Applicationsin Finance,T. Schneeweis and D. Ho, (eds), Kluwer Academic Publishing, 1993. Kunitomo, N.(1992), "Improving the Parkinson Method of Estimating Security s, Price Volatilities",~ o ~ r n aofZ~ ~ s i n e s65,295-302. Marsh, T. and Rosenfeld, E.(1986), "Non-tradin~,Market Making, and Estimates of Stock Price Volatility",~ o ~ r nofa FinanciaZ l E c o n o ~ ~ c15,395-472. s, Parkinson, M. (1980), T h e Extreme Value Method for Estimating the Variance of the Rate of Return", ~ o ~ r ofn ~a ~~s i n e s53,61-66. s, Rogers, L.C.G. and Satchell, S.E. (1991), " E s t ~ a ~Variance g from High, Low, and Closing Prices",The ~ n n a l of s AppZied P r o ~ a ~ i Z1,504-512. i~~, Rogers, L.C.G. and Satche11,S.E. and Yoon,Y. (1994), "Estimating the Volatility of Stock Prices: A Comparison of Methods thatUse High and Low Prices", AppZied Financial Econo~ics,4,241 -47.

Spurgin, R.(1994)"Proposal for a derivative security based on the range of anoth security", Working Paper, Universityof Massachussetts. Torney, D.(1986) "Variance of the Rangeof a Random Walk",~ournulof Stu~is~icu~ Physics, 44,49-66. Weiss, G.H. and Rubin, R.J. (1983), "Random Walks: Theory and Selected Applications", A ~ ~ u n cin e s~ h e ~ i cPhysics, ffl 33,363-505. Wiggins, J.B. (1991), "Empirical Tests of the Bias and Efficiency of the ExtremeValue VarianceEsthator for Common Stocks",~ournulof~usiness,64(3). Wiggins, J.B.(1992), "Estimating the Volatilityof S&P500 Futures PricesUsing the Extreme-Value Method", The ~ournulo f ~ u ~ u r~e us r ~ e t12(3), s , 265-273.

~ h i n n - Lin, ~ ~ hJohn ~ n i ~an h t

The purpose of this chapter is to present a model for intra-day prices and volatility generation for equity. In particular, we consider alternative choices of conditio~ngvariables, i.e. exogenous variables, to help us in modelling. Although our methodology is general, we restrict ourselves to two US stocks,IBM and INTEL. We use tickby tick data for January 1994 for these stocks, which were chosen on the basis of their high liquidity. It might be argued that this is insufficient information to carry out our analysis; our response is that our useof data here is illustrative and that a full analysis involving many stocks and longer time periods could be carried outby researchers following the methodology presented he (NUSE) Trade an data comes from the New York Stock Exchange (TAQ) database. This database contains virtually every trade andofquote every stock traded on major American stock exchanges. Relevantliteratureonchoosingsuitableconditioningvariablesis reviewed on p We 28. present ourinitialmodels, investigate their statistical properties, and identify certain problems on p 30. We present details of the data,e ~ ~ m a t i o techniques n ande s t ~ a t i o nresults on p 32. Wefindthatourinformationvariables do notsatisfytherequirement,assumed in themodelonp 30, of beingindependentlyand identicallydistributed. On p 45, we addresstheproblemsdiscussed at p 44 by presentingextendedmodelsbased on doublystochastic processes.Surprisingly,thesemodelsarestraightforwardtoestimate €inancia1~ a r ~Tick t sby Tick Edited by Pierre Lequeux.0 1999 John Wiley & Sons Ltd

for all information variables except volume. We find that volume does to be a suitable variable for measuring information flow, whilst r of trades or the number of price changes seem to work very well. Finally, and importantly in our opinion, we find no evidenceof volatility GARCH models measured on the same sistence. This indicates, to us at least, stence of volatility maybe an artifact of the choiceof model and does not reflect a market opportunity or a forecastable featu of the data.

OICE

ITE

OF ~ O N ~ I T I O ~ I N ~

lied research in high frequency data, as claimed by Gourieroux, Jasiak LeFol(1996), has a rather short history, it may be dated from the First nternational Conference on High Frequency Data in Finance, sponsored en and Associates, March 29-31, 1995. Due to the recent availof data sets, researchers are now able to uncover more interesting atures of asset dynamics at intra-day frequency. Goodhart and O’Hara of e reviewed a host of literature which contains the availability ,statistical properties, problems and difficulties involved with .There is also a long history of using trading informacondi~oningvariable to explain returns/vola~itydynamics at lower frequency. This literature dates back to Osborne (1959) and the (197’3) on stochastic subordination. However, is available on high frequency intra-day data, xplain thereturn/volatili~ processes. not observable practically, difficulties arise proxy forit. In addition to trading er of price changes, other variables e trading volume (trading volume quote changes, and executed irical analysisof stock returns.

~odellingIntra-day ~ 9 ~ iPrices t y and Volatility

Tauchen (1992);’ and to examine stock returns volatility, see Andersen (1996) and Lamoureux and Lastrapes (1994). However, in a recent paper,Jones, Kaul, and Lipson (1994) show that trading volume has no informational content beyond that contained in the number of trades. The useof number of trades as the informational proxy dates back to Osborne (1959), who modified Bachelier’s (1900) random walk modelby incorporating a diffusion process into the evolution of stock prices, withan instantaneous variance dependent on the numberof trades sampled from a uniform distribution. The uniform distribution assumption on the number of trades is however dubious, because transaction time intervals are certainly not uniformly distributed, see Oldfield, Rogalski, and Jarrow (1977). Recently, several researchers have revitalized the use sup~ort for using them of the numberof trades and have given empirical as alternative informational proxies.2 Marsh and Rock (1986) find that the net number of trades (numberof seller-initiated minus buyer-initiated transactions) explain as much as does the net volume. Geman and h 4 (1996) demonstrate that the moments of the time chan needed to induce returnsnormalitymatchthemoments of thenumb of tradesforthe S&P500 one minute returns. Madan and Chang (1997 ropose a variance gamma stock price process and confirm that normality is attainedin the trade-based measure of time. All of this new evidence indicates that the n m b e r of trades could be a better instrument for the no^-quantifiable information than trading volume. In addition to trading volume and number of trades, average tradin~ volume is also usedin the empirical analysisof stock returns. Jones, Kaul and Lipson (1994) actually use average trading volume, instead of total trading volume, in their comparison of the explanatory power of different information proxies. Their justification comes from the obse~ationthat both the number of trades and the average trading volume are highly correlated with the total trading volume; however, is little there correlation Inother words, between the number of trades and average trading volume. the number of trades and the average trading volume seem to contain different information. Quote changes arethe number of times the valid market quotedprices changed throughout the day for a certain security. Although Bollerslev and Domowitz(1993)find that market activity, as measuredby the number of quote arrivals, has no statistically sig~ficanteffect on returns volatili~, Smaby (1995) and Takezawa (1995) suggest that the number of quotes is positively and si~ificantlyrelated to the intra-day volatility of foreign exchange rates,in their recent studies. The equivalent measure to the number of quote changesis the number of pricechangesforthetradedataset.Again,thisvariablehasnot

30 S ~ i n n - ~ Lin u k et al.

often been used as a proxy for information arrival, possibly due to its unavailability. Both the number of quote changes and the number of price changes seem to be intuitively good instruments for the discrete pricejumps often observed in equity markets. In addition to the aforemention formation variables, Locke and Sayers (1993) have also examined the impact of executed orderhbalance in reducing volatilitypersistence, All of these variables are observable and have empirical implication of a random rate of flow of information. In this chapter, we only examine the performance of two most often used information variables, trading volume and the number of trades, and the number of price changes.

A HO~O~ENEO ~O UM S ~ O U N]POISSON D MODEL

In what follows, we assume that there is some variable, N(t), that measures the arrival of information. Such a variable could be volume, number of trades, or numberof price changes, etc, see p29. We shall show that log-returns are conditionally normal with mean variance being linear functions of AN(t), which is the model of Harris (1987). This can be derivedfromthestandarddiffusionandPoisson stochastic differential equation describing the evolution of asset prices, a model motivated by the finding of occasional jumps in the empirical timeseries.Therationale of usingamixedjump-diffusionprocess is that it reveals systematic discontinuities. We define our price generating equation next:

where P(t) denotes the price of an asset at time t, a and CT are parameters, z(t) is a standard Brownian motion, N(t) is a (homogeneous) Poisson process with parameterh, Q is a normal variate with mean EA.Qand variance in the interval(t, t At]. It is straightforward to derive the probability densi function (pdf)of logarithmic returns,X(t) = ln(P(t)/P(t -l)),since

06

+

+ a(z(t) -z(t -1))+

ANff)

2

i=l

We see thatX(t) is independent and identically distributed (i.i.d.) and

+

$(EA.

EA.Q

j , o2

+ 0-6* j )

3

where EA, = a -a2/2,and @(a, b) is the normal density with mean a, and variance b. We nextcomputepdf(X(t)l AN(t)). Simplemanipulations with Equation 2 show that

+

pdf(x(f)lAN(f))= @(p

@Q

AN(f),U2

+

; 0

AN(f))

'

4

It follows, in regression notation, that

where ~ ( t )is distributed N(0,l) and is independent of AN(t). Equation 5 may also be interpreted as a linear regression model with linearheteroskedasticity in AN(t), assuchitis an example of the heteros~edastici~ models popular in econometrics, seeJudge e f al. (1985, 419) for a survey. As long as @Q and :0 are found to be significant,our assumption that AN(t) influences the conditional mean and variance of returns is not rejected. To test if AN(f) influences the rate of returns, an appropriate test would = 0 . We now devote be the joint hypothesis, that is, PQ = 0 and some arguments to testingour various hypotheses.We shall considerthe different hypotheses in turn. Let the three tests be

06

H10 :pQ = 0 Hz0

:0; = 0

06 # 0

vs.Hu :

6

H 3 0 : ~ ~ , ~ = O a n d o h = Ov s . H 3 ~ : p ~ # O o r ~ ; # O A test that p~= 0 implies thatthe number of trades does not influence the expected rateof returns, whilst it increases the volatility of the asset, an assertion investigated by Lamoureux and Lastrapes (1990). A test that : 0 = 0 implies that jump magnitudes are of constant size, albeit unknown to the econometricians. In this case, eacharrival of new information has the same kind of effect on stock prices, i.e. each transaction generates the same amount of trading volume and, consequently, the same impact on prices. Similarly,the joint hypothesis implies that the AN(f) is completely independent of price changes. Since all these hypotheses are interesting, it is worthwhile estimatingand testing our model. We note two of the tests above have the difficulty that the point:0 = 0 lies onthe boundary of the parameter space,so that the asymptotic distributionof the one-sided test ~ ( 1For ) .this reasonthe Lagrange Multiplier will be non-standard, i.e. ~not tests, since (LM) test would be preferred to Wald or Likelihood(LR) Ratio 1) under H0 it is well-known that the LM test retains its ~ ~ (distribution

even for boundary points. Here, computational ease is required at the cost of potential lossof power. It is s~aightforwardto derive theLNI test/ see Breusch and Pagan (1980). The derivation of the score statistics for the three hypotheses is shown in Appendix 1. We present the resultsas a theorem. he LNI tests for our hypotheses given in Equation 6 are:

where ht:= Z2 A

+Z

6

~

~

~

where LMi is the test appropriate for It should be noted that our test procedures are a s ~ p t o t i c a l l yx2(1) for test statistics LMl/ LA42 and x2(2)for test statistics L ~ 3This . is true despite the fact that the alternative hypothesis involves being positive, so that one may wish to use this infor~ationexplicitly. This has been done in general by Rogers (1986) in which he proposes a test procedure based on theKdm"ucker test of Gourieroux, Holly and Monfort (1982). This should lead to a more powerful test but involves substantially ore om put at ion; we shall not investigate this point any further.

06

Data used in this chapter are extracted from theJ atabase, which is produced month1 E).This databasecont.

employed in several intra-d~ytr

Mo~ellinglntra- ay E

~Prices ~ and iVolatili~y ~ 3

Russell (1994) and Engle (1996), and INTEL had the highest total trading volume amongall the stocks availablein that particular month. In January 1994, there were 21 trading days, which are treated separately in the following study. From these 21 days, we use only the "trade" information. Variables recorded for each observation include: a time stamp, a traded price and the associated trading volume (share). Only those transactions that occurred between 9:30 a.m. 4:OO to p.m. are extracted, because both NYSE and NASDAQ, where IBM and INTEL were mostly trade respectively,wereopenduringthatperiod.Althoughtherearesome transactions that happened before 9:30 a.m. and some that happened after 4:OO p.m., the percentageof these exceptions is quite small. Therefore, we decided to delete these observations,We ended up with 13,095 observations of IBM stock traded on NYSE, and 72,831 observations of I N T E ~ traded on NSADAQ.To facilitate our analysisin fixed time intervals, we then sample these tick-by-tick data every minute. The sampling procedure, adopted by Locke and Sayers(1993,117) is described as the following: 1. Select the first recorded tradeas the observation for each minute. 2. Retaintheprevioustradeinformationforthosefollowing ~nutes

with no trades.

This yields roughly 390 observations per day. While sampling l-minute the data, not only havewe extracted the price series,we have also calculated total trading volume (TVol), total number of trades (N), and total number of price changes (NPG) in each one-minute interval. We first examine the returns processes. The stock return concept used in this paper i s the one-minute log-return, defined as the difference of the logarithm prices of two consecutive minutes. By examining Tablel a and lb, we notice that most return processes are highly kurtotic and nonnormally distributed. We compute the Bera-Jarque normality test statistics, which is asymptotically distributedas ~ ~ ( 2"'his ) . test is rejected in most cases for INTEL and is rejected in every case for IBM. Also, most returns processes are not independent and identically distributed (i.i.d.), judgin from the BDS test statistics proposedby Brock, Dechert and Schein (1986). Another salient feature of the returns series is that most of have s i ~ f i c a n t l ylarge negative first-order autocorrelation, as reporte in Table 2. Most literature on the intra-daily analysis of exchanges and stock indices has indicated s i p s of significant negative but first-order autocorrelation, see Andersen and Bollerslev19917), and ~ i ~ l i u o(1991), li and ~oodhartand O'Hara (1997). observes s i ~ f i c a nlarge t negative first-order autocorrel and JPY/US$ exchange rates.He uscri~esthe h~~~u~tucor~e~u~iun tu the noisy

34 S ~ i n n - Lin ~ u et ~ al.

Table l a Descriptive Statistics of Log-Returns (INlTL). MeanNOBS Date

Jan/03/94 Jan/04/94 Jan/05/94 Jan/06/94 Jan/07/94 Jan/10/94 Jan/11/94 Jan/12/94 Jan/13/94 Jan/14/94 Jan/17/94 Jan/18/94 Jan/19/94 Jan/20/94 Jan/21/94 Jan/24/94 Jan/25/94 Jan/26/94 Jan/27/94 Jan/28/94 Jan/31/94

Varia~ce

389 389 389 389 389 389 389 389 389 389 389 389 389 389 389 389 389 389 389 389 389

-4.1632e -005 1.0284e -004 4.0012e -005 -4.001 2e -005 9.9267e -005 2.9048e -005 -4.8095e -006 0.0000e 000 0.0000e 000 -9.3650e -006 -8.6177e -005 3.3667e -005 -1.0221e -005 4.0644e -005 2.9834e -005 I .9775e -005 -9.8683e -006 -5.0554e -006 1.0061e -005 3.9857e -005 0.0000e 000

+ +

+

8.6799e -006 7.9931e -006 7.6488e -006 8.5810e -006 8.6340e -006 8.6570e -006 8.5162e -006 9.9269e -006 6.0198e -006 9.1 142e-006 9.8798e -006 1.4521e -005 2.0207e -005 1.1315e -005 7.1557e -006 1.3870e -005 1.1452e -005 9.2477e -006 8.1759e -006 8.5613e -006 7,1059e -006

-

Skewness BJK~rtosis

-0.074 0.118 0.121 0.096 0,030 -0.267 0.116 -0.180 -0.000

0.018 0.196 0.281 -0.354 0.023 -0.033 0.018 0.036 -0.027 -0.078 0.041 0.023

0.261 1.326 0.997 0.656 1.665 2.862 0.442 4.449 -0.435 0.890 4.878 8.987 17.402 2.334 1.116 27.735 36.546 0.599 0.243 -0.001 -0.373

-

BDS

1.5 29.4 17.1 7.6 45.0 137.4 4.0 322.9 3.1 12.9 388.2 1314.1 4916.5 88.3 20.3 12467.9 21647.6 5.9 1.4 0.1 2.3

8.468 7.546 8.993 7.886 6.582 7.809 6.837 7.961 7.623 6.877 8.440 8.817 8.748 6.910 10.691 7.934 8.044 8.903 7.744 9.430 8.876

*NOBS = Number of Observations,BJ = Bera-Jarque Normality Test Statistics x2(2) with5%critical value = 5.99. BDS = Brock-Deche~-Sche~~an N ( 0 , I)asy, Embedding Dimension = 3, Epsilon = Standard Deviation/Spread

Table l b Descriptive Statistics of Log-Returns (IBM). MeanNOBS Date

Jan/03/94 388 Jan/04/94 387 Jan/05/94 387 Jan/06/94 385 Jan/07/94 388 Jan/10/94 388 Janllll94 Jan/12/94 387 Jan/13/94 389 Jan/14/94 389 Jan/17/94 388 Jan/18/94 389 Jan/19/94 388 Jan/20/94 389 Jan/21/94 386 Jan/24/94 387 Jan/25/94 388 Jan/26/94 382 Jan/27/94 389 Jan/28/94 388 Jan/31/94 389

Variance

389

3.3764e -005 2.7636e -005 2.7229e -005 -5.4684e -005 2.1935e -005 2.1796e -005 -2.1 832e -005 -2.2085e -005 2.1972e -005 -1.0962e -005 -5.0047e -005 -1.6857e -005 -4.0048e -005 -2.8917e -005 5.8679e -006 1.2993e -004 -2.1982e -005 -1.0802e -004 1.6968e -005 -2.2266e -005 -5.0462e -005

1.3221e -006 8,6563e -007 7.3054e -007 1.0495e -006 8.2120e -007 8.8587e -007 8.6346e -007 9.0234e -007 7.9455e -007 7.1724e -007 1.2484e -006 1.1145e -006 9.7589e -007 1.1908e -006 9.1534e -007 2.7850e -005 4.3778e -006 1.6862e -006 1.1024e -006 9.9185e -007 1.0457e -006

Skewness

0.229 -0.123 0.331 -0.206 0.061 0.048 -0.053 -0.048 -0.143 0.197 -0.600 -0.164 -0.084

-0.169 0.016 -0.146 -2.478 -0.352 -0.119 -0.039 -0.077

K ~ rBDS ~ o s i s BJ

2.210 3.260 4.317 1.859 2.590 2.146 2.299 2.133 3.809 4.554 3.788 2.000 2.075 1.856 2.644 25.620 20.380 2.498 3.206 1.768 1.606

82.4 172.3 307.6 58.2 108.7 74.6 85.9 73.5 236.5 338.6 255.3 66.6 70.0 57.7 112.4 10585.2 7112.0 107.2 167.5 50.6 42.2

9.6480 6.0146 3.6733 4.0637 3.9882 1.2024 3.6332 3.3904 4.5641 5.8922 4.2779 7.2851 2.3455 4.2923 2.9040 9.9298 2.9699 3.0933 3.4422 5.7514 2.3132

*NOBS = Number of Observations,BJ = Bera-JarqueNormalityTestStatistics ~ ~ (with 2 ) 5% critical value = 5.99. BDS = Brock-Dechert-~heinkm~N(0, l)asy,Embedding Dimension = 3, Epsilon == Standard Deviation/Spread " ,

1

~odellingIntra-day Equity Prices and Volatility 35

Table 2 First Three Lags Autocorrelationsof the Log-Returns. 1NT'E.L (NASDAQ)

lBA4 (NYSE) Date

Jan/03/94 Jan/O4/94 Jan/05/94 Jan/06/94 Jan/07/94 Jan/10/94 Jan/11/94 Jan/12/94 Jan/13/94 Jan/14/94 Jan/17/94 Jan/18/94 Jan/19/94 Jan/20/94 Jan/21/94 Jan/24/94 Jan/25/94 Jan/26/94 Jan/27/94 Jan/28/94 Jan/31/94

-0.399 -0.302 -0.255 -0.283 -0.272 -0.237 -0.271 -0.316 -0.236 -0.292 -0.227 -0.248 -0.287 -0.309 -0.275 -0.554 0.030 -0.105 -0.203 -0.330 -0.274

0.008 -0.070 0.032 0.077 -0.042 0.012 -0.136 -0.041 -0.074 0.129 0.008 0.022 -0.041 0.109 0.028 0.154 0.102 0.187 0.023 0.024 -0.028

Lug3

Lag1

Lug2

1ag3

0.055 0.027 0.029 -0.125 0.083 -0.027 0.053 0.026 0.058 -0.113 -0.060 -0.012 -0.028 0.010 -0.072 0.001 -0.028 0.008 0.011 -0.073 -0.048

-0.485 -0.464 -0.538 -0.452 -0.476 -0.491 -0.462 -0.467 -0.467 -0.527 -0.477 -0.577 -0.497 -0.514 -0.465 -0.309 -0.524 -0.474 -0.481 -0.536 -0.488

0.047 0.081 0.114 -0.081 0.091

-0.014 -0.091 -0.029 0.124 -0.103 0.080 -0.075 0.067 -0.063 -0.063 0.014 -0.018 0.137 -0.070 0.019 0.074 0.047 -0.019 0.092 0.013 0.069

-0.018

0.055 -0.057 -0.027 0.081 -0.011 0.169 -0.045 0.060 -0.010 -0.216 0.032 0.009 -0.019 0.088 -0.011

* A s s u ~ Gaussian g white noise, the95% confidence intervalof the sample ACF couldbe is about calculated as rfr2/8, where T is the sample size. For our data set, this number rfr0.1014.

s t ~ ~of cthe~~ ~~ ~~ eOther ~ e t possible s . explanations for the negative firstorder autocorrelation include: bid-ask bounce, nonsynchronous trading, and brokers' inventory considerations. We will not model autocorrelation in our data set. However, in later work, we will try to account for this feature. We providesurnrnary statistics of the three trading variables in Tables 3 to 5 below.Weobservethat all threetradingvariablesarepositively skewed, highly kurtotic, and non-normal. Similar to return processes, we reject that these trading variables are i.i.d. in most cases. Comparing IBM and INTEL,two noticeable differences arise: 1. Mlhen all 21 trading dates are pooled together, as shown in Table 6 below, IBM has a lower average number of trades and price changes in each one-minute interval.On average, INTEL has 8.69 trades and 4.54 price changes in each l-minute interval,while IBM onlyhas 1.60 trades and0.33 price changes. Accordingly, thereis a less severe IBM. discreteness problem with INTEL than with

36 S ~ i n n Lin - ~et~al. ~

Table 3a Descriptive statistics of trading volume (INTEL). Mean NOBS Date

Jan/03/94 Jan/04/94 Jan/05/94 Jan/06/94 Jan/07/94 Jan/10/94 Jan/11/94 Jan/12/94 Jan/13/94 Jan/14/94 Jan/17/94 Jan/18/94 Jan/19/94 Jan/20/94 Jan/21/94 Jan/24/94 Jan/25/94 Jan/26/94 Jan/27/94 Jan/28/94 Jan/31/94

V~riance

390 390 390 390 390 390 390 390 390 390 390 390 390 390 390 390 390 390 390 390 390

7257 9778 11397 7621 15337 10946 7537 15224 8333 9849 12826 18317 62260 20868 22353 14056 10155 18585 14278 9490 8823

Sk~ness

+ + + +

1.223e 008 2.395e 008 3.374e 008 1.069e 008 5.906e 008 2.030e + 008 1,382e 008 5.142e 008 1.412e 008 1.944e 008 3.947e 008 9.682e 008 3.883e 009 1.624e 009 1.183e 009 5.656e 008 3.237e 008 9.252e 008 3.513e 008 2.760e 008 3.594e 008

3.644 3.648 3.812 2.665 3.342 2.912 3.436 2.838 2.433 3.253 4.077 6.341 1.855 10.006 3.627 3.912 4.593 3,815 2.237 3.104 9.540

+

+ + + + + + + + + + + + + + +

~ u r ~BDS osis

21.873 16.515 18.175 9.259 13.695 11.483 14.815 9.350 6.476 14.595 25.435 58.890 4.394 146.424 19.061 20.940 28.431 20.081 5.746 11.694 134.074

BJ

8637.6 5297.2 6312.7 1854.9 3774.0 2693.7 4334.1 1944.0 1066.4 4149.4 11593.6 58968.3 537.4 354906.8 6759.0 8120.2 14506.7 7499.1 861.6 2848.2 298024.3

-

7.126 5.192 3.325 6.787 9.360 4.903 5.152 10,873 7.549 7.880 4.045 3.598 11867 8.643 3.134 5.407 3.246 7.063 5.617 8.858 5,950

*NOBS = Number of Observations,BJ = Bera-Jarque Normality Test Statistics ~ ~ ( with 2 ) 5% critical value = 5.99. BDS = Brock-Dechert-S~e~anN ( 0 , l)asy,Embedding Dimension = 3, Epsilon = Standard Deviation/Spread

Table 3b Descriptive statistics of trading v o l w e (IBM). Mean NOBS Date

Jan/03/94 Jan/04/94 Jan/05/94 Jan/06/94 Jan/07/94 Jan/10/94 Jan/11/94 Jan/12/94 Jan/13/94 Jan/14/94 Jan/17/94 Jan/18/94 Jan/19/94 Jan/20/94 Jan/21/94 Jan/24/94 Jan/25/94 Jan/26/94 Jan/27/94 Jan/28/94 Jan/31/94

389 388 388 386 389 389 390 388 390 390 389 390 389 390 387 388 389 383 390 389 390

2781.0 3450.3 5356.4 4816.8 3085.9 3654.2 3145.6 3162.6 3854.6 3101.8 2930.3 4350.8 3928.5 4535.1 5741.l 8797.4 22047.8 8929.5 4600.8 2800.3 3863.3

Variance

Skewness

38576492.9 63085762.2 152626185.8 268190079.3 5454619.2 54680838.9 98414312.6 53332475.8 67155595.6 51283929.9 4088~07.4 73731863.2 78357611.5 76137246.1 240149318.0 241820200.1 1171816934.6 504852033.0 94950462.1 53443427.8 54664127.7

6.289 7.724 6.238 13.836 7.091 3.889 7.440 4,968 4.136 6.687 3.894 3.593 4.496 3.632 9.786 6,477 5.040 12.836 6.040 7.468 4.067

-

~ur~ BDS osis

54.822 93.777 52.767 230.536 69.460 18.455 64.934 31.639 24.657 66.937 19.024 16.041 24.729 19.634 134.918 71.144 43,368 209.362 59.584 82.443 23.864

B]

51278.1 146029.0 47530.5 867097.6 81459.2 6500.7 72113.6 17778.7 10991.3 75715.5 6848.8 5020.3 11222.5 7121.5 299699.6 84540.7 32130.5 710007.0 60062.0 113782.7 10328.7

-

1.4729 3.6474 5.5166 3.2387 2.6029 0.51352 1.5224 2.6565 6.2859 2.2052 4.0623 4.4550 1.3763 0.51898 1 .M53 3.8537 4.8469 0.72295 4.2846 1.3955 1.0396

*NOBS = Nuber of Observations, BJ = Bera-Jarque Normality Test Statistics x2(2) with 5% critical value = 5.99. BDS = ~rock-Dechert-Sche~m~ N ( 0 , l)a7y,E m b e d d Dimension ~~ = 3, Epsilon = Standard Deviation/Spread

Mudelli~gIntra-day Equity Prices and VolatiZi~ 37

Table 4a Descriptive statistics of number of trades (I~EL). Date

Jan/03/94 Jan/04/94 Jan/05/94 Jan/06/94 Jan/07/94 Jan/10/94 Jan/11/94 Jan/12/94 Jan/13/94 Jan/14/94 Jan/17/94 Jan/18/94 Jan/19/94 Jan/20/94 Jan/21/94 Jan/24/94 Jan/25/94 Jan/26/94 Jan/27/94 Jan/28/94 Jan/31/94

NOBS

Mean

Variance

390 390 390 390 390 390 390 390 390 390 390 390 390 390 390 390 390 390 390 390 390

5.6 6.1 7.2 5'9 9.8 8.5 6.1 9.8 5.4 7.4 6.8 9.2 33.9 13.1 9.5 7.4 5.6 7.4 6.6 5.5 5.8

31.1 34.6 62.6 31.1

174.6 59.2 30.9 131.9 19.9 60.1 36.6 50.5 571.6 157.0 69.6 43.0 25.3 40.1 31.0 36.3 55.8

S~wness

3.427 3.565 5.135 2.963 6.211 4.565 3.255 3.911 2.148 4.589 2.553 2.199 1.976 3.828 4.554 3.058 2.868 2.244 2.177 4.436 8.253

Kur~osis

BT

BDS

18.411 19.955 38.847 16.141 61.325 31.809 16.362 18.829 6.437 32.046 9.529 8.540 5.341 21.546 34.008 15.990 11.437 6.740 7.164 27.983 97.587

6271.7 7296.8 26237.0 4804.5 63620.0 17796.9 5039.2 6755.5 973.3 18056.4 1898.9 1499.5 717.3 8496.3 20142.5 4762.7 2660.3 1065.4 1142.0 14003.7 159180.7

13.588 13.303 11.534 13.435 15.036 7,959 9.268 15.674 12.973 12.364 13.493 14.661 23.569 18.091 13.215 13.266 10.706 12.842 14.248 13.969 9,917

-

*NOBS = Number of Observations,BJ = Bera-JarqueNormality Test Statistics x2(2)with 5% critical value = S.99, BDS = Brock-Dechert- hei ink man N (0, l)asy, Embedding Dimension = 3, Epsilon = Standard Deviation/Spread

Table 4b Descriptive statistics of number of trades (IBM). Date

Jan/03/94 Jan/04/94 Jan/05/94 Jan/06/94 Jan/07/94 Jan/10/94 Jan/11/94 Jan/12/94 Jan/13/94 Jan/14/94 Jan/17/94 Jan/18/94 Jan/19/94 Jan/20/94 Jan/21/94 Jan/24/94 Jan/25/94 Jan/26/94 Jan/27/94 Jan/28/94 Jan/31/94

NOBS

Mean

389 388 388 386 389 389 390 388 390 390 389 390 389 390 387 388 389 383 390 389 390

1.4 1.4 1.6 1.6 1.3 1.5 1.2 1.3 1.3 1.2 1.3 ,

1.6 1.4 1.5 1.5 2.6 3.7 2.1 1.4 1.3 1.5

Va~iance

1.8 2.1 2.3 2.3 1.5 2.0 1.6 1.7 1.8 1.7 2.0 2.2 1.9 1.9 2.2 4.9 6.8 3.4 1.9 l .9 1.6

KS~~rw~nuessiss

1.115 1.315 1.835 1.040 0.918 1.020 1.159 1.267 1.381 1.243 1.290 1.060 1.386 1,360 1.292 1.329 0.774 1.131 1.166 1.407 1.035

1.320 1.953 8.170 0.815 0.419 0.979 1.553 2.049 2.008 1.542 1.367 1.144 2.781 2.535 2.246 2.465 0.306 1.407 1.184 2.504 1.605

Bf

BDS

108.8 173.5 1296.7 80.2 57.5 82.9 126.5 171.7 189.4 139.1 138.2 94.2 250.0 224.6 189.1 212.4 40.4 113.2

8.2054 5.1818 4.0105 6,6650 2.2061 2.3133 2.3101 5.0166 6.8374 5.4762 4.1208 4.8445 1.3596 4.5076 5.0264 7.4080 6.9422 5.1466 3.1581 5.4202 0.40399

111.1

230.0 111.5

38 S ~ i n n - Lin ~ uet~ al.

Table 5a Descriptive statisticsof number of price changes (IN'IEL), NOBS Date

Jan/03/94 Jan/04/94 Jan/05/94 Jan/06/94 Jan/O7/94 Jan/10/94 Jan/11/94 Jan/12/94 Jan/13/94 Jan/14/94 Jan/17/94 Jan/18/94 Jan/19/94 Jan/20/94 Jan/21/94 Jan/24/94 Jan/25/94 Jan/26/94 Jan/27/94 Jan/28/94 Jan/31/94

390 390 390 390 390 390 390 390 390 390 390 390 390 390 390 390 390 390 390 390 390

Mean

Vari~nce

3.0103 3.1231 3.6564 3.2026 4.6513 4.4718 3.2103 5.3821 3.0923 4.0282 3.5103 4.8179 17.4667 6.7846 4.9538 4.1231 2.9077 3.9333 3.3846 2.7769 2.9000

7.2081 6.8691 10.7762 8.6504 18.9835 10.1110 9.3696 38.4012 7.5647 12.1560 10.7441 12.9668 144.3575 60.7915 19.9824 11.3833 6.3771 12.4480 7.7334 6.1481 7.5298

S ~ ~ n e s s K~rtosis

1.8741 1.7950 2.1332 1.8694 1.8770 1.5547 2.7147 3.4377 2.1491 2.4341 2.4117 1.8517 1.9079 5.4213 3.6961 1.9455 1.9183 2.1248 1.4063 2,0716 3.5401

-

6.1987 4.3085 6.8053 5.1254 5.0057 3.7506 11.4268 14.5498 7.0446 10.0782 9.8556 5.6909 5.2997 44.4658 22.7886 6.7445 5.5525 6.5732 2.7355 7.7779 23.4108

B1

BDS

852.7 511.1 1048.4 654.0 636.2 385.7 2600.8 4208.2 1106.6 2035.6 1956.5 749.1 693.0 34040.0 9326.9 985.2 740.2 995.6 250.1 1262.0 9720.7

10.415 4.761 9.888 9.497 1.617 6.312 8.374 16.204 10.286 8.647 8.796 10.623 21.771 19.549 9.765 9.573 5.993 9.121 7.200 10.788 4.137

-

*NOBS= Number of Observations, BJ = Bera-Jarque Normality Test Statistics xY(2)with 5% critN ( 0 , I)asyr Embedding DimensionL=: 3, Epsilon = ical value= 5.99. BDS = Brock-~echert-schein~~ Standard Deviation/Spread Table 5b Descriptive statisticsof number of price changes (IBM). Date

Jan/03/94 Jan/04/94 Jan/05/94 Jan/06/94 Jan/07/94 Jan/10/94 Jan/11/94 Jan/12/94 Jan/13/94 Jan/14/94 Jan/17/94 Jan/18/94 Jan/19/94 Jan/20/94 Jan/21/94 Jan/24/94 Jan/25/94 Jan/26/94 Jan/27/94 Jan/28/94 Jan/31/94

NOBS

Mean

Variance

Skewness

Kur~osis

B1

BDS

389 388 388 386 389 389 390 388 390 390 389 390 389 390 387 388 389 383 390 389 390

0.3907 0.2526 0.2088 0.3264 0.2468 0.3033 0.2795 0.2526 0.2000 0.1897 0.3188 0.3487 0.2725 0.3308 0.2248 0.6005 0.8483 0.4099 0,2410 0.2853 0.3154

0.5531 0.3391 0.2896 0.3867 0.3101 0.3614 0.3613 0.2926 0.2272 0.2364 0.3878 0.4950 0.3585 0.3710 0.2887 0.9589 1.1445 0.4938 0.2862 0.3282 0.4016

2.0608 2.8183 3.1204 2.1803 2.4469 2.0460 2.5780 2.1648 2.5245 2.8549 2.0971 2.4629 2.4906 1.9398 2.8445 2.2807 1.5243 1.7337 2.6665 2.0529 2.0610

3.9658 9.4730 11.6976 5.5918 6.0589 3.8552 7.8688 4.2024 6.5558 8.8073 4.3028 6.9755 6.9718 3.7109 9.9777 6.8562 2.3735 2.4843 9.2613 3.8841 3.7419

530.3 1964.4 2841.8 808.7 983.2 512.3 1438.2 588.6 1112.7 1790.3 585.2 1185.0 1190.0 468.4 2127.2 1096.3 242.0 290.4 1856.0 517.8 503.6

9.0637 4.9650 3.7918 3.8511 4.6989 3.9948 3.8849 2.7193 4.1656 5.3012 4.0809 6.4851 0.42557 5.3026 2.3997 4.2912 8.0694 4.1158 2.5195 5.7743 2.3853

*NOBS= Number of Observations, BJ = Bera-Jarque Normality Test Statistics x2(2)with 5% critical value = 5.99. BDS = Brock-Dechert-Schein~an N ( 0 , I)Bsy, Embedding Dimension = 3, Epsilon = Standard Deviation/Spread

12

~odellingIntra-dayEquity Prices and Volatili~y 39 Table 6 Summary statisticsof Tvol, N, and W C when all trading dates are pooled together. INTEL

Mi~mum 1st Quartile Median Mean 3rd Quartile Max~um

TVoE

N

0 0 1000 5184 5000 400000

0.000 0.000 1.000 1.596 2.000 14.000

NPC

TVol

0.0000 0 0.0000 1800 5700 0~0000 15010 0.3259 16180 0.0000 6.0000 92.000 170.000 643000

N 0.000 3.000 6.000 8.689 10.000

NPC 0.000 2.000 3.000 4.542 6.000

2. The IBM stock has a higher percentage of no trade (N = 0) in the

one-minute intervals, as shown in Table 7 below. Overall, there is about 30 percent of no-trades forIBM, while only about2 percent for INTEL. In other words, the nonsynchronous trading problem is much more pronounced for IBM. Therefore, we expect that INTEL would be more suitable for the intra-daily study conducted in thischapter, Table 7 Frequency of No trade. INTEL (NASDAQ) Date

Jan/03/94 Jan/04/94 Jan/05/94 Jan/06/94 Jan/07/94 Jan/10/94 Jan/11 /94 Jan/12/94 Jan/13/94 Jan/14/94 Jan/17/94 Jan/18/94 Jan/19/94 Jan/20/94 Jan/21/94 Jan/24/94 Jan/25/94 Jan/26/94 Jan/27/94 Jan/28/94 Jan/31/94 Overall

N=O

TOT

%

0,321 0.327 0.273 0.282 0.293 0.280 0.338 0.338 0.351 0.359 0.362 0.259 0.314 0.254 0.307 0.160 0.082 0.206 0.285 0.347 0.208

12 8 9 15 9 0 15 2 10 4 11 3 0 0 2 5 13 3 12 17 12

390 390 390 390 390 390 390 390 390 390 390 390 390 390 390 390 390 390 390 390 390

0.031 0.021 0.023 0.038 0.023 0.000 0.038 0.005 0.026 0.010 0.028 0.008 0.000 0.000 0.005 0.013 0.033 0.008 0.031 0.044 0.031

8190

162

N=O

TOT

%

125 127 106 109 114 109 132 131 137 140 141 101 122 99 119 62 32 79 111 135 81

389 388 388 386 389 389 390 388 390 390 389 390 389 390 387 388 389 383 390 389 390 0.0198

We now turn to the estimation of Equation 5, which can be estimated by the following iterative feasible generalized least squares procedure: First, we regress X(t) on AN(^) by ordinary least squares (OLS), and calculate the residuals, say g’s, from the resulting OLS e s t ~ a t e @ s, @Q. In other words,.G = Z(t) = X(t) -6 -A ~ ( t ) 6 ~ .

z2,

Regress g2 on AN(t) by nonlinear least squares (NLS) to obtainand

G5 .3 ).G

,,/m.

Apply generalised least squares on Equation 5 after dividing both sides of the equationby This will produce another set of estimates, @ and @Q. Based on these estimates, calculate the new squared residual E2 = [X(t)-@ -AN(t) @Q) and iterate on step 2 and step3. v

Estimates derived from the above procedure will converge to maximum likelihood estimates by a familiar linearized m a x i ~ u mlikelihood argument.Usually,onlythefirstthreeiterationsarerequiredtoproduce convergent estimates. Initially,weuseallthreetradingvariables,Tvol, N, andNPC, as i~ormational proxies. However, we find that wheneverisTvol employed, theaboveiterativeprocedurehastroubleconverging.Thisindicates that Tvol is not a suitable informational proxy in our model. It is also consistentwithrecentempiricalfindingontheinformationalrole of trading volume, as described on p 29. Therefore, we exclude Tvol as one of the informational proxiesin the following analysis. The estim~tionresults are reported in Tables 8 and 9 below. Judging from the t-statistics, neither the number of trades nor the nurnber of price changes significantly influence the mean and the variance of returns of INTEL. In contrast, both number of trades and numberof price changes have significant impact on the variance of returns of IBM. This may be related to the data in Table6 (on p. 39 above) where the higher numbers of tradesand price changesforELrelativetoIBMmeanthattheir impact is less import~nt.Technic it is as if the Poisson process maybe converging to Brownian motion again. To further investigate the effect of the r of trades and the in number of prices changes,weconducttheLroceduresdetailed Theorem 1.Theresultsarereported in Table 43 below. In general, we cannot reject the null hypothesis that @Q = 0 for neither INTEL nor IBM. The null hypothesis that 2; = 0 cannot be rejected for most INTEL cases, while it is rejected for all IBM cases. These findin and LM2 are consistent with the findings based on the previously rep

~odellingIntra-day Equity Prices and Vol~t~lity 41

Table 8a Estimati~nresults of equation 5 tioning variable.

Jan/03/94 Jan/04/94 Jan/05/94 Jan/06/94 Jan/07/94 Jan/10/94 Jan/11/94 Jan/12/94 Jan/13/94 Jan/14/94 Jan/17/94 Jan/18/94 Jan/19/94 Jan/20/94 Jan/21/94 Jan/24/94 Jan/25/94 Jan/26/94 Jan/27/94 Jan/28/"94 Jan/31/94

1.99623e -004 1.52690e -004 -1.84342e -004 1.29885e -005 -8.05041e -005 -3.71835e -005 -3,31412e -005 1.14074e -004 1.16775e -005 -3.3867% -006 -4.19065e -005 2.1825963 -004 4.65174e -005 -3.57308e -004 -1.07105e -004 -5.80505e -005 7.78902e -005 4.80304e -005 -4.31422e -005 -1.59122e -004 -7.14433e -005

(INTEL)using number of trades as the condi-

-4.32879e -005 -8.23055e -006 3.11915e -005' -8.96784e -006 1.83980e -005 7.85007e -006 4.73006~3-006 -1.17300e -005 -2.19872e -006 -8.20296e -007 -6.56040e -006 -2.00913e -005 -1.67423e -006 3.07089e -005 1.45346e -005 1.05155e -005 -1.58708e -005 -7.18525e -006 8.07815e -006 3.63564e -005 1.23353e -005

8.83767e -006$ 8.42912e -006$ 7.84127e -006$ 9.26315e -0063 9.46621e -006$ 9.05212e -006$ 8.63678e -006$ 1.13677e -005% 4.49239e -006$ 9.38246e -006$ 8.99713e -006$ 9.17177e -006$ 1.09268e -005 5.67778e -006$ 7.75926e -006$ 1.73548e -005$ 1.06255e -005' 1.04176e -005$ 8.16688e -006$ 8.95511e -006$ 7.52336e -006$

1.31247e -015 3.86243e -014 5.28197e -017 4.19801e -014 1.24699e -014 5.3i904e -014 9.96011e -014 1.40633e -014 2.84593e -007$ 4.41067e -016 1.26637e -007 5.78238e -0071 2.72177e -007 4.36347e -007$ 6.49908e -014 1.29492e -018 1.43156e -007 1.68758e -015 3.52169e -017 4.18263e -015 5.03337e -020

*Superscripts 'and $indicatesignificant estimates at95% and 99%, respectively

Table 8b Estimation resultsof equation 5(ISM) using number of trades as the c o n d i t i o ~ g variable.

Jan/03/94 Jan/04/94 Jan/05/94 Jan/06/94 Jan/07/94 Jan/10/94 Jan/11/94 Jan/12/94 Jan/13/94 Jan/14/94 Jan/17/94 Jan/18/94 Jan/19/94 Jan/20/94 Jan/21/94 Jan/24/94 Jan/25/94 Jan/26/94 Jan/27/94 Jan/28/94 Jan/31/94

8.47565e -005 3.54840e -005 1.72855e -005 -2.50751e -005 -3.13990e -005 2.20453~3-005 -1.92161e -005 3.58428e -005 -5.14234e -005 1.31612~3-005 3.24073e -005 -1.05613e -005 -1.17997e -004+ 3,80987e -005 3.71696e -005 -6.19560e -005 1.10154e -004 -4.42420e -006 -6.88986e -005 -5.00181e -005 -1.13336e -004

-3.76855e -005 -5.71942e -006 6.370910 -006 -1.88733e -005 4.06552e -005 -1.62833e -007 -2.09804e -006 -4.62232e -005 5.83860e -005 -1.96320e -005 -6.37293e -005 -3.88737e -006 5.77178e -005 -4.57353e -005 -2.09765e -005 7.47843e -005 -3.54554e -005 -4.93458e -005 6.11763e -005 2.18855e -005 4.08308e -005

5.06112e -007$ 4.60576e -0073 4.39923e -007$ 6.22927e -007$ 3.46749e -007-1: 4.95472e -0071 4.51786e -007$ 5.31211e -007$ 4.26358e -0071: 3.78540e -007$ 6.45929e -007$ 5.49921e -007$ 6.74295e -007$ 7,20096e -007$ 5.73443e -OO~$ 7.40194~3-006 1.90160e -006 3.84824e -007 6.88203e -007$ 6.06988e -007$ 6.94322e -007$

6.01113e -007$ 2.94003e -007$ 1.84871e -007$ 2.70108e -007$ 3.59012e -007$ 2.53087e -007$ 3.28378e -0073: 2.93435~3-007$ 2.80516e -007$ 2.74229~3-0071 4.56689e -007$ 3.46613e -007$ 2.17576e -007$ 3.17926e -007$ 2.27453e -007$ 7.93658e -006$ 6.62572e -0071 6.16205e -0 0 s 2.82413e -007$ 2.99088~3-007$ 2.23505e -007$

*Superscripts 'and $indicate significant estimates at 95% and 99%, respectively

Table 9a Estimation results of equation 5 (INTEL) using number of price changes as the conditioning variable. Date

Jan/03/94 Jan/04/94 Jan/05/94 Jan/06/94 Jan/07/94 Jan/10/94 Jan/11/94 Jan/12/94 Jan/13/94 Jan/14/94 Jan/17/94 Jan/18/94 Jan/19/94 Jan/20/94 Jan/21/94 Jan/24/94 Jan/25/94 Jan/26/94 Jan/27/94 Jan/28/94 Jan/31/94

E 2,79663e -004 1.29022e -004 -6.15312e -005 -1.28360e-005 -2.41464e -004 -1.61753e -004 -2.86262e -005 1.66276e -005 6.76371e -005 1.38520e-004 -1.54735e-004 1.83196e -004 7.12381e -005 -8.547470 -005 -1.68194e -004 -9.62666e -005 2.47163e -004 5.77888e -005 -4.74702e -006 -8.70525e -005 -1.22786e-004

OQ

-1.07282e -004? -8.41686e -006 2.77974e -005 -8.49783e -006 7.32694e -005? 4.27296e -005 7.48359e -006 -3.1 1267e-006 -2.20174e -005 -3.68765e -005 1,95953e-005 -3.10886e -005 -4.66889e -006 S.86825e -005 4.00588e -005 2.82125e -005 -8.90341e -005 -1.60304e-005 4.37727e -006 4.58381e -005 4.24943e -005

8.90754e -006% 7.85017e -006% 7.20774e -006% 8.23181e -006t 9.62282e -006% 7.12497e -006% 7.87763e -006% 1.08291e -005% 4.87449e -006' 8.52530e -006% 7.3076Se -006% 9.06530e -006? 1.097 12e-005 5.81773e -006% 6.81423~3-006% 1.77812e -005% 8.93494e -006? 9.59514e -006% 6.80104e -006% 7.87873e -006% 7.32542e -006%

1.19623e-014 3.91817e -008 1.12640e-007 1.02132e-007 2.47580e -014 3.35560e -007 1.93200~3-007 4.44068e -014 3.65487e -0075 1.37965e -007 7.28027e -007+ 1.12965e -006? 5.25986e -007 8.16752e -007% 6.0042963 -008 1.68407e-021' 8.32762e -007 2.0816k -016 4.00279e -007' 2.30541e -007 5.38475e -014

*Superscripts 'and *indicate significant estimates95% at and 99%, respectively Table 9b Estimation results of equation 5 (IBM) using n u b e r of price changes as the c o n d i ~ o variable. ~g

Jan/03/94 Jan/04/94 Jms/05/94 Jan/06/94 Jan/07/94 Jan/10/94 Jan/11/94 Jan/12/94 Jan/13/94 Jan/14/94 Jan/17/94 Jan/18/94 Jan/19/94 Jan/20/94 Jan/21/94 Jan/24/94 Jan/25/94 Jan/26/94 Jan/27/94 Jan/28/94 Jan/31/94

6.02553e -005 7.58812e -005% 1.57119e -005 -5.64767e -005 -5.25477e -006 3.78886e -005 -4.20013e -005 4.23226e -005 -1.2703le -005 3.21602e -005 -3.23145e -005 -6.46973e -005' -6.73258e -005' -4.76575e -005 2.32398e -005 3.60744e -005 4.23419e -005 -2.17494e -005 -1.46022e-005 -2.89158e -005 -1.95561e -005

-6.76212e -005 -1.90518e -004 5.50246e -005 5.47722e -006 1.09891e-004 -5.29159e -005 7.19807e -005 -2.54346e -004t 1.72930e-004 -2.26685e -004 -5.54844e -005 1.36837e -004 9.98472e -005 5.65120e -005 -7.70753e -005 1.55893e-004 -7.56287e -005 -2.09918e -004 1.30648e -004 2.32429e -005 -9.77442e -005

6.17154e -007' 3.69360e -007' 3.68642e -007' 4.85193e -007% 3.22471e -007% 4.15901e -007% 3.31776e -007% 3.93Me -007% 2.66369e -007% 3.07903e -007% 5.33500e -007% 4.75595e -007% 5.11553e -007% 5.80608e -007% 4.42601e -007% 7.16779e -016 1.93266e -006 7.66423e -007% 4.41051e -007g 3.72453e -007% 4.68804e -007%

*Superscripts ?and %indicate significant estimates 95% and at99%, respectively

1.76942e -006' 1.9058le -006% 1.72722e -006% 1.71530e -006% 2.00768e -006% 1S2934e -006% 1.88459e -006% 1.91947e -006% 2.60886e -006% 2.08081e -006% 2.22843e -006% 1.83636e -006% 1.66638e-006% 1.82703e -006% 2.07353e -006% 5 . 0 6 W -005% 2.86363e -006% 2.21501e -006% 2.68871e -006% 2.15573e -006% 1.82722e -006%

Modelling Intra-day Equity Prices and Volatili~ 43

Table 10a LM test statistics for equation 7( I m L ) .

N Date

LM3 LMZ

LM2

Jan/03/94 Jan/04/94 Jan/05/94 Jan/06/94 Jan/07/94 Jan/10/94 Jan/11/94 Jan/12/94 Jan/13/94 Jan/14/94 Jan/17/94 Jan/18/94 Jan/19/94 Jan/20/94 Jan/21/94 Jan/24/94 Jan/25/94 Jan/26/94 Jan/27/94 Jan/28/94 Jan/31/94

2.532 0.108 3.067 0.104 2.418 0.156 0.030 0.603 0.005 0.002 0.051 0.387 0.022 2.083 0.720 0.105 0.180 0.077 0.096 2.073 0.437

0.1 0.6 0.3 1.2 4.0 0.4 0.0 5.7 7.6 0.2 1.1 15.1 20.1 27.2 1.2 9.6 0.7 2.3 0.0 0.7 1.3

NPC 295879.2 14125.0 402552.7 13133.5 309104.0 18768.8 3365.8 67172.3 10195.2 177.7 12681.1 20391.7 6482.2 107935.9 110417.5 9392.3 16355.5 9226.3 11746.1 251009.2 65353.1

LMl

LM2

3.650 0.023 0.373 0.027 4.164 0.713 0.018 0.013 0.176 0.620 0.117 0.234 0.044 0.286 1.524 0.196 1.161 0.130 0.006 0.399 0.720

0.2 0.0 0.4 0.2 2.9 2.9 0.8 2.2 5.0 0.5 11.1 13.8 19.0 37.7 0.2 10.6 5.8 0.3 3.5 0.7 0.2

LA43

418504.9 3160.5 73346.8 2903.5 533145.9 86954.4 14412.3 1424.7 101769.2 64249.0 6623.7 579.5 10397.1 11080.5 214287.9 18242.3 333147.9 14433.2 288.3 142218.1 103814.5

Table lob LM test statistics for equation 7 (IBM).

N Date

m 2

M 1

LM1

NPC LM3

m 2

LM3

187.0 323.4 303.4 193.6 353.8 207.2 334.2 268.4 473.6 398.6 236.6 231.7 203.6 169.5 288.4 612.7 93.6 155.2 330.1 297.7 226.6

3123695.1 8061427.7 33255.6 311366.5 918565.3 1960846.2 1095101.6 13537747.4 3264487.8 11441683.4 75114.1 396386.6 3656040.4 495987.9 1846759.1 23515.9 70446.8 2172753.1 3873798.5 77788.5 75852.5

.~

Jan/03/94 Jan/04/94 Jan/05/94 Jan/06/94 Jan/07/94 Jan/10/94 Jan/11/94 Jan/l2/94 Jan/13/94 Jan/14/94 Jan/17/94 Jan/18/94 Jan/19/94 Jan/20/94 Jan/21/94 Jan/24/94 Jan/25/94 Jan/26/94 Jan/27/94 Jan/28/94 Jan/31/94

0.607 0.022 0.038 0.243 1.007 0.000 0.003 1.188 2.092 0.273 1.863 0.010 1.975 1.016 0.330 0.328 0.675 1.734 1.785 0.247 0.840

69.5 41.4 26.6 24.8 55.1 32.0 44.4 34.1 41.9 48.5 51.5 40.9 18.4 25.4 27.5 77.9 29.4 84.3 21.8 29.9 14.1

390700.6 420148.4 236021.4 215394.7 2286418.3 196540.3 49963.4 1571143.4 14511511.5 295116.3 3808498.9 427964.1 3463029.2 1136782.0 330127.5 24908.4 96354.9 2195317.4 7945025.7 2203044.0 2574018.3

0.289 1.539 0.119 0.002 0.502 0.168 0.256 2.618 0.813 1.548 0.138 1.139 0.482 0.168 0.206 0.111 0.375 2.314 0.496 0.024 0.519

t-statistics. However, theLM3 test is s i ~ f i cin~all t cases. This provides an evidence that return processes are indeed influenced by the trading processes. The insignificant results are possibly due to modelling 116sspecification. One wayof examining the modelling mis-specification is to examine the independenceof the fitted residuals from the model. Since o ~running in Equation 5 is assumed to be i.i.d., we test this a s s u ~ p t i by the BDS test on the standardized estimated residuals, namely

&(t)=

h

X(t) -j 2 -@Q 'A N ( t )

8

From those results reported in Table 11 below, the i.i,d. assumption on is clearly rejected for most cases. This leads us to the doubly stochastic modelling in the next section. Table 11 BDS tests on the estimated standardized residuals.

Jan/03/94 Jan/04/94 Jan/05/94 Jan/06/94 Jan/07/94 Jan/10/94 Jan/11/94 Jan/l2/94 Jan/13/94 Jan/14/94 Jan/17/94 Jan/18/94 Jan/19/94 Jan/20/94 Jan/21/94 Jan/24/94 Jan/25/94 Jan/26/94 Jan/27/94 Jan/28/94 Jan/31/94

8.617 7.546 9.205 7.886 6.410 7.809 6.837 7.725 7.647 6.877 8.231 7.615 6.974 6.255 10.590 7.959 7.923 8.903 7.744 10.022 8.953

8.838 7.546 8.993 7.886 6.387 7.928 6.897 7.961 7.698 6.930 8.258 7.146 8.734 6.997 10.271 7.658 6.734 8.903 7.822 9.524 8.878

9.1082 5.9476 3.6334 4.0592 3.9121 1.1909 3,5701 3.3836 4.5410 5.8569 4.0395 7.2038 2.3438 4.2690 2.8906 9.7016 0.57503 2.4349 3.3970 5.7085 2.3108

X ( t ) -E -EQ A N ( t )

Standardized Residual:

-4

-

BDS = Brock-Dec~ert-Sche~manN(0, l)asy, Embedding Dimension = 3, Epsilon = Standard Deviation/Spread

7.3948 3.3806 3.5866 2.9863 3.5160 0.99458 3,8187 2.0141 3.0972 4.3545 2.5102 5.3387 2.5470 3.9588 2.8051 -0.41132 0.045457 -0.58903 1,2443 5.0051 1.9430

~ o ~ e l l i ~nng~ r a - dEquity ay Prices and V o l a ~ i l i 45 ~

We have shown that the information variables are not consistent with the (homogeneous) Poisson process with afixed parameter. h order to accommodate the non-homogeneous nature of the information data, we now introduce a more complex (non-homogeneous) Poisson process which allowsh(t)to vary. By an appropriate choice of h(t),we can model / 1)). ~ ("here t are many candidates the marginal distri~utionof ~ ( ~ ( t )for the process of h(t).Here, h ( t )is definedin the form of a GARC model as: ~ ( t=) av2(t -1) ~Var(X(t-1)/1(t-2)) 9

+

where v(t) is N(0,l) u~conditionalin N(t), in fact, it is Az(t) = z(t) -z(t 1)and I(t) contains informationup to the end of the minute? Heuristica~ly, the expected numberof jumps depends upon the previous volatility and the deviation from fundamental Az2(t), see Equation2. It follows from Equation5 that 10 since h ( t )is known givenI(t -1).To simplify our model we shall assume that X(t) and b N ( t ) are (weakly) stationary. Under the assumption of weak stationarity Equation 9 becomes

+

where 8 = EA,^ G;), 0 0 1. Wecan calculatethemeanand variance of h(t),detailed in Appendix 2, as follows: Bo2

the

+a

E[h(t)]= 1-8

2a2 Var(h(t)) = 1 -e2

12

using the factthat v2( t )has a ~ ~ (-1 distribution. ) Given our model, the information variable is not purely exogenous any more:its intensity is dependent upon the past history of prices. This framework is attractive because it allows a feedback effect through the variables. It can explain certain phenomenain financial time seriessuch

46 S ~ i n n - ~ Lin u h et al.

as volatility clustering where large price changes tend to bunch together. This non-homogeneous Poisson process also resolves the restrictive asp of the homogeneous Poisson distribution which implies that mean and variance are equal. From the moment generating function of AN(t) derived in Appendix 2,” (1 -2ad(exp(s) -

13

we can derive the mean and the variance of AN(t),

Var(AN(t)) =

Bo2+A!

1 1- 0- 0 ~

14

+2a2

Note that the mean and the variance are not equal in the presence of stochastic h(t). Moreover, the serial correlation of A N ( t ) can be shown to be Corr(hN(t), AN(t -S)) =

Cov(BSh(t-S ) , AN(t -S ) ) Var(AN(t))

-OsVar(h(t -S ) )

15

Var( hN(t))

i f s = 1 , 2 , ...

ThenfromEquation5,themomentconditions Appendix 3, are obtained as follows:

of X(t), detailed in

Corr(X(t), X(t -S ) ) =

I19

ifs=O

Modelling Intra-day Equity Prices and Volatili~ 47

Given that AN(t) is observable, the joint likelihood functionof X(t) and AN(t) can bewritten as T

T

pdf(X(t),AN(t)lI(t -1)) pdf(X(t)lAN(t),I(t -1)) pdf(AN(t)lI(t-1)) *

17

k l

Thus, the log-likelihood function becomes

Although the joint density is tractable, the marginal density of X(t) and AN(t) cannotbederivedexplicitly.Thisisone of thecharacteristics of a mixture of distributions whichcontainseveralrandom variables. We estimate Equation 17 and report the results in Table l 2 and 13 below. In terms of the significance of coefficients, @Q and l?& in the varianceequation,weobtainsimilarresultsasthoseobtainedfrom 5. estimating the homogenous compound Poisson model in Equation Namely, G; is highly significant for all IBM cases, while @Q is insignifIn addition, 3 ishighlysignificantforallcases. icantinmostcases. From Equation 9, we know that 3 measures the sensitivity of information arrival intensity h(t) to theprevious period's realized volatility. This is a strong evidence for the existence of a stochastic arrival intensity process. Another way of examining the performance of the doubly stochastic Poisson model is to compare the sample moments (mean and variance) of

I

I

I

I

I

I

I

I

I

I

I

I

I

I

I

I

I

l

I

I

I

I

-

I

t

&%U",wawww%wwm I

388888888888

I

t

I

I

I

I

I

I

I

I

";t(U",U",lnU",U",%w

U

,

I

I

l

I

I

I

I

I

88888888

I

I

I

I

I

I

I

I

I

I

I

I

I

I

I

I

I

I

I

I

~ o d e l l i n ~ntra-day g Equity Prices and Volatili~ 49

I

m m * m m * ~ m m m m m m m m m m ~ m m m I

I

I

I

I

I

I

/

I

I

I

l

l

I

I

I

I

I

/

I

/

0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

m m m * m * m m m m m m + m m * w m m m m I

I

I

I

I

I

I

I

I

I

I

I

I

I

I

I

I

I

I

I

I

888888888888~88888888

50 S h i n n - ~ ~Lin h et al.

l

l

l

l

l

l

l

l

l

l

l

l

l

l

l

l

l

l

l

l

l

l

l

l

l

I

l

I

l

I

l

l

I

l

.

i

I

.

l I

.

I

.

I

l

I

I

l

I

l

I

l

I

l

I

l

l

I

l

I

l

I

l

l

l

l

l

I

l

I

l

l

l

I

l

I

l

l

l

/

l

l I

I

1

~ o d e l l i n gl ~ ~ r a - d E aqyuityPrices and Volatility 51

I

l

52 S ~ i n n Lin - ~et~al.~

trading variables with those implied by the model in Equations 9 to 12. The results are reportedin Tables 14 and 15 below. h interesting result shorn in these tables is that the implied expected values, ~ ( ~ of~the~ ) ~ ( ~ quite ~ ~well, ) To , trading variables match their sample counterparts, give a quick measure of how close they are, we calculate ax2 test statistics TCF as

All of the TCFs are well inside the critical region under conventional significance levels. By comparing the ~ a ~ i t ~ofd TCFs, e s we also find that,forbothINTELandIBM,thenumber of tradesseemto be a slightly better proxyof information arrival.h addition, for IBM, implied Table 14a Moments and persistence level of information arrival (N of INTEL).

Jan/O3/94 Jan/04/94 Jan/O5/94 Jan/06/94 Jan/07/94 Jan/10/94 Jan/11/94 Jan/12/94 Jan/13/94 Jan/14/94 Jan/17/94 Jan/18/94 Jan/19/94 Jan/20/94 Jan/21/94 Jan/24/94 Jan/25/94 Jan/26/94 Jan/27/94 Jan/28/94 Jan/31/94

5.600 31.058 5.914 6.079 34.567 6.716 7.223 62.621 7.189 5.928 31.090 5.896 9.779 174.604 9.834 8.467 59.211 8.387 6.051 30.861 6.116 9.831 131.946 9.709 5.362 19.908 6.314 7.359 60.112 7.286 6.787 36.626 6.746 9.213 50.461 9.258 33.931571.5611395.703 13.064 156.955 38.734 9.456 69.565 9.386 7.438 42.987 7,386 5.579 25.283 5.503 7.423 40.132 7.486 6.597 30.951 6.668 5.500 36.256 7.991 5.808 55.806 6.079

5.866 6.459 7.166 5,896 9.749 8.384 6.002 9.631 5.574 7.235 6.744 9.177 28.760 12.924 9.370 7.376 5.499 7.354 6.537 5,458 5.732

4.7703e -002 1.1138e -002 1.2062e -002 2.5707e -001 9.1279e -003 2.2356e -002 2.2876e -002 1.8145e -002 4.5339e -004 -004 6.1 175e -005 1.7368e -004 1.1370~: 8.4888e -002 1.1503e -004 9.2317e -005 3.1 142e -003 9.0166e -005 8.2168e -004 1.1343e -001 2.3370e -005 4.0003e -004 7.7256e -002 5.6944e -005 4.1533e -003 7.4047e -001 2.6326e -002 8.0632e -003 5.1082e -002 7.7275e -003 2.1252e -003 2.5314e -003 1.5969e-001 2.7417e -004 8.1291e -002 1.5376e -001 1.4122e -004 1.3669e 003 9.5107e -002 9.2974e -001 2.5810e 001 4.9379e -001 1.5166e -003 1.6071e -002 2.1208e -004 7.8933e -004 9.1233e -003 6.5947e -005 5.2115e -004 3.6946~3-003 2.5099e -004 l .1638e -003 1.3198e -001 2.7106e -002 6.4740e -004 1.3129e -001 6.7783e -004 5.5071e -004 2.5329e 000 3.8894e -002 3.2320e -004 3.4652e -001 2,9079e -002 1.0077e -003

+ +

+

TCF

9.8737e-001

~ o d ~ l lIntra-day i n ~ Eqrrity Prices and Vola~ili~y53

Table 14b Moments and persistence level of information arrival (N of IBM).

Jan/03/94 Jan/04/94 Jan/05/94 Jan/06/94 Jan/07/94 Jan/10/94 Jan/11/94 Jan/12/94 Jan/13/94 Jan/14/94 Jan/17/94 Jan/18/94 Jan/19/94 Jan/20/94 Jan/21/94 Jan/24/94 Jan/25/94 Jan126194 Jan/27/94 Jan/28/94 Jan/31/94

1.352 1.371 1.562 1.568 1.316 1.537 1.246 1.257 1.256 1.228 1.296 1.618 1.352 1.467 1.504 2.568 3.728 2.129 1.403 1.270 1.538

1.780 2.053 2.324 2.256 1.505 2.048 1.579 1.655 1.847 1.724 2.013 2.216 1.863 1.869 2.271 4.885 6.750 3.539 1.871 1.852 1.596

1.351 1.353 1.564 1.588 1.311 1.516 1.248 1.258 1.235 1.235 1.277 1,661 1.349 1.463 1.496 2.562 3.753 2.257 1.409 1.251 1.550

1.336 1.347 1.554 1.573 1.311 1.512 1.244 1.253 1.234 1.228 1.272 1.634 1.346 1.458 1.495 2.562 3.715 2.121 1.405 1.249 1.545

1.4869e -002 6.0051e -003 9.8332e -003 1.5193e -002 5.0583e -004 4.6982e -003 3.2484e -003 4,0820e -003 4.0109e -004 7.6988e -003 4.5813e -003 2.7289e -002 2.7381e -003 5.1599e -003 1.4632e -003 1.1785e -004 3.8233e -002 1.3656e-001 3.7588e -003 1.8914e -003 4.2753e -003

0.6775 0,6881 0.6195 0.6830 0.8801 0.7322 0.7263 0.7427 0.6367 0.6955 0.6434 0.6504 0.5076 0.6404 0.6493 0.3052 0.5157 0.5969 0.5917 0.6889 0.6025

1.9162e -004 4.2762e -004 4. 1184e -005 1.5893e-005 1.9069e -005 4.1336e -004 3.2154e -006 1.2769e -005 3.9222e -004 0,OOOOe 000 4.5283e -004 1.5667e -004 2.6746e -005 5.5556e -005 5.4181e -005 1.4052e-005 4.5491e -005 3.0174e -005 2.8470e -006 3.5308e -004 3.1715e -005

TCF2.7403e E = Mean, V = Variance,

E = Estimated Mean,

CF = Criterion Function =

EtNI

,TCF ==

+

-003

= Estimated CFt

variances, p ( A N f ) ,of the trading variables are also very close to their sample counterparts,V (ANt). Also reported in Tables 14 and 15 are the g values, which represent the degree of dependence of h(t) on h(t -1) from Equation 11. This is an equivalent measure of volatility persistence in GARCH models. On average, when the number of trades is used as the informational proxy, 8 ctr 0.05 for INTEL, and g zz 0.64 for IBM. Similar results are obtained. when the numberof price changesis used as the informational proxy.h that case, ctr 0.1 for INTEL, and ctr 0.64 for IBM on average. In other words, the volatility persistence implied by our model is much smaller than those impliedby the GARCH-type models. To comparethepersistence of GARCH-typemodelsversusthat of informational volatility models as in this chapter, we fit a GARCH(1,l) model on the same data sets. We compare gthe values listedin Tables 14

h

g

54 S h i ~ ~ Lin - ~etual. ~

+3

and 15 with the valueof G reported in Table 16 below. We recall that Table 14 describes the models for the number of trades, whilst Table 15 describes the models for the number of price changes. In Table 14a, for INTEL, there are no valuesof G greater than 0.5, and there are only three values greater than 0.1. Similarly, in Table 15a, there are onlysix values of 5 greater than 0.1. However, for the GARCH(1,l) model for INTEL in greater than 0.9, and 12 values Table 16, there are four values of G no values greater than0.5 out of the 21 days. Likewise, for IBM, there are of G greater than 0.8 in either Table 14b or15b. But, for the CARCH(1,l) greater than 0.8. model for IBMin Table 16, there aresix values of G This indicates, to us at least, that the claimed persistence of volatility may be an artifact of the choice of model and does not reflect a market opportunity or a forecastable featureof the data.

+3

+3

Table 15a Moments and persistence level of information arrival (NPCof INTEL).

Jan/03/94 3.010 7.208 Jan/04/94 3.123 6.869 Jan/05/94 3.656 10.776 Jan/06/94 3.203 8.650 Jan/07/94 4.651 18.983 Jan/10/94 4.472 10.111 Jan/11/94 3.210 9.370 Jan/12/94 5.382 38.401 Jan/13/94 3.092 7.565 Jan/14/94 4.028 12.156 Jan/17/94 3.510 10.744 Jan/18/94 4.818 12.967 Jan/19/94 17.467 144.357 Jan/20/94 6.785 60.792 Jan/21/94 4.954 19.982 Jan/24/94 4.123 11.383 Jan/25/94 2.908 6.377 Jan/26/94 3.933 12.448 Jan/27/94 3.385 7.733 Jan/28/94 2.777 6.148 Jan/31/94 2.900 7.530

3.338 3.197 1.417e -001 1.665e -004 1.0938e -002 3.094 3.093 8.258e -004 8.841e -002 2.9098e -004 3.666 3.666 1.049e -005 9.184e -002 2.7278e -005 3.945 3.391 5.539e -001 3.672e -002 1.0423e -002 6.022 4.888 1.134e 000 5.615e -002 1.1491e -002 4.490 4.474 1.599e -002 2.753e -001 8.9405e -007 4.088 4.019 6.868e -002 1.844e -004 1.6285e -001 5.302 5.291 1.093e -002 1.464e -005 1.5651e -003 3.528 3.233 2.950e -001 2.212e -002 6.1494e -003 3.989 3.984 5.078e -003 5.926e -004 4.8594~3 -004 3.492 3.492 1.188e -004 3.322e -001 9.2784e -005 4.809 4.803 5.880e -003 1.941e -001 4.6846e -005 000 329.588 13.673 3.159e 002 8.409e -002 1.0528e 15.638 6.729 8.909~3 000 3.041e -001 4.6604e -004 4.915 4.915 7.154e -006 3.686e -002 3.0946e -004 4.118 4.116 2.119e -003 2.328e -004 1.1905e -005 2.890 2.889 7.848e -004 3.398e -001 1.2496e -004 4.617 4.189 4.280e -001 4.561e -002 1.5645e -002 3.383 3.378 5.064e -003 1.579e -001 1.4506e -005 3.065 2.993 7.227e -002 1.683e -002 1.5588e -002 2.973 2.857 1.163e -001 2.698e -002 6.4718e -004

+

+ +

+

TCF 1.2899e

E = Mean, V = Variance,

E = Estimated Mean,

CF = Criterion Function = IE(NE)

h

= Estimated Variance,

-E ( ~ p c ) TCF ] ~ ,= ~~~1 C F ~ E(NPC)

+ 000

5 = /3(i;& + G;)

~ o d e l l i n gIntru-day Equity Prices and Volatility

55

Table 15b Moments and persistence levelof information arrival (NPC of IBM).

Jan/03/94 Jan/04/94 Jan/05/94 Jan/06/94 Jan/07/94 Jan/10/94 Jan/11/94 Jan/12/94 Jan/13/94 Jan/14/94 Jan/17/94 Jan/18/94 Jan/19/94 Jan/20/94 Jan/21/94 Jan/24/94 Jan/25/94 Jan/26/94 Jan/27/94 Jan/28/94 Jan/31/94

0.391 0.253 0.209 0.328 0.247 0.303 0.279 0.252 0.200 0.190 0.319 0.349 0.272 0.331 0.234 0.607 0.848 0.420 0.241 0.285 0.315

0.553 0.339 0.290 0.387 0.310 0.361 0.361 0.292 0.227 0.236 0.388 0.495 0.359 0.371 0.309 0.971 1.144 0.523 0.286 0.328 0.402

0.395 0.265 0.202 0.333 0.249 0.302 0.290 0.256 0.208 0.201 0.314 0.356 0.277 0.335 0.225 0,595 0.874 0.424 0.239 0.283 0.313

0.390 0.261 0.202 0.332 0.248 0.301 0.288 0.256 0.207 0.198 0.313 0.349 0.276 0.334 0.225 0.595 0.854 0.411 0.238 0.280 0.313

5.042e -003 0.666 2.5641e -006 3.802e -0030.6412.4521e -004 -004 5.748e -0050.5902.4257e -005 1.339e -0030.6624.8193e -006 1.15063 -003 0.692 4.0323e -005 6.543e -0040.6501.3289e 1.458e -0030.7102.8125e -004 2.260e -004 0.662 6.2500e -005 1.168e -003 0.719 2.3671e -004 2.424e -003 0.653 3.2323e -004 -004 7.578e -0040.5991.1502e 6.080e -003 0.566 0.OOOOe 000 6.334e -0040.6005.7971e -005 2.980e -004 0.667 2.6946e -005 1.500e -0040.6583.6000e -004 2.023e -0060.6592.4202e -004 2.008e -002 0.686 4.2155e -005 1.300e -0020.5121.9708e -004 1.233e -003 0.707 3.7815e -005 -005 3.275e -0030.6148.9286e 2.361e -0040.6881.2780e -005

+

TCF 2.6406e -003

Table 16 Estimation results of GARCH(l,l)t. INTEL

Date Jan/03/94 0.13962 Jan/04/94 0.11542 Jan/05/94 Jan/06/94 0.03984 Jan/O7/94 0.13667 Jan/10/94

h

a!

0.09664 (3.28851) (6.86519) (3.05787) 0.15985 (6.31624) (0.62048) (3.15266)

B

h

0.92212 0.82548 (11.36881) 0.59413 (24.56781) 0.47626 (14.25865) 0.68766 (32.01800) -0.17610 (-3.08832) 0.36975 0.23308 (7.25938)

IEM

G+j

0.73375 0.59168 0.84751 -0.13626

h

a !

0.3665 (4.7914) 0.2045 (5.1389) 0.1017 (3.6133) 0.1506 (5.0494) 0.0572 (8.4115) 0.0188 (0.4940)

B

A

0.4917 0.1252 (3.8129)j 0.6867 0.4822 (28.8343) 0.6124 0.5107 (30.6564) 0.7111 0.5605 (27.5990) 0.9455 0.8883 (188.5023) 0.5126 0.4938 (1.6193)

G+P

Jan/11/94 Jan/12/94 jan/13/94 Ja;li/14/94 Jan/17/94 Jan/18/94 Jan/19/94 Jan/20/94 Jan/21/94 Jan/24/94 Jan/25/94 Jan/26/94 Jan/27/94 Jan/28/94 Jan/31/94

0.12756 (4.28403) 0.34599 (7.54474) 0.02529 (1.05012) 0.31604 (3.55535) 0.32243 (7.88337) 0.13544 (5.09211) 0.62457 (22.39771) 0.10307 (7.73352) -0.02205 (-0.94791) 0.07347 (9.96571) 0.42942 (37.31836) 0.14975 (7.07011) 0.05535 (0.99949) 0.15494 (3.81609) 0.03917 (0.75202)

0.65105 (26.90697) 0.08378 (1.80773) -0.91167 (-12.03280) -0.11722 (-1.52218) 0.10302 (2.12959) 0.50834 (28.72999) 0.33569 (17.95127) 0.78258 (80.69002) -0.54658 (-7.26724) 0.74165 (194.19238) 0.56253 (69.63233) 0.72163 (44.24845) -0.04139 (-0.52909) 0.77262 (10.94175) 0.32467 (5.13231)

+

+

0.77861 0.42977 -0.88638 0.19882 0.42545 0.64378 0.96026 0.88565 -0.56863 0.81512 0.99195 0.87138 0.01396 0.92756 0.36384

0.0848 (7.5351) 0.0758 (4.1435) 0.1824 (5.7604) 0.1104 (13.1450) 0.2475 (8.9954) 0.3266 (5.7305) 0.0723 (1.4920) 0.1321 (3.6792) 0.0974 (1.6897) 0.2565 (21.5278) 0.1222 (27.1364) 0.1611 (6.7358) 0.1889 (6.6651) 0.2110 (3.1765) 0.0673 (2.5352)

0.8245 (116.0896) 0.7035 (52.5527) 0.5146 (29.6367) 0.8378 (194.2958) 0.4167 (17.9314) 0.4013 (17.8341) -0.1454 (-0.4893) 0.3700 (12.8733) -0.1417 (-2.8914) 0.7196 (266.0425) 0.8393 (209.7340) 0.6434 (37.7690) 0.5634 (32.6528) 0.1839 (5,1117) -0.7187 (-6.6678)

0.9093 0.7793 0.6970 0.9481 0.6642 0.7279 -0.0731 0.5021 -0.0442 0.9760 0.9616

0.8045 0.7523 0.3949 -0.6514

'GARCH(1,l) model: a : = a0 GE:-~ ,&CF~-~ bJumbers in the parentheses are t-values h

This chapter has had three objectives. They were (1) to compare different proxies for informational variables in high frequency equity data; (2) to (3) to model dynamic processes using doubly stochastic Poisson models; investigate intra-day volatility persistence. We find that the number of trades and the number of price changes seem to be the best choices for formational variables, volume being decidedly inferior. Secondly, we find that our model does seem to be estimable without undue difficulty

~ o d e l l i l~tra-day ~g Equity Prices and V~l~tility

and finally we find that persistence in volatility is much reduc our model is used rather than ~ ~ ~ ~ H ( The 1 , use l ) .of infor variables seems tosubstantially eliminate much of the ~ersiste persistence in volatility isa stylized fact that seems so terms of theoretical explanations,the results lead towar m d understand in^ of intra-day volatility.

ESTS

Our equation can bewritten as:

where ~ ( t )conditional , on N(t), is N(0, a2+ a ~ A N ( t ) )Let . Y be a (T X matrix Y’ =

[

],y’= [p, p Q j , /= ~ 102, ~ ~ j . ~ ~ e n

*‘*’

ANI, AN2, ANT the likelihood function L can bewritten as: ‘ 7

* a - ,

L = L(y, B)=

1 2

“

1.

where ht = P’Y;, Yt is the t-th row of Y, et = -Yfr, an term is ornitted. It foliows that the first derivatives of the function are:

1.3

The secondderivatives follow i m e

1.

58 S h i n n - ~ ~Lin h et al.

Given the above information, constru~tedas:

the Fisher’s formation matrix can be

1.5

~e now calculate

(

= El-lj -,j = 1,2,3where0’, = ’’E ~

~

f

)

p’], that is,

1.6

The score test forthe hypothesis H’, 1 which will be asymptotically distributed as ~ ~ (for 1 j)= 1 , 2 and x2(2)for j = 3. In turn, from Equation1.3:

1.7

where is equal to

P‘Yt (evaluated under Ho).

1.8

1.9

~odellingIntra-day Equity Prices and Volatility 59

OMENT ~ E N E ~ A T FUNCTION IN~ ANI)

AN(t)lh(t) Poisson (h(t)) E[exp(s AN(t))jh(t)]= mgf of a Poisson Process

2.1

A Poisson distributed random variablex with densityf u n c t i o ~ ~ ( x ) Axe-’ f(x) =,x =0,1,2, . * *

X!

2.2

has the following moment generating function:

= e-’ exp (heS )

2.3

= exp(-h(1 -e’)) = exp(h(es -1)) Therefore, E[exp(s MV(t))lh(t)]= exp(h(t)(eS -1)). Now, a

Po2 h(t) = 1-8

00

j=O

2.4

1

1

E[e~p(~aei~~(l))] 2.6 t=l

(S

A ~ ( t ) )= ] E[exp(h(t)(es -l))]

pa2

+a

E[h(t)] = 1-8

2.9

2.10

2.11

~ o d e l Z i ln~~t r f f - d aEquity y Prices and Vol~ti~ity 61

ai(-1)(1

-2qaSi)-2(-2a~)lq=o

2.12

- 2a2

"

1 -o2

Therefore, €[AN(t)] =

Bo2

+a

~

1-6 pu2 + a

2.13

+

2a2 Var[AN(t)] = 1-0 1-82

From our model, COV(AN(t),AN(t -1)) = Elct-l)[COv(AN(t)ll(t-l)), Cov(AN(~ -l)ll(t -l))]

+ CovI(t-l)(E1AN(t>Il(t-?l, E[AN(t -1)lW -l)]) = COVQt-l) (av2(t -1)+ pa2 + 6h(t -l),h(t -1)) = @Covl(t-l,(h(t-l),h(t -1))= OVar~(~-l~(h(t -1)) - 26a2 -

1-o2

3.1

Then we derive the following recursively, Cov(N(t),N(t -S ) ) =

2@b2

3.2

1-o2

~

Finally, it follows that Corr(N(t),N(t -S ) ) =

N(t -S ) ) COV(~(t), Var(N(t)) 2a2

(1,

+

+

2a20S (Bo2 a ) ( l +6)' ' .

ifs = 1 , 2 , . .. ifs=O

3.3

62 S ~ i n n - Lin ~ u et ~ al.

For the momentsof X(f),

it follows that

The correlation between X(t)and X(t -S) can be recursively derived by r d show that using Cov(X(t),X(t -1)).It iss ~ a i g h t f o r ~ ato

Cov(X(t), X(t-1))

(

+

= cov CTZl(t)

+

Qi, av(t -1)

3.6

=; p COVI(t-l) ( h N ( t ) ,AN(f -1))=; p cov (h@), h(t -1)) 2a20 1 -O2

= pQ-

Then we derive the following recursively, Cov(X(t),X(t -S)) =

2p;a20S ~

1 -O2

3.7

~odelling Intra-day Equity Prices and Volafility 63

Finally, it follows that Corr(X(t),X(t -S ) ) =

Cov(X(t),X(t -S ) ) Var(W)>

3.8

OTES

1. See also Karpoff (1987) for a survey of previous studies on the price-trading volume relationship. 2. This is probably due to the availabilityof data. A s pointed outby Jones,Kaul and Lipson (1994), although trading volume for the NASDAQ securities have been available for many years, historical data on the number of transactions were not available until recent years. 3. To avoid getting negative variance estimates, we use NLS to obtain 2, and $Q. S2,and are then derivedby &-method. 4. A GARCH-type model would involve interpreting$ ( t -1) as Var(X(t-l)iI(t -2)) x Az2(t) for z(t) i.i.d. N(O,1).This complicates the model without adding its to explanatory power. We shall refer to this asGARCH-typeeffects, although the model in Equation 9is closer to astochastic volatility model. 5. We have assumed thatA N ( t )is stationary, an alternative expressioncan be calculated if we startfrom a fixed starting point.

$6

-

IBLIOGRAPHY

Andersen, T.G. (1996), "Return Volatility and Trading Volume: An Information Flow Interpretationof Stochastic Volatility",~ o ~ r n aofZ Finunce,51,169-204. Andersen,T.G.and Bollerslev, T. (1997), "IntradayPeriodicityand Volatility Persistence in Financial Markets",~ o ~ r nofa 2E ~ ~ i r i c uFinance, Z 4,115 -158. Bachelier, L. (1900), Theorie dela S ~ e c ~ ~Paris: a ~ i oGauthier-Villars. ~, Bollerslev, T. and Domowitz,I. (1993), "Trading Patterns and Prices in the Interi ~ u ~ ce 1443. , bank Foreign Exchange Market",~ o ~ r n a Z o f ~48,1421 Breusch, T. and Pagan, A. (1980), "The Lagrange Multiplier Test and its Applications to Model Specification in Economics", Review of E c o ~ o ~ i c S ~47, ~d~es, 239 -253. Brock, W., Dechert, W.and Sheinkman, J. (1986), "A Test for Independence Based on the Correlation Dimension", mimeo, University of Wisconsin, Madison. Clark, P. (1973), "A Subordinated Stochastic Process Model With FiniteVariance 41,135- 155. for Speculative Prices", Econo~e~rica, Engle, R.F. and Russell, J.R. (1994), "Forecasting Transaction Rates: The Autoregressive ConditionalDuration Model", UCSD 94-27.

Engle, R.F.(1996), ”The Econometricsof Ultra-High Frequency Data”, NBER 5 Gallant, A., Rossi, P. and Tauchen, G. (1992), ”Stock Prices and Volume”, re vie^ of Financial Studies, 5,198-242. Geman, H. and Ant?, T. (1996), ”Stochastic Subordination”,Risk, 9,145-149. Goodhart,C.A.andFigliuoli,E. (1991), ”EveryMinute Counts in Financial Markets”, ~our#al o~~nter#ationaZon^ and ~inance,10,23-52, Goodhart,C.A.E.and OHara, M. (1997), ”HighFrequencyData in Financial Markets: Issues and Applications”, ~ournaZof E ~ p i ~ i c Finance, aZ 4,73-114. Gourieroux, C., Holly, A. and Monfort, A. (1982), ”Likelihood Ratio Test, Wald Test, and Kuhn-Tucker Test in Linear with InequalityConstraintson the Regression Parameters”,Econo~etrica,50,63 -80. Gourieroux, C.,Jasiak, J. andLeFol, G. (1996), ”Intra-day Market Activity”,SEE, No. 9633. Harris, L. (1987), ”Transaction DataTests of the Mixture of Dis~ibutionsHypothve 22,127- 141. esis”, ~ournalof ~inancialand ~ ~ a n t i t a t iAnalysis, Jones, C., Kaul, G. and Lipson, M. (1994), ”Transactions, Volume and Volatility”, R e ~ i ofe ~inancial ~ Studies, 7,631-651. Judge, G., Hill, C., Griffiths, W., Lutkepohl, H.and Lee, T-C (1985), The Theory of E c o ~ o ~ e t r i cJohn s , Wiley & Sons. Karpoff, J. (1987), ”The Relation Between Price Changes and Trading Volume:A Survey”, ~ournaZ of ~ina#cialand ~uantitativeAnalysis, 22,109- 126. Lamoureux, C. and Lastrapes, W. (1990), “Heteroskedasticity in StockReturn Data: Volumeversus GARCH Effects”,~ournalof Fi#ance, 45,7-38. Lamoureux,C.and Lastrapes, W. (1994), ”EndogenousTradingVolumeand Momentum in Stock-Return Volatility”,~ournalof ~usiness and Econo~ic Sta~istics, 12,253-260. Locke, P.R. and Sayers, C.L. (1993),”Intra-day FuturesPrice Volatility:~ f o r ~ a t i o n Effects and Variance Persistence”, ~~ur#aZ of AppZied Econo~etrics,8,15-30. Madan, D. and Chang, E. (1997), ”The Variance Gamma Option Pricing Model”, mimeo. Marsh, T. and Rock, K.(1986), ”The TransactionalProcess andRational Stock Price Dynamics”, University of California at Berkeley. Oldfield, G., Rogalski, R. and Jarrow, R. (1977), ”AnAutoregressive JumpProcess ~ournalof F~nancia~ Econo~ics, 5,389-418. for Common Stock Returns”, er at ions R e s e a ~ c ~ , Osborne, M. (1959), ”Browning Motion and the Stock Market”, 7,145-173. Rogers, A. (1986), ”Modified Lagrange Multiplier Tests for Problems with One31,341 -361. Sided Alternatives”,~our#alof Econo~e~rics, Smaby, T. (1995), ”AnExamination of the Intraday Behavior of the Yen/Dollar Exchange Rate: The Relationshipbetween Trading Activity and Returns Volatility”, ~ o ~ r n of a ZEcono~icsand Finance, 19/39-50. Takezawa,N. (1995), ”Note on Intraday ForeignExchange Volatility and the 48,399-404. ~ o r m a t i o n aRole l of Quote Arrivals”,Econo~ics ~etters, Zhou, B. (1996), ”High-Frequency Dataand Volatilityin F o r e i ~ - E x c ~ a nRates”, ge and ~ c o n o ~Statistics, ic 14 (l), 45-52. ~ o u ~ nofa~usiness l

The volatility of a spot exchange rate S c m be defined for many price models by the a ~ u a l i s e dstandard deviation of the change in the logarithm of S during some time interval. For a diffusion process defined by d(lnS) = pdf \3/(f)dW, with*(f) a deterministic function of time and W(t) a standard Wienerprocess, the deterministic volatilityo(0, T) from time 0 until timeT is definedby

+

Options traders make predictions of volatility for several values of cf. These forecast horizons typically vary between a fortnight and a year and are defined by the times until expiration of the options traded. Insights into these predictions can be obtained by inverting an option pricing formula to produce implied volatility numbers for various values of T. Xu and Taylor (1994) show that these volatility expectations vary s i ~ ~ c a n t l y for exchange rates, both across expiry times 71 and through time. Options markets are often considered be markets to for trading volatility. It then follows that implied volatilities are likely to be good predictors of subsequent observed volatility if the options market is efficient. As options traders have more information than the historic recordof asset prices it may also be expected that implied volatilities are better predictors than forecasts calculated from recent prices using ARCH models. Day and Lewis (1992) investigate the information content of implied volatilities, calculated from call options on theS&P 100 index, within an 0 1997 Elsevier Science. of € ~ z ~ i r i c~~ iZ n a ~ 4c, epp , 317-340. Printed with. permission from ~o~rnaZ

66 S ~ ~ hJ. e Taylor n and Xinzhong Xu

ARCH framework. They conclude that recent stock index levels contain incremental volatility information beyond that revealed by options prices. Lamoureux and Lastrapes (1993) report a similar conclusion for individ US stocks. Xu and Taylor(1995), however, usedaily data to conclude that exchange rates do not contain incremental volatility information: implie volatility predictions cannot be improvedby mixing them with conditional variances calculated from recent exchange rates alone. Jorion(1995) also finds that daily currency implieds are good predictors. The superior efficiencyof currency implieds relative to implieds calculated from spot equity indices has at leasttwo credible explanations. First, there is the theoretical argumentof Canina and Figlewski(1993) that efficiency willbe enhanced when fast low-cost arbitrage trading is possible. S&P 100 indexarbitrage,unlikeforexarbitrage,isexpensivebecause many stocks must be traded. Second, as Jorion (1995) observes, index option implieds can suffer from substantial measurement error becaus of the presenceof some stale quotes in the index. This paper extends the studyof Xu and Taylor (1995), hereafter XT, by using high-frequency exchange rates to extract more volatility inform from the historical record of exchange rates. From probability theory it is known that it may be possibletosubstantiallyimprovevolatility estimates by using very frequent observations. Nelson(1992) shows that it is theoretically possible for volatility estimates tobe made as accurate as required for many diffusion models by using ARCH estimates and sufficiently frequent price measurements. As trading is not continuous and bid/ask spreadsexist,thereare, of course,limitstothebenefits obtainable from high-frequency data. The definition of implied volatility and the low-frequency results of XT are reviewed on pp 67 and 68. Our estimatesof ~eutschemark/dollar volatility obtained from the high-frequency dataset of Olsen & Associates are described on p 70. The results from estimating ARCH models when the conditional variance is a function of implied volatilitiesand/or highfrequency volatility estimates are presented on p 75. Further evidence about the incremental information content of options prices and the O&A quotations database is providedby evaluating the accuracy of volatility forecasts, see p86. Our conclusions are summarised on p 89.

IMPLIED VOLATILITY

The implied volatilities usedin this paper are calculated from the prices of nearest-the-money options on spot currency. These options are traded

~ ~ ~ r ~ Volatility e n t a l~ n ~ o r ~ a t i67 on

atthePhiladelphiastockexchange(PHLX).Standardoptionpricing formulaeassumethespotratefollowsageometricBrownianmotion process. The appropriate European pricing formula for thec price of a call option is then a well-known functionof the present spot rateS, the time until expiration T,the exercise priceX, the domestic and foreign interest rates,respectively r and q andthevolatility 0 (seeforexampleHull, 1995). The Philadelphia options canbe exercised early and consequently the accurate approximate formulaof Barone-Adesi and Whaley (1987) is used to define the price C of an American call option. This price can be written as c+e, *, S-x, SS 2< SS*,

.=(

with e the early exercise premium and S* the critical spot rate above which the option shouldbe exercised immediately. The implied volatility is the number 01that equatesan observed marketprice CM with the theoretical price C: CM = c(s, T , r, 4, 01).

x,

There will be a unique solution to this equation when CM > S -X. As K/&T>0 when S S*(5), thesolutioncan be foundveryquickly by an interval subdivision algorithm. Similar methods apply to put options. A typical matrix of currency implied volatilities calculated for various combinations of time-to-expiry T and exercise price X will display term structure effects as T varies for a fixed X near to the present spot price. These effects have been modelled by assuming mean reversion in implied volatilities (Xu and Taylor, 1994). Matrices of implieds also display smile effects as X varies for fixedT (Taylor and Xu, 1994). Tradersknowthatvolatility is stochastic,neverthelesstheymake frequentuse of impliedestimatesobtainedfrompricingmodelsthat assume a constant volatility. The implied volatility canbe interpreted as a volatility forecast if we follow the analysis of Hull and White (1987) and make three assumptions: first that the price S ( t ) and the stochastic is not volatility o(t) follow diffusion processes, second that volatility risk priced and third that spot price and volatility differentials are uncorrelated. The first and second assumptions are pragmatic and theisthird consistent with the empirical estimates reported by XT. With these assumptions, let FTbe the average variance (l/T) :J a2(t)dt. Also, let c(02) represent the Black-Scholes, European valuation function for a constant level of (1987) show the fair European call price volatility, 0.Then Hull and White is the expectation € [ c ( ' l f ~ ) ] ,which is approximately c(E[VT])when X is near S. Thus the theory can supportbelief a that the implied volatility for time-to-expiry T is approximately the square root of €[TT]. Traders might

S t ~ h 1~.TayZor n and Xinzhong Xu

obtain efficient prices if they forecast the average variance and then insert its square rootinto a pricing formula that assumes constant volatility.

The evidence for incremental volatility information can be assessed by makingcomparisonsbetweenthe m a x ~ m ~ m ~ i k ~attained l~o~d bys different volatility models. An ARCH model for returns Rt based upon information setsQ will specify aset of conditional variancesht and hence conditional distributions &/;IZt-.l, from which the l i ~ e l ~ o oofdobserved returns canbe calculated. We consider informati~nsets If,Jt and Kt respectively defined by (a) all returns up to time t, (b)implied volatilities up to time t and (c) the union of these two sets, We say that an orm mat ion source has incremental information if it incredses the l o g - ~ k e l ~ o oofd observed returnsby a statistically significant amourit. The following maximum log-li lihoods are reportedby Xu and Taylor (1995, Table 3) for a model defined below, for five years (1985-1989) of daily DM/$ returns from futures contracts:

4

4327.31

It

4349.64

Kt = It

+It

4349.65

Source It has incremental information because its addition to It adds 22 to the log-likelihood with only one extra parameter includedin the ARCH model. This is significant at very low levels. Source 4 , however, does not contain incrementalidormation because its addition toIt only adds 0.01 to the log-likel~ood.Thus, in this low-frequency example, there is only incremental informationin options prices. The models estimated in XT use daily conditional variances ht that reflect higher levelsof volatility for ond day and holiday returns. These seasonal effects are modelled by multiplicative seasonal parameters, respectively denoted by M and H . The quantityh: represents the conditional variance with seasonality removed:it is definedby ht if period t ends 24 hours after period t -1, h t / ~if t fallsonaMondayand t -1onaFriday,1 ht/H if a holiday occurs between the two prices.

A general specification forh; that incorporates information at timet -1 about daily returns &-l, implied volat~itiesit-1 and their lagged values

is givenby

The quantity it-1 here denotes the implied volatility for the nearest-themoney call option, for the shortest maturity with more than nine calendar days to expiration. Althoughit-1 is an expectation for a period of at least ten days it is used as a proxy for the market's expectation for the single trading period t. The standard deviation measure it-1 is an annualised quantity. It is converted to a variance for a 24"our return in the above equation by assumin there are 48 Mondays, 8 holidays and 196 normal weekdays in a year. An appro~riateconditional distribution for daily returns from DM/$ futures is the generalised error distribution (GED) that has a single shape parameter, called the thickness parameter U . The parameter vector for the general specification is then8 = (a, b, c, d, M, H,U ) . All conditional means are supposed tobe zero. is obtainedby assuming The maximum likelihood for information 1, sets d = 0 followed by maximisatio~of the log-likelihood over the remaining parameters. This gives:

The estimate of U has a standard error less than 0.1 and therefore fattailedconditionaldistributionsdescribereturnsmoreaccuratelythan conditional normal distributions ( U = 2), as has been shown in many otherstudies of dailyexchangerates.Theestimates of M and H are more than one but their standard errors, respectively 0.12 and 0.33, are substantial. The maximum like~hoodfor options information Jt is obtained when aand b areconstrained to be zeroand all theotherparametersare con strained. MLE then givesc = 0 and: The incremental ~ p o r t a n c eof previous returns and options orm mation is assessed by e s ~ a t i n the g general specification without parameter constraints. The MLE estimates of a and c are zero, with b estimated as 0.04 (t-ratio 1.43) and d as 0.93 (t-ratio 3.11). Any conventional statistical tests accept the null hypotheses a = 0, b = 0, c = 0 and d = 1. They also . reject d == 0 at very low significancelevels,

'70 S t ~ k e I. n Taylor and Xinzhong Xu

XTconcludethat all therelevantinformationfordefiningthenext period's conditionalvarianceiscontained in themostrecentimplied volatility. This conclusion holds despite using a volatility expectation for at least a ten-day period as a proxy for the options market's expectation for the next trading period. XT also present results for volatility expectation forthenextdaycalculatedfromatermstructuremodelforimplieds studied in Xu and Taylor (1994). These expectations are extrapolations (T = 1day)fromseveralimplieds (T 10 days).Suchextrapolations provide both the same conclusions as short-ma~rityimplieds and very similar maximum levels of the log-likelihood function, However, these extrapolations are biased. Out-of-sample forecastsof realised volatility during four-week periods in1990and1991confirmthesuperiority of theoptionspredictions compared with standard ARCH predictions based upon previous returns alone.

\JOLATILI'IY ESTI~ATES AND E X P E ~ T ~ T I O N S

Estimates of Deutschemark/dollar volatility have been obtained from the dataset of spot DM/$ quotations collected and distributedby Olsen & Associates.Thedatasetcontainsmorethan1,400,000quotationson the interbank Reuters network between Thursday 1 October 1992 and Thursday 30 September 1993 inclusive. It is our understanding that the dataset is an a h o s t complete recordof spot DM/$ quotations shown on using GMT. We Reuters FXFX page. The quotations are time stamped converted all times to US eastern time which required different clock adjustments for winter and summer. Volatility estimates have been calculated for 24-hour weekday periods for comparison with daily observations of implied volatilities. The options market at theP~iladelphiastock exchange closesat 14.30 US eastern time, whichis 19.30GMTin thewinterand18.30 GMTinthesummer. A 24-hour estimate for a winter Tuesday is calculated from quotations made between 19.30GMT on Monday until 19.30 GMT on Tuesday. We follow GMT Andersen and Bollerslev (1997) and ignore the 48 hours from 21.00 on Friday until 21.00 GMT on Sunday, because less than 0.1 percentof the quotations are made in this weekendperiod, Thus a 24-hour estimate for a winter Monday uses quotations from 19.30 to 21.00 GMT on the previous Friday and from 21.00 GMT on Sunday until 19.30 GMT on Monday.

l n ~ r e ~ e n t~aoi l a t i ilnfor~tion i~ 71

~ e ~ i t i and o n ~otivation of the Estimates

The realised volatility for day t is calculated from intra-day returns R,,i with i counting short periods during day t, in the following way: 5

Here m is a multiplicati~econstant that converts the variance for one trading day into an annual variance and ut is an annualised measure of realised volatility. The number of short periods in one trading day is chosen tobe n = 288 corresponding to five-minute returns. We follow the methods of Andersen and Bollerslev (1997), hereafter AB, when five-minute returns are calculated. Their methods use averages of bid and ask quotations to define rates. They define the rate at any required time by a linear interpolation formula that usestwo quotations that immediately precede and follow the required time. As in AB, suspect quotations are filtered out using the methods of Dacorogna et al. (1993). AB note that there isvery little autocorrelationixi the five-minute returns: the first-lag coefficient is -0.04. Negative dependencehas previously been documentedby Goodhart and Figliuoli(1991). Somemotivationfortheabovemethod of volatilityestimationis provided by supposing that spot exchange rates S(z) develop in calendar time t according to a diffusion processdescfibed by d ( h S(t)) = pdz

+s ( z ) C ( t ) d ~ ( ~ )

6

with ~ ( an t ) annualised stochastic quantity and ~ ( za)deterministic quantity that reflects the strong intra-day seasonal pattern in ‘volatility. This pattern has been investigatedin detail by AB and has been described in earlier studies that include Bollerslev and Domowitz (1993)and Dacorogna ._ et al. (1993).The squareof thg’seasonalmu1 s ( t ) averages one overa completeseasonalcycle, so if and z2 denidenticalpositioninthe cycle thens(q) = s(t2) and Jr: s2(t)dt = z2 -XI..When the volatility is constant during a oncl-day cycle, o years, and the multipliers are Constants Si during intra-dayint ~”*,_

,- , ....’

var(Rt,ila(t)) = h--

s;C2(t) = ho2(t),

7

i=l

with z the calendar time associated with trading period t. The quantity v: is the estimate of o2(z)obtained by setting m = 1/h and using R$ to estimate the above conditional variance of R k j . We set m = 260 which is

appropriate whenit can be assumed that there is no volatility durin weekend and a year contains exactly 52 weeks. The estimate ztt will not be the optimal estimateof O(Z) when volatility isconstantwithincycles.However,theestimate is consistent (vf. -+ a(z)as n -+ cm) and it does not require estimation of intra-day seasonal volatility terms. st~matesfrom Intra-

Figure 1 isatime-seriesplot of thevolatility es~imatesut forthe253 days that the PHLX was open between October 1, 1992 and September 30,1993inclusive.Theaverage of theseestimates is 12.5percentand their standard deviation is 3.6 percent. Further descriptive statistics are U.S. presented in Table 1. The estimates have also been calculated for holidays and are smaller numbers as should be expected. The two extreme holiday estimates are 2.5 percent on Christmas Day and 1.9 percent on New Year’s Day; the other six holiday estimates range from 6.7 to 10.5 percent, The estimates are higher in October 1992 thanin any other month, with the two highest estimates, 32 and 26 percent, respectively, calculated for Friday 2nd and Monday 5th October. The October average is 19.3 perce compared with 14.4 percent for November and 11.7 percent for the other ten months. The difference maybe associated with events that followed the departureof Sterling from theEMS in September 1992.

5 0 Oct

Nov

Dec Feb Jan Mar May Apr June July Aug Sep

Volatility estimates from intra-day quotations,

lncre~ental~ o l a t i l il tn~f o ~ a t i o n 73 '

Table 1 Summary statistics for volatility estimates v calculated from intra-day price quotations and impliedvolatilities i calculated fromoptions prices. l~tra-day es~imates Oct./Sept.

v

Dec./Sept.

l ~ p l i e dv~lat~lities i

Oct./Sept.

Dec./Sept.

253 12.53 3.57

211 11.66 2.61

253 13.57 2.68

211 12.64 1.27

5.83 10.30 11.86 13.90 32.05

5.83 9.76 11.26 12.81 20.32

9.71 11.92 12.82 13.90 24.24

9.71 11.81 12.45 13.36 16.69

11.43 11.96 12.03 13.14 14.12 0.001

10.42 11.19 11.31 12.26 13.09 0.000

13.74 13.68 13.53 13.53 13.36 0,966

12.67 12.72 12.73 12.66 12.42 0.816

Autocorrelation Lag 1 Lag 2 Lag 3 Lag 4 Lag 5

0.628 0.444 0.392 0.382 0.382

0.386 0.042 0.077 0.038 0.120

0.914 0.863 0.821 0.777 0.734

0.800 0.699 0.632 0.603 0.565

Partial autocorrelation Lag 2 Lag 3 Lag 4 Lag 5

0.083 0.140 0.123 0.109

0.184 0.150 0.119 0.172

0.169 0.067 0.001 -0.017

0.165 0.094 0.127 0.037

Sample size Mean Standard deviation ~ ~ u Lower quartile Median Upper quartile Maximum Monday mean Tuesday mean Wednesday mean Thursday mean Friday mean p-value, ANOVA

m

Summary statistics are calculated for the 12 months from October 1992 to September 1993

and for the 10 months commencing December 1992.

The estimates display a clear day-of-the-week effect. The average estimate increases monotonically as the week progresses, from 11.4 percent on Monday to14.1 percent on Friday. This pattern reflects the predominance of important scheduled macroeconomic announcements on Fridays and less important announcements on Thursdays. Parametric (ANOVA) and 0.2 percent for non-parametric (~ruskal- all is) tests have p-values below tests of the null hypothesis that the distributionof the estimates is identical for the five days of the week. Removing the high volatility months of October and November reduces the mean estimateby about 1.0 percent for each day but the monotonic pattern and the low p-values remain. The autocorrelations and partial autocorrelationsof the volatility estimates are similar to those expected from an AR(1) process. The first-lag

74

Stephen 1.Taylor and Xinzhong Xu

autocorrelation is 0.63 for all the estimates but it falls to 0.39 when October and November are excluded. ~ m ~ lVolatilities i e ~

Figure 2 is a time-series plotof implied volatility estimatesit for the same days as are used to produce Figure 1. Each estimate is the average of two implied volatilities, one calculated from a nearest-the-money (N'I") call option price and the other from a NTM put option price. The last options prices before the PHLX close at 14.30 local time are used. These are the only useful options prices supplied to us by the PHLX: high and low options prices are supplied but theydo not usually define high and low implied volatilities. The spot prices used for the calculations of the implieds are contemporaneous quotations supplied by the PHLX. On each day, the shortest maturity options with more than nine c days to expiration are selected. The time to maturity of the options is We onlyusetheestimates alwaysbetween 10 and45calendardays. it to represent options information about volatility expectations. We do not seek shorter-term expectations from the term structure of implieds because this involves extrapolations that produced no statistical benefits in Xu and Taylor (1995). The averageof the estimatesit is 13.6 percent, which is slightly more th the average of the intra-day estimates. Table 1 provides ~ o r ~ a t i for on comparisons of the distributions of the implied and intra-day estimates. 25

20

5

0' Oct

Nov

Dec

Jan

Feb Sep Aug Jul May Jun Apr Mar

Figure 2 Impliedvolatilities.

I

~ n c ~ E ~ eVolfftility n ~ f f 1 rnfo~ation 75

20 25

I

0

a

15

10 0

5

20

3525

30

Intra-day estimate

Figure 3 Comparison of implied volatilitiesand intra-day estimates.

Figure 2 shows that traders expected a higher levelof volatility in October and November and thereafter had expectations that werewithin an unusually narrow band. There are no day-of-the-week effects because the implieds are expectations for long periods that average 25 calendar days. The implied estimates it are markedly less variable than the realised estimates zlf again because the impliedsare a medium-term expectations measure. This also explains why the serial correlation in the implied volatilities is substantial:0.91 at a lag of one-day, using all the data and 0.80 when the first two months are excluded. Thecorrelationbetweentheimpliedvolatilities and the intra-day 0.66. These two volatility measurements are plotted volatility estimates is against each other on Figure 3.

ESTIMATED FROM INTRA-DAY QUOTATIONS

ODELS WITH vOLATILITY

Models and results are first discussed for daily returns and are subsequentlydiscussedfor hourly returns. Dailymodels are

76 StephenJ.Taylor and X i ~ z ~ oXu ng

straightforwardbecausetheyavoidestimation of intra-day,seasonal volatility patterns. Hourly models, however, are more incisive because of the much larger numberof observed returns. odel for Daily R e t ~ ~ ~ s

ARCHmodelsareestimatedfor daily spotreturns, Xt = In(St/St-l), obtained from rates when the PHLX closes. All the ARCH models are estimated using data for theofset PHLX trading days. Our of set253 daily returns is small. The results are unusual and only needbetodiscussed when the conditional distributionof returns is normal with mean zero and a conditional variance ht that depends on the information &-l, given by combining the information from options trades with the set of five-minute returns up to time t -1. The options information is summarised by the implied volatility term&-l. The volatility information provided by the five-minute returns is summarisedby the estimateut-1. The following model makes use of conditional varianceshl; appropriate for 24-hour periods after removing multiplicative Monday and holiday effects, definedby Equation 1: 8

9a

lob 1oc The parameter vector is 6 = (a, b, c, d, e, M, H).The terms and i:-l are into quantities appropriate divided byf to convert these annual quantities for a 24-hour period.

Table2presentsresultsforthegeneralmodelandsevenspecial cases.Whenthehistory of five-minutereturnscontains all relevant information about future volatility, the options parameter e is zero. An estimationwiththisconstraintproducesasurprise,whenthe initial

)

~ n c r e ~ e n tVa ol l ~ ~ i~nfor~lation lit~ 77

Table 2 Parameter estimates for daily ARCH models that include intra-day volatility estimates and short-ma~rityimplied volatilities.

c x 105 3.484 (3.55) 0.204 (1.95) 0.203 (1.94) 0.203 (1.89) 0.897 (0.64) 0.897 0.897 (0.64)

0.000 0.000 0.000 0.000

d

b

a

max. En(L) 868.66 871.05 871.05 871.05 870.95 873.04 873.04 873.04

0.329 (2.01) 0.000

0.956 (45.87) 0.956 (45.78) (44.38) 0.956 0.000 0.000

e

0.000

1.000

(2.88)

0.683 (2.88) 0.683 0.000 (2.88) 0.683

The numbers in parentheses t-statistics, are estimated using the Hessian matrix and numerical second derivatives. t-statistics arenot reported when an estimate is less than The 24-hour conditional variance hi is the product of the 24-hourdeseasonalisedconditional variance h: and a multiplier that is either 1, M (for Mondays) or H (for holidays). The deseasonalisedconditionalvariance is defined by h: = c lht-1) bhF-, d~21_~ei;"_,. The terms Rt-1, and Z i L1are, respectively,daily returns, intra-day volatilityestimatesand the squares of scaled impliedvolatilities. All parameters are constrained to be non-negative. In the fifth row,e is constrained to equal one."he estimates of M and H for the most general model are 1.44 and 1.74, standard errors 0.34 and 0.89, respectively.

+

+

+

+

value h: is an additionalparameter. As a = d = 0, theconditional variancesaredeterministic,hence if theunconditionalvarianceis p h = c / (1-b) then: hf = &h b ~ ( h~ph). 11

+

This result is less surprising when we recall the volatility estimates plotted on Figure 1. The twelve months begin with high volatility followed by a long period during which volatility does not change much. The above edge solution is unlikely to be estimated if the period of exceptionally high variance is anywhere other thanat the beginning of the sample. Ex post, the selectionof dates for the sample period is rather unfor~ate! The edge solution is a consequence of an unusual volatility pattern found in a small sample. Small samples can give more ordinary results, for examplea = 0.035 and b = 0.917 for GARCH(1,l) estimated from the daily DM/$rate from September1994 to August1995. Next, consider models that make use of the information in implied volatilities. The specification

has a maximum likelihood that is 1.99 above that of the edge solution. Estimation of the most general model simply produces the linear function of squared implied volatility above; the estimates of a, b and d are all zero.

78 S t e p h ~1. Taylor and Xinzhong Xu

The results are compatible with the hypothesis that there is no incremental volatility informationin the dataset of five-minute returns, when calculatingdailyconditionalvariances.However,thehypothesis that there isno incremental volatility information the in implied volatilities is dubious.

intra-^^^ Seasonal ~ ~ l t i ~ l i e r s We now multiplythe number of returns used to estimate models24.byThe much larger sample sizeprovides a reasonable prospect of avoiding the unsatisfactory edge solutions found for dailyreturns. Before estimating conditional variances for hourly returns we must, however, produce estimates of the intra-day seasonal volatilitypattern. We present simple estimates here. Our estimates ignore the effects of scheduled macroeconomic news announcements; we discuss the sensitivity of our conclusions to this omission on83. p Andersen and Bollerslev (199’7)provide different estimates basedupon smooth harmonicand polynomial functions. It may be helpful to review some notation before producing the seasonal estimates. The timet is an integer that counts weekdays,n is the number of five-minute returns in oneday (= 288) and Rt,i is a five-minute return; i = 1identifies the return from 14.30 to 14.35 US eastern time (ET) on the previous day (i.e. t -1) ...i = 288 is the return from 14.25 to 14.30ET on day t. Returns over 24 hours and over 1hour periods indexed by j are respectively given by 13 Sums of squared returns provide simple estimatesof price variabilityand averages across similar time periods can be usedto estimate the seasonal volatility pattern. Let N be the number of days in the sample. It would be convenient if the seasonal pattern could be described by 24 one-hour, sf = 24. A natural multiplicative, seasonal variance factors sf, with estimate of the variance multiplier for hour j is given by

x$

14 However, the seasonal pattern varies by day of the week, as might be expected from Table 1 and thus it appears preferable to estimate 120 multiplicative factorsthat average one over a complete week.

A second way to estimate variance multipliers takes accountof the day of the week. Let St be the set of all daily time indicesthat share the same day-of-the-week as time index t. Let Nt be the number of time indices to be found in St. Then a setof 120 factors are given by: 15 Figure 4 is aplot of standard deviation multipliers, 3,,j. The final hourly interval, j = 24, is the hour ending at 14.30 ET (19.30 GMT,winter) when the options market closes. The first interval,j = 1, is the hour b e ~ n i n g after the previous day’s options close. The multipliersare generally higher for intervals 13to 24, corresponding to 07.30 until 19.30 local time in London, with the highest levelsin i n t e ~ a l s 18 to23 when both US and European dealers are active. The Thursday and Friday spikes,at intervalj= 19, reflect the additionalvolatility when many US macroeconomic newsreports are released in the hour commencing at 08.30 ET.Ederington andLee (1993,1995)provide detailed documentation of this link with macroeconomic news. The lower local maximum, atj = 13, occurs whentrade accelerates in Europe in the hour commencing at 07.30 local timein London. The Mondayspike earlier in the day,at j = 6, is the start of a new weekin the Far East markets.

1 2345

6

7 8 9 10 11 12131415

16 17 18 19 20 21 22 2 3 2 4

Hourly interval

Figure 4 DM/$intra-day standard deviation multipliers.

An ARCH specification for hourly returns that is similar to that considered for the daily returns involves hourly returns rf,j, information sets Kf,i-1, recent five-minute returns R t , i , one-hour realised variances Vt,j, one-hour conditional variances h t ~one-hour , deseasonalised conditional variances and the multipliers$t,i. The specification alsoincorporatesthe annualised implied volatility it-1 calculated at the previous close; hourly implieds are not available to us,although we would not expect them to contributemuch becausethe implieds change slowly.The formation set Kt,j-1 is defined to be all relevant variables known at the end of hour j -1 on day t, namely the implieds it-2, ... ,the latest five-minute return Rt,12(j-l) and allprevious five-minute returns. The most general ARCH model that has been estimated for hourly returns is: rf,jIKt,i--1 mt,i

~ u ( ~ t ht,i), , j ,

= #rt,i-1,

16a 16b 16c 16d

12(j-l)

16e i=120"-2)+1

16f for some numberfthat does not need tobe estimated; wesetf equal to the number of annual hourly returns (24 x 252). The subscript pair ,C i refers to the timeinterval t -1,n -i whenever i is not positive. Thedistribution ~ u ( m t , jht,j) , is GED with thickness parameter U, mean ~ t , and i variance hf,i.The parameter vector is 8 = (a,b, c, d , e, ct,,U). As the autoregressive, mean parameter ct, is alwaysinsi~ificant,we only discussresults when ct, is constrained to be zero. Equation 16d contains terms, with coefficients a and d, that are both measures of hourly return variability. Both measures are included to permit comparisons of the information contentof five-m~uteand hourly returns.

Table 3 presents results for6049 hourly returns when 120seasonal volatility multipliers are included in themodels.Themaximum

lncremental V o l a ~ i l~nfor~ation i~ 81

Table 3 Parameter estimates for ARCH models of hourly returns with 120 seasonal terms. c x 105

b

a

Panel A:normal distribution 0.1321 (27.29) 0.0012 (2.35) 430.01960.0012 5.58) (3.24) 800.00450.0012 1.14) (2.44)

d

0.2875 (13.38) 0,0352 (6.49)

0.0319 (5.38)

0~0000

0.0000 0.29260.1046 (3.78) (6.98) 0.0000

0.1766 (1.70)

0.0000 0.18120.0713 (2.14) (4.51) Panel B: GET) distribution 0.1232 (19.12) 0.0008

0.9434

(1.25) 070.02270.0013 4.45) (2.38) 310.00280.0009 0.52) (1.31)

0.1437 (7.01) 0.0876 (4.22)

0.6528 (54.99) 0.6528 (54.88) 0.3935 (8.20) 0.4141 (6.76) 0.4127 (8.30)

0.3197 (10.68) 0.0408 (5.53)

0.0387 (4.65)

0~0000

0.0000 0.25900.1221 (2.66) (5.74) 0.0000

0.0000

e

0.0808 (3.64)

0.2212 (1.88) 0.1801 (183)

0.1678 (5.96) 0.1099 (3.87)

0.6482 (39.05) 0.6482 (38.98) 0.4034 (6.71) 0.3641 (5.38) 0.3887 (6.71)

V

max. ln(L)

2

31848.67

2

31963.41

2

31933.21

2

31964.09

2

31945.15

2

31945.15

2

31999.55

2

31997.85

2

32011.87

1.1025 (42.42) 1.1460 (41.90) 1.1376 (41.88) 1.1463 (41.89) 1.1418 (41.58) 1.1458 (41.58) 1.1598 (41.40) 1.1594 (41.51) 1.1638 (41.45)

32193.31 32253.21 32230.24 32253.36 32231.90 32231.90 32266.90 32268.01 32277.00

The numbers in parentheses are t-statistics,estimated using the hessian matrix and numerical second derivatives. All parameters are constrained to be non-negative. t-statistics are not reported when anesthate is less than Theone-hourconditionalvariance is definedby htj = $;jhtj, =c bh;j-l 12G-1) d(Ci=12+-2)+1 R;i)/';j-l+ eir-1. The terms rt,j-l, Rt,i and iTmlare, respectively,

+

+

+

one-hour returns, five-minute returns and the squares of scaled implied volatilities. The conditional distributionsare normal distributions in panel A and aregeneralisederror distributions, with thickness parameterv, in panelB.

82 S ~ ~ h 1. e Taylor n and Xinzhong Xu

log-likelihood increases substantially when 120 day-of-the-week multipliersreplace24hourlymultipliers,typically by about 65 for conditional normal distributions and by about 22 for conditional GED distributions. Consequently, our discussionof the results is based upon models with 120 intra-day seasonal multipliers. All our observations and conclusions are also supported by numbers in a further table, available upon request, for models that have only 24 seasonal multipliers. The lower panel of Table 3 shows that the conditional distribution of the hourly returnsis certainly fat-tailed. The GED thickness parameter is estimated tobe near 1.15 with a standard error less than 0.03. Conditional normal distributions are rejected for the most general specification and all the special cases. The log-likelihood ratio test statistic is 130.26 for x:. A thickness the general specification with the null distribution being parameter of 1 defines double negative-exponential distributions so the hourly returns have conditional distributions that are far more peaked and fat-tailed than the normal. Ourassumption of theGEDfortheconditionaldistributionsdoes notensureconsistentparameterestimatesandstandarderrors if the assumption is false. The quasi", estimates in the upper panel of Table 3 are consistent although they are not efficient. The resultsin the lower panelof Table 3 fall into three major categories, and are discussed separately. The conclusions are the same if we focus on the upper panel for normal distributions. First, consider models that only make useof returns orm mat ion. The models that incorporate information more than one-hour old, through parameter b, havesignificantparametersforbothrecentinformation (the last hour, through a and d ) and old i~ormation.This is the usual situation when ARCH models are estimated and so we no longer have the curious edge solutions discussed for the daily returnsat p 77. M e n f i v e - ~ ~ ureturns te are used, but hourly returns are not (a = 0; 6,d >0), the maximum of ln(L) is 23 more than the maximum when only hourly returns are used (d = 0; a, b >0). There is thus more relevant volatility i~ormationin fiveminute returns than in hourly returns.i~ormation This comes from more than twelve five-minute returns, as expected, because the maximumof ln(L) decreasesby 60 when olderi~ormationis excluded (a = b = 0; d >0). When all the returns variables are included in the model, a is i n s i ~ f i c a nand t much smaller than d. The persistence estimates, given by the sum a b d, are between 0.984 and 0.993 when old information is included (b >0). Second, consider models that only make of daily use implied volatilities. Thevariable ir-l is biasedbecauseestimates of themultiplier e are significantly smaller than 1. Some of this bias is presumably due to an

+ +

unsuitable choice for the constant f that converts annual variances into hourly variances. Whene and f are unconstrained the maximum of In@) is 21 less than the maximum when spot price quotations alone are used. This shows that five-minute returns are more informative than implied volatilities, at least when estimating hourly conditional variances. Third, consider models that make use of five-minute returns, hourly returns and daily implied volatilities. The most general modelin the final row of Table 3is estimated tohave a zerointercept c and theparameters a, d and e have C-ratios above 3.5 andthus are significant at very low levels. Deleting the implied volatility contribution from the most general model would reduce the maximum of ln(L) by 24. Alternatively, deleting the quotations terms would give a reduction of 45. It is concluded that both the quotations and the implied volatilities contain a significant amount of incremental information.

The hourly seasonal volatilitymultipliers are particularlyhigh in the hour commencing at 08.30 ET when many US macroeconomic news reports are released. This effect is most prominent on Fridays. The volatility multipliers used in the preceding analyses are, for example, the same for This all Fridayhours commencing at 8:30 regardless of any news releases. me tho do lo^ might induce systematic mis-measurementsof the volatility process. We have assessed theimportance of this issue bycomparing the results when there are 120 volatility multipliers with further results when either 121 or 144multipliers are used. Our first set of 121multipliers contains two numbers for Friday 08.30to 09.30 ET: one multiplier for those Fridays that have a relevantreport and another multiplier for the remaining Fridays. Our first set of 144 multipliers contains two numbers for eachof the 24 hours from Thursday 14.30 to Friday 14.30 ET, one used when there is a relevantreport and the other when there is not. We have defined a relevantreport as a newsannomcement about one or more of the six sigdicant macroeconomic variables listed by Ederington and Lee (1993, ~1189):employment, merchandise reports were trade, PPI,durable goods orders, GNP and retail sales. These issued on25 of the Fridaysin our sample. We find that the maximum of the log-likelihood function increases by similar amounts when there are more multipliers whichever model is estimated. Consider the nine log-likelihood values reported in Table 3, panel B, for nine specifications with 120 multipliers. Thesevalues increase by between 6.6 and 8.6 when 121 multipliers are used and by between 16.7and19.6when144 are used. Consequently, as our conclusions

84 S ~ ~ 1.~Taylor e nand Xinz~07zg Xu

depend on substantial log-likelihood differences across specifications the conclusions do not change when the additional multipliers are used. It may be objected that the Fridays have been partitioned by announcements rather than by the impactof unexpected news. Second sets of 121 and 144 multipliers have been calculated by separating the 25 Fridays having the hi~hestrealised volatility from 08.30to09.30fromthe rema~n~ Fridays. The results are then similar,as 19 of the 25 high-volatility hours include a relevant a~ouncement.The increases in the lo~-li~elihoods from the values reported in Table 3, Panel l3 are now in the ranges 8.4 to 10.7 and 9.3 to 11.7, respectively, for 121 and 144 multipliers. We note that the first, second, fourth and fifth Fridays in the ranked list coincide with employment reports but the third ranked Friday has no relevant announcements. The estimates of the parameters a, b, d and e change very little when the number of multipliers is increased above 120. The magnitudesof the changes are all less than 0.03 when the most general model is estimated. When the general modelis constrained by ignoring the options information (e = 0), the persistence measure a b d is always betweep 0.984 and 0.985.

+ +

uarterly ~ u ~ ~ e r i o ~ s

It could be possiblethatsome of theconclusionsareonlysupported by the data during part of the year studied. The higher than average realisedvolatilityduringthefirstquarter,fromOctobertoDecember 1992, mightbe an unusual period whose exclusion would reverse of some the conclusions. The models whose parameter estimates have been givenin the lower panel of Table 3 for the whole year in the datasets have been re-estimated for the four quarters of the year c o ~ e n c i n in~October 1992 and in January, April andJuly 1993. The same 120 seasonal multipliers are used for the whole year and for each of the four quarters, All the conclusions for the whole year are supported by each quarter of the data: (1) when quotations alone are considered, five-minute returns have more volatility information than hourly returns and the relevant information is not all in the most recent hour (li~elihood-ratiotests, 5 percent significance level), (2) five-minutereturnsaremoreinformativethanimpliedvolatilities when estimating hourly conditional variances and (3) there is signi~cant incremental informationin both the quotations and the implied volatilities (li~elihood-ratiotests, 5 percent significance level). The reductions in the maximum of the log-li~elihoodwhen the quotations information is removed from the most general model are 14.7 for

l n c r e ~ e n tVolutility ~l lnfor~ution 85 Table 4 Parameter estimates for the most general ARCH model of hourly returns.

S u ~ ~ l e c x 105 Full

0.0000

Q1

0.0000

Q2

0.0000

Q3

0.0164 (2.06)(0.32) 0,0000

Q4

a

b

d

e

0.0808 (3.64) 0.0775 (1.78) 0.0435 (1.19) 0,1071

0.1801 (1.83) 0.0000

0.0750 (1.47)

0.4420 (2.33)

0.1099 (3.87) 0.1591 (2.54) 0.0995 (1.84) 0.0877 (1.53) 0.1185 (2.50)

0.3887 (6.71) 0.4841 (11.01) 0.3797 (2.82) 0.4334 (2.05) 0.2120 (2.18)

0.2391 (1.05) 0.0000

V

1.1638 (41.45) 1.1621 (20.94) 1.2155 (21.04) 1,1480 (20.05) 1.1421 (20.65)

max. In@,)

32277.00 7726.38 8622.58 7676.79 8255.79

The generalmodel has conditional variances defined Table in 3. There are 120 seasonal multipliers and the conditional distributionsare generalisederror dis~ibutions. The estimates are for the full year (October 1992 to September 1993) and for the four quarters that commence in October 1992, January 1993, April 1993 and July 1993. The numbers in parentheses are t-statistics.

the first quarter, 8.0 for the second quarter, 8.4 for the third quarter and 17.3 for the fourth quarter. The incremental information in the quotations information is thusof a similar orderof magnitude in all the quarters and the first quarter is not clearly different to the other three quarters. The reductions in the m a x i ~ of u the ~ log-likelihood when the information in implied volatilities is removed from the most general model 4.2 for arethe first quarter, 4.4 for the second quarter, 2.4 for the third quarter and 2.1 for the fourth quarter. These reductions are much smaller and are similar across quarters. Table 4 presents the quarterly parameter estimates for the most general model. The estimates change little from quarter to quarter. The sum of the maximuml o g - l i ~ e l ~ o ofor d s the four quarters is only 4.54 more than the maximum when the same parameters are used for the whole year. Twice this increase in the log-likelihood is less than the number of extra parameters when four quarterly models are estimated compared with oneannualmodel.Thereisnostatisticalevidence,therefore,thatthe parameters of the general model changed during the year. The variations in estimated parameters,by quarter, are minor relative to their estimated standard errors. nostic n tat is tics and Tests

A time series of standardised residuals from our most general model for hourly returnsis defined by:

86 Stephen I. Taylor and X i n ~ ~ o nXug

The conditional variances are calculated using the maximum likelihood estimates of the model parameters for the whole year. In the unlikely ev of our model being perfect we would expect the standardised residuals to be approximately independent and identically distributed observations from a zero-mean and unit-variance distribution. {zT}.Their mean is 0.005 and "here are 6049 numbers in the time series is -0.01 and their kurtosis their standard deviation is 1.004. Their skewness is 5.30, both of which are close to the values expected from a generalised error distribution with thickness parameterv near one (skewness = 0 for all U, and kurtosis = 6 when v = l). A histogram of the ZT shows fat tails and a substantial peak around zero, which is a feature of the GET>when U is near one. Twenty of the standardised residuals are outside f 4 although all of them are insidef5.5. ,;z from lags 1 to 10, are all within The autocorrelations of ZT, IZT/ and rfr0.025 = sfr1.96/dm and therefore provide no evidence against the 5 i.i.d. hypothesis, since all 30 tests accept this null hypothesis at the percent level. The first-lag autocorrelations of the three series are 0.003, 0.007 and -0.007. Statistically significant dependence is found at lags that are multiples of 24: for ZT at lag 96 (correlation = 0.056), for ~ Z T I at lags 24, 48, 72, 96 and 120 (the correlations are 0.045, 0.042, 0.049, 0.056 and ;z at lags 24/72and 96 (correlations0.032,0.022 and 0,062). 0.026) and for These correlations show that the model is not perfect, presumably because of estimationerrorsinthehourlyseasonalmultipliers.Nevertheless, with all autocorrelation estimates within f0.07 the model is considered a satisfactory approximationto the process that generates hourly returns. Estimates of spectral density functions, calculated from the autocorrelations at lags 1 to240 of ZT, lz~land,;z confirmthisconclusion. No statistical evidence against the i.i.d. hypothesis canbe found in the estimates at frequencies corresponding to either 24-hour or 120-hour cycles. lz~l There is a significant spectral peak at zero frequency for the series (t-statistic = 3.49) that may simply reflect very small positive dependence at several lags.

FORECASTS OF REALISED VOLATILITY

A comparison of volatility forecasts. can provide further evidence about incrementalinformation. Wedividethewholeyear of dataintoan

in-sample period from which ARCH parameters and intra-day seasonal volatilitymultipliersareestimatedand an out-of-sampleperiodfor which the accuracyof forecasts of hourly realised volatility is evaluated. Wesplittheyearinto a nine-monthin-sampleperiodfollowed by a three-monthout-of-sampleperiod.Relativeaccuracymeasuresfor fiveforecastingmethodsarecalculatedusing120seasonalvolatility multipliers.Therelativemeasuresarenotsensitivetothetreatment of Friday macroeconomic announcements. Using our firstof set 121 or 144 multipliers, definedat p 83, has no effect on the rankings of the forecasts. Two measuresof hourly realised volatility are forecast, defined first by 18 and secondby the same quantity adjusted for intra-day seasonality usin 120 volatility multipliers: 19 Three forecasts of at,j,l are defined by conditional variances ht,j obtained from variations on the most general ARCH model for hourly re~rns defined in Section 5.4. The first forecast excludes all quotationsinformatio~ by imposing the restriction a = b = d = 0 on the ARCH model. The second forecast excludes the options information by requiring e = 0 in theA model. The third forecast is calculated from the general model without any parameter restrictions. These three forecastsdenotedfi,j,l,l, are I = l,?,3. A fourth forecast,ft,j,l,4,is definedby multiplied by the in-sample average of the quantities at,j,2. Four forecastsf;,j,2,1of at,j,2 are defined in a similar way. The first threeof these forecasts are now defined by deseasonalised conditional variances for the three ARCH specifications and the fourth forecast is the in-sample average of at,j,2. The accuracyof a set of forecastsft,j,k,lof the outcomes,at,j,kis reported here relative to the accuracy of a reference forecast givenby the previous realised volatility

Table 5 presents valuesof the relative accuracy measures 21

Table 5 Measures of relativeforecast errors whenforecasting hourly realised volatility out-of-sample.

Forecast 1 quotations 2 3 4 5

Error ~ e t ~ ~ c

A~sol~te

Seasonal f f ~ ~ ~ s t ~ e ~ t NO

options only only optionsquotations and in-sample average lagged realized volatility

0.767 0.781 0.731 1.055 1

A ~ s f f ~ ~ t e S 9 ~ ~ r e Square Yes NO Yes 0.772 0.769 0.731 1.080 1

0.807 0.812 0.797 0.821 1

0.795 0.783 0.776 0.803 1

The accuracy of forecasts is measured by either the absolute forecast error or the squared forecast error. Hourly realized volatility is forecast, either withoutor with a seasonal adjustment. Nine months are used for in-sample calculations and then three months for out-of-sample evaluations. The numbers tabulated are C la - j l P / C la -f5lP with a the realized volatility number. fi forecast I and p either 1 or 2.

for powers p = 1,2.The summations are over all hours in the out-ofsample period. The bestof a set of five forecastsft,i,k,l,I = 1, ,.,5, has the least value of F. The least value of F is considered for each of the four colums in Table 5. The colums are defined by all combinations of p (1 or 2) and k (1 or 2). When accuracy is measured by absolute forecast errors, so p = 1, the best forecasts come from the general ARCH specification for both realised measures (k = 1,2). Thisisfurtherevidencethatthereisincremental volatility informationin both the spot quotations and the options prices. The average absolute forecast error from the general specification is 5 percent less than that from the next best s~ecification.The second best set of forecastsarefromquotationsalonewhenthequantityforecast is adjusted for seasonality, but are from options prices alone when the q~antityforecast is not adjusted, although the differences between the accuracies of the second and third best forecasts are small. The results are similar but less decisive when accuracy is measured by squared forecast errors(p = 2). The most general ARCH specification again gives the bestout-of-sam~le forecasts. However, the avera squared forecasterrors for the best forecasts are only slightly less than for the next best forecasts.This may be attributed to the marked skewness Po the rightof the distributionof the quantities tobe forecast: this inevitably produces some outliersin the forecast errors whose impactis magnified when they are squared.

The evidence from estimating ARCH models usin ratequotationsforoneexchangeratesupports two concl~sions.First ~ve-minutereturns cannotbe shown to contain any information when estimating daily conditional var result may simply be a consequence of the S n u m ~ eof r daily return^ availableforthisstudy.Second,whenestim h o ~ r l yconditional variancesthereisa s i ~ ~ i c a amount nt of inforn in five~minutereturn^ that is incrementaltotheoptionsinformation. rt~ermore,thequotationsinformationthenappearsto be moreinfotivethantheoptions or mat ion. Thusthereis s i ~ i ~ c a increme nt olatilityinformation in one million forei exchange ~ ~ o t a ~ oThis n s conclusion . is c o n ~ i r ~ e by out-of-sample comparisonsof volatility forecasts. Forecastsof hourly realised volatility are more accurate when the~uotations informatio~ is used in a~ditionto options formation.

The authors thank the two referees and the editor for their very and extended versio Frequency Data in F ~ a n c in March 1995. Thea ~ t ~ o r ~ co~erence,the 1995 Euro~eanFinanc ciation conference and the Aarhus at he ma tical Finance confere theircomments.Theyalsothank ~articipantsatseminarsheld Isaac Newton Institute amb bridge, City Univ~rsityLondon, L a n ~ ~ ~ t e r University, Liverpool~niversity,Warwick ~ ~ i ~ e r and s i t ytheUniversit~ of Cergy-Pontoise.

Andersen, T.G. and Efollerslev,T. (1997), "IntradayPeriodicityandVolatility Persistence in Financial Markets", ~ o ~ ro n~ ~E l~ p i ~~i ~ n ~c n~4,115-158. cZe / Barone-Adesi, G. and Waley, R.E. (1987), "Efficient Analytic A ~ ~ ~ o x i ~ aoft i o n American Option Values",~ o ~ ro n~ €~ i ln ~ 42,301 ~ c e , -320.

90 Stephen 1.Taylor and Xinzhong Xu

Bollerslev, T. and Domowitz, I. (1993), ”Trading Patterns and Prices in the Interbank Foreign Exchange Market”, ~ournalof Finance, 48,1421 -1443. Canina, L. andFiglewski, S. (1993), ”TheInformational Content of Implied Volatility”, Review of Financial Studies,6,659-681. Dacorogna, M.M., Muller, U.A., Nagler, R.J., Olsen, R.B. and Pictet, O.V. (1993), “A Geographical Model for theDaily and Weekly SeasonalVolatility in the FX Market”, ~ournalof lnternational Money andFinance, 12,413-438. Day, T.E. and Lewis, CM. (1992), “Stock Market Volatility and the I n f o ~ a t i o n Content of Stock Index Options”,~ournalof Econo~etrics,52,289-311. Ederington, L.H. and Lee, J.H. (1993), ”How Markets Process Information: News Releases and Volatility”, ~ o u of~F~nance, a ~ 49,1161 -1191. Ederington,L.H.and Lee, J.H. (1995), ”TheShort-runDynamics of the PriceAdjustment to New Information”,~ournulof Finuncial and ~ u a n t i t a t iAnulysis, ~e 30,117-134. Goodhart,C.A.E.andFigliuoli, L. (1991), ”EveryMinute Counts in Financial Markets”, ~ o ~ r nofa lnternational l Money andFina~ce,10,23-52. Hull, J. (1995), lntroduction to F~turesand O~tions Mur~ets, 2nd edn., Prentice-Hall, Englewood Cliffs, NJ. Hull, J. and White, A. (1987), ”The Pricing of Options on Assets with Stochastic Volatilities”, ~ournalof Finance, 42,281 -300. Jorion, P. (1995), ”Predicting Volatility in the Foreign Exchange Market”, ~ o u r ~ u l of Finance, 50,507-528. Lamourem, C.B. andLastrapes, W.D. (1993), ”Forecasting Stock Return Variance: ~~ ~ i~ ~ ~n a Toward an Understandingof Stochastic plied Vola~lities”,R Studies, 6,293-326. Nelson, D.B.(1992), ”Filtering and Forecasting with Misspecified ARCH models I: Getting the Right Variance with the Wrong Model”, ~ou~nal o~Ec~~ 52,o ~ e ~ 61 -90. Taylor, S.J. and Xu, X. (1994), ”The Magnitudeof Implied VolatilitySmiles: Theory and Empirical Evidence for Exchange Rates”, Review of future^ M a r ~ e t 13, ~, 355-380. Xu, X. and Taylor, S.J. (1994), ”The Term Structureof Volatility Impliedby Foreign Exchange Options”,~ournalof Fi~ancialand ~uantitativeAnalysis, 29,57-74. Xu, X. and Taylor, S.J. (1995), ”Conditional Volatility and the Informa~onalEfficiency of the PHLX Currency Options Market”, ~ournalof ~ a n ~ and ~ nFina~ce, g 19,803-821.

REFACE

This chapter addresses three problematic issues concerning the application of the linear correlation coefficient in the high-frequency financial datadomain.First,correlation of intra-day,homogeneoustimeseries derived from unevenly spaced tick-by-tick data deserves careful treatment if a data bias resulting from the classical missing value problem is to be avoided. We propose a simple and easy to use method which corrects for frequency differentials and data gaps by updating the linear correlation coefficient calculation with the ofaid co-volatility weights. We view the method as a bi-variate alternative to time scale transformations which treat heteroscedasticityby expanding periods of higher volatility whilecontractingperiods of lowervolatility.Secondly, it isgenerally recognized that correlations between financial time series are unstable, and we probe the stability of correlation as a function of time for seven years of high-frequency foreign exchange rate, implied forward interest rate and stock index data. Correlations estimated over time in turn allow for estimationsof the memory that correlations have for their past values. Third, previous authors have demonstrated a dramatic decrease in correlation as data frequency enters the intra-hour level (the "Epps effect"). We characterize the Epps effect for correlations between a number of financial time series and suggest a possible relation between correlation attenuation and activity rates.

The estimationof dependence between financial time series is of increasing interest to those concerned with multivariate decision formation. ~inancial~ a r ~Tick t sby Tick Edited by Pierre Lequeux.0 1999 John Wiley & Sons Ltd

92 Mark Lundin, Mi~helV a c o r o ~ aand ~ l r i A. ~ hM ~ l l e r

ce is often characterized numerically using the linear correl The p o p u l a r i ~of this estimation technique stems from its ~ i m pdefinition/ l~ practical ease of use and from its straightforward results 'ch are easily interpreted/~ t l e sand s directly com~arable.Despite the tive simplicity of its d e f ~ t i o n there / exist a number of unresolved ation and ~terpretationof results in the .l

The data input for the correlation coe~icientcalculation are two time ith equal, usually homogeneous, spacing between observati This necessity is easily satisfied where low frequency (5 one week)dataisconcerned,However,formulation of intra,homogeneous eserves more careful treatment if a ulting estimation avoided. This especially is the case e series of unevenly spaced tick-by-tick data encies or opening(overlap)hours. ~e propose lied normalization method which corrects rvationfrequencies and for data gaps.This dates the correlation calculation only when ata exist and not when there is none, ensuring that there is no ias resulting from the miss in^ value problem (Krzanowski t, 1994, 1995) and when time series with largely different c~aracteristicsare correlated.h addition this approach remains scale free and straight~orwar~ to implement and interpret. ~e view the ethod asa bi-variate alternative to time scale ~ans~ormations which y expanding periods of higher

tion coefficient does not account for of two ces over time. The variances

Correlationo ~ ~ i g ~ - F r e q uFei ~ ac ~ y c iTime a l Series 93

of correlation” is proposed as the basis for the formulationof a long term correlation forecast. Foranygivenfinancialdistribution,therole of thefrequency of timeseriesdataontheestimation of correlationshould be clearly established. This is especially relevant as higher frequency data become more widely available and more often used in order to boost statistics. Previous authors have demonstrated a dramatic decrease in correlation estimations as data frequency enters the intra-hour level, for both stock (Epps, 1979) and foreign exchange returns (~ui~laum et eal, 1994 Low eC al, 1996). We follow the suggestion of Low et al, (1996) by referr this phenomenon as the ”Epps” (Epps, 1979) effect, In this discu an attempt is made to characterize and investigat more deeply the Epps effect in a number of financial time series throu h the examination of seven yearsof high-frequency financial returns. The discussion which follows will cover the following points. First the characteristics of the financial data used for the study are specified. a covolatility normalized method of adjust in^ the standard corre coefficient for differences in data frequency between time series and as a compensation mechanism for data gaps or non-overlapping periods is to illusdescribed. Monte Carlo and high-frequency financial data are used trate the characteristic features of this alternative methodology. Estimation of the relative time dependent variance of correlations betweena number of instrumentsis then considered and the dependence of these correlations on their past values is determined. A simple parameterization of correlation self memory, tested on various financial instruments, is proposed. This is followed by the estimation of correlation as a function of varying data frequency; the previously mentioned Epps effect is examined in detail. relation is drawn between the attenuation of correlation in financial time series and the activities of the correlation constituents. Finally, conclusions are drawn regarding the correlation of high-financial time series. ~ E C I F I C ~ T I OOF N THE

Starting with an observed, non-equally spaced tick-by-tick time we series, use linear interpolation to construct an equally spaced times series with time between observations equalAt.toLinear interpolation may introduce some formof undesirable dependence in the data aand correction method applicable for estimation of correlation willbe discussed later. We address threetypes of differentintra-daypricetimeseries in thisdiscussion, Foreign exchange prices are reportedin terms of the bid-ask spread and

94 Mark Lundin, Michel~ a c a r a gand ~ a Ulrich A.Muller

we definethe logarithmic middle price (Guillaumeet al, 1994), x(ti), as the arithmeticmean of the logarithmic bid andask quotations and Pa& respectively) as given in Equation 1:

where At is the duration of a fixed timeinterval and x(ti) is the sequence of equally spaced (byAt), logarithmic prices. Stock indexdata arrives in terms of a simple price which we transform as:

2 Defining a time seriesof forward interest rates is less straightfo~ard. As a basis, we use interest rate futures (Eurofutures) with fixed expiry dates which represent the forward interest rate starting at the expirydate and ending after a fixed period. A time series of prices coming from a a study, single Eurofuturescontract is not large enough to allow for serious therefore it is necessary to join contracts together in anappropriate way. A suitable continuous-time series that can bederived from Eurofuturesis the series of forward interest rates. Here the beginning of the forward interest period (which we choose to be of three months) is always at a fixed time interval, e.g. six months, from the quotation time. Sometimes, when the expiry of a 3-month Eurofutures contract happens to be six months in the future (in the aforementioned example), the resulting forward interest rate coincides with the interest rate implied bythat Eurofutures contract. forward interest rate x For a time t i where this is the case, the logarithmic is definedas follows: 3 where f is the Eurofutures quote and (1--f/lOO%) is the annualized implied forward interest rate (which could be multiplied by 100 percent to obtain theusual form of an interest rate). In allother cases, the forward period to be taken overlaps with the forward periods of more than one ~urofuturescontract, so we need an empirical method to compute the forward interest rate from several contracts. A number of such methods exist and typicallyattempt to simulate what a trader does when holding a contract and switching (rollingover) to another, some time before the expiry of a contract. Other methods try to obtain the best estimate of the true forward interest rate through linear or non-linear interpolation of implied interest rates of Eurofutures contracts. The methodused here belongs tothe latter type and is described by Muller, (1996). However, the choice of this method is not the focus of this paper; other methods may be

Correlationof ~ig~-Frequency Financial TimeSeries 95

suitable. The final forward interest rate is in logarithmic form asin shown Equation 3. h the following study, we are less interested in the (logarithmic) prices et al, 1994)which we of these instruments than in their returns (Guillaume define as: -. AXi

E

Ti

r(At; ti) E [X(ti)

- ti -At)]

4

The discussion which follows involves the use of intra-day foreign exchange, stock index and implied forward interest rate return values ranging from January 9,1990January to 7,1997. Correlations between the following financial instruments were considered: USD/DEM-USD/~LG USD/DEM-USD/GBP USD/DEM-USD/ITL DEM/GBP-USD/GBP USD/FRF-USD/GBP USD/JPY -DEM/ JPY Dow Jones Industrial Average (DJ1A)-AmericanStockExchange Index (Amex) USD 3 to 6 month implied forward interestrate-DEM 3 to 6 month implied forward interest rate DEM 3 to 6 month implied forward interest rate-DEM 9 to12month implied forward interest rate USD impliedforwardinterestratedataconsisted of transactionsat theChicagoMercantileExchange(CME) and DEM impliedforward interestratedatadescribetransactions at theLondonInternational Financial Futures and Options Exchange (LIFFE). Further characteristics of these implied interest rate data is described in Ballocchi et al, (1998).For thesake of comparison,wenegatethereturnsfromthe "cable" crossrate GBP/USD. Thus they are reported asUSD/GBP instead. ThecrossrateGBP/DEMwasconvertedtoDEMIGBP in thesame manner. The correlations between instruments of fundamentally different financial natures are not discussed in study this since they were observed to be largely uncorrelated, making them less interesting for the purposes of this discussion. Tick activity averages for the instruments used in this discussion are shown in Table 1. These relative numbers are not actual volume figures, but the mean number of price quotations observed per day (including weekendsand holidays) and for the database which was

96 Mark ~ u n d i nMichel , Dacorogna and ~ l r iA. c M~ller ~

Table 1 Activity (interms of the average number of quotations per day) for the financial ~ s t ~ m e nconsidered ts inthis study. The sampling period was from January 9, 1990 to January 7, 1997 (2,555days). Financial ins~r~~ent

USD/DEM USD /JPY USD/GBP USD/FRF USD/NLG USD/ITL DJIA DEM/ JPY Amex DEM/GBP DEM 3-6m IR USD 3-6m IR DEM 9-12m IR

Mean price q~otutionsper day

Mean price quotes per b~sinessday

3390 1492 1217 708 594 432 385 328 319 280 156 97 89

4715 2060 1697 991 831 604 539 454 446 390 218 136 125

used. For some instruments, multiple data sources were merged to form more populated time series while for others only one source was used. All foreign exchange (FX), stock index and implied forward interest rate returns were obtained in physical time unless otherwise stated and included weekdays as well as weekends and any holidays. A relatively two to reliable set of data filters were also applied to the data, removing three percentof the dataas obviously false outliers.

EIGHTI~G IN THE ~ O R R E L A T I O ~

Estimation of the correlation coefficient is straightforward but some inc venience is introduced via its simple definition. The usual definition of the sample correlation requires two equally spaced stationary time series as input. This necessity is easily satisfied when low frequency (5 one tick per week) data are concerned. However, the problem requires more careful treatment for higher data frequencies and where one cannot dictate the observations times or number of observations (Maller et al, 1990). Oneoftenfaces two mainproblemswhenestimatingcorrelation between two high-frequency financial time series. First, the two observed time series usually have completely different frequencies. If both time series happen to occur at completely regular time grids but with differen

Correla~io~ o ~ ~ i g ~ - F r e ~fin u ean~cia^ c yTime Series 97

frequencies, one might want to generate from them two equally spaced time series with equal frequency, just by taking the time grid with the smaller frequency as joint the time grid for both series. However, this easily satisfied situation does not occur very often in practice. More frequently, USD/DEM foreign exchange rate, one is faced with time series such as the which can vary from one thousand or more quotes per hour (from a single data supplier) to tens of quotes per hour, sometimes within a 24-hour to statistically measure or estimate period. at is then a reasonable way the dependence between this foreign exchange rate, with its associated varying arrival times, and another one which is perhaps less active or with activity peaks and valleys at completely different times of the day or week? Ideally, one would prefer to update the correlation estimation more often when more information exists and less often whenno or little ~ f o r ~ a t i is o navailable. Formulating two equal but unevenly spaced time series grids for both instruments wouldbe a possible solution. In fact, a time scale transformation which treats seasonal heteroscedasticity, known in the literature as @("theta")-time (Dacorogna et al, 1993), has already been demonstrated for high-frequency foreign exchange data. This time scale models the intra-daily deterministic seasonal patterns of the volatility caused by the geographical dispersion of market agents. Weekends and holidaysarealsoaccountedfor.One of its characteristicsconsists of compressing physical time periods of inactivity while expanding periods of greater activity. Although this method has proven useful for a number A of applications, its implementationistimeconsuminginpractice. multivariate formulation of @-time even increases the complexity of the situation. In addition, one of the characteristics of @-time is to remove seasonalities in order to measure more subtle underlying effects. However, deseasonalizing is not necessarily suitable for estimating correlation; one risks to eliminate pertinent information from the time series. A second problem one faces when estimating correlation between two high-frequency financial time series is that of missing values or data gaps. Large data gaps are actually a border of the casefirst problem (varying and non-matching data arrival frequencies) but there is no harm in discussing them separately for purposes of motivation. Despiteone's best efforts, data gaps sometimes occur due to failure in the data acquisition chain. One two time can only makean educated guess about the correlation between series when such a gap occurs,it cannot be estimated empirically. More commonly, there exist financial instruments whose time series exhibit regular and large data gaps by virtue of their definition. Consider, for example,attemptingtoestimatethecorrelationbetween an exchange reported stock index such as the Row Jones Industrial average, which exists for 6.5 hours per day, five days per week (usually), and another

98 Murk ~ u n d ~Michel n, ~ a c oand~ Ulrich o ~A.~M ~ l l e r

instrument which exists fora similar amount of time eachday but with a relatively large time shift(perhaps the Financial Times 100 index). There are a number of different schools of thought regarding the correlation between two financial instruments when one or both are not actually active. One solution might to carry be out the estimation with the derivativ of an instrument as an inputproxy for the underlying, but these are often sigruficantly different in character and a number of application-limi~g assumptions would be involved. Another possible solution might involv estimation of intra-day correlation using all available data, but after applying a simple time shift to oneof the geographically offset markets so that, for example, the closing price in New York occurs at the same time as the closing price in London. However, this is actually a form of time lagged correlation estimation and we consider it a different issue entirely. When confronted with varying activity rates and data gaps, it often seems convenient to use some form of data interpolation to solve ones problems. Unfortunately, the experienceof many practitioners has not been reassuring (Press et al, 1992). There exista number of methods for approximation of an equally spaced time series from an irregularly spaced, tick-by-tick data set. Most of these methods involve some form of data imputation. Methods of imputing data vary in complexity and effectiveness; most have been found to be beneficialunder at least some set of conditions and assumptions. However, all forms of imputation rely on a model and a standard supposition is that characteristics of the data do not change between in-sample and out-of-sample periods. There is always the possibility that imputation will introducea bias into variance and co-variance estimations but nevertheless it is difficult to avoid some form of it in cases where data is not of an infinitely high-frequency. Some useful attempts have been made to circumvent imputation altogether. Oneinteresting and recent example is described in De Jong and Nijman (1997). This workbuilds primarily on et al. (1983), and Lo and MacKinlay, (1990a,b). efforts described in Cohen The authors develop a covariance estimator which uses irregularly spaced data whenever and wherever it exists in either of two time series, However, methods such as this one rely on the assumption that process generating transaction timesand prices themselves are independent. This assumption may be well-suited,depending on theinstruments and application being considered. However, testing for this independence is rarely trivial and we prefer to avoid the assumption altogether. Instead we developwhat we see asa complementary technique. In this discussion, we proposeand illustrate a simple estimateof correlation which avoids imputation based on data models that are constructed from out of sample data or on distributional assumptions. Although the

Correlation of ~ i g k - F r e Financial ~ u ~ ~ Time Series 99

inputs for this alternative estimator are equally spaced time series derived through simple linear interpolation, the method filtersout any underestimation of variances and co-variances caused by the lack of sampling variation which results from the over-interpolation of data. In addition, rather than making the strong assumption that prices and transaction times are independent, this method makes use of arrival times in order to compensate for the (sometimes large) differences in the observation frequencies of financial time series. Data gaps of varying size are c o m o n and we avoid any discussion of whether correlation actually exists during this period, since in any case we cannot estimate it directly. Our goal is rather to develop a statistical measureof dependence (correlation) when information exists and to avoid updating this estimation when data are not available.This should be recalled whenit the time comes to interpret results. This approach implies that a lower data frequency or data gap in one time series may limit the use of another one. The unavoidable price to pay for such a methodology is the loss of statistics in the estimation. However, this method is specifically meant for estimationof correlation at higher data frequencies where computational statistics become of aless constraint on accuracy. Formulation of an Adjusted Correlation Estimation

The standard linear correlation coefficient isan estimator of the correlation between two time series, A X i and A y i whose definition is given in Equation 5:

5

where

n

n

Thesampleis of size T with n = T / b t equallyspacedobservations. -1 (completely antiCorrelation values are unitless and may range from correlated) to1 (completely correlated).A value of zero is indicativeof no correlation between the two time series.

100 Mark L~ndin, Michel Dacorogna and ~ l r i A. c~ Miiller

h estimate of the local sample covolatility2 for each of these regions isdefined by furtherdividingeachtimeinterval of length At (on which Axi and Ayi are evaluated) into m disjoint subintervals of equal length Ai = &/m, and we may obtain sub-return values, Ani and APj, j = 1, .. ,m in the usual way. This refined time series now consists of G = ?I/Ai equally spaced (via linear interpolation) observations on retur

?(",I

n(Ai;Zj> = [Inpask(ij)

+ InBid(Zj)l/?

7

Then for each of the periods [ti-l, t i ] with corresponding coarse returns, Axi (as for Atlyi), we can define an estimator of the covolatility between the two refined time seriesof returns:

where A2l.m =

j=1

I

m

and Aiji.m =

j=1

112

9

The most obvious choice for a is 0.5, though the choice of the value of a can be studied as a way to magnify or demagnify the weight given to farther outlying return values. A value of 0.5 is used inall cases described in the discussions which follow. Equation 8 defines covolatility around the mean rather than around and wi = 0 for the caseof returns derived fromtwo linearly inte~olated prices existing outsideof our regionof interest, At. This follows from the fact that eachof the j sub-return values wouldbe equal to each other and also equal to their mean value. The difference between theseis equal to zero and the sum of the products of zero and any other quantity will also be equal to zero.3 These covolatility estimates can be inserted as weights into the sample variances and covariances to obtain an estimator of correlation:

Notice that Axi and Ayi in Equations 10 and 5 are the same since they are taken over the same time period, [ti-l, ti]. The return values on the

C o ~ e l a t iofo ~~igh-Fyeq~ency Financial Time Series 101

coarser grid with span At c m then be defined as the sum of the return values on the refined grid with span At: 11 j=1

There are two sums left in Equation10 which areupdated with linearly interpolated returns even in cases when no actual information exists: ~Ay, Ax. That is the case if we define them according to Equation6. One would rather prefer that mean values, and G,be calculated again in a weighted fashionso that true return values are input to the calculation where they exist and not where theydo not exist.h addition, thereshould be assurance that weighted means are calculated over the same data sample used for the rest of the correlation calculation. Therefore we define weighted mean values for both time series which are also covolatility weighted: T/At

TlAl

12 i d

i=l

Equation 12then defines mean values which are unbiased by a i if all Axi (or Ayi) are equal. The weights correctly adjust for periods of lower or higher activityand in periods when data does not exist the mean value is numerically equal tothe usual mean value calculation but with data gaps removed. that It was already notedthat Equation 8 is formulated in such a way ~ t )= i 0 for the case of returns derived fromtwo linearly interpolated prices existing outside of our region of interest, At, and in that casethe summations of Equations 10 and 12 are not updated. The covolatility adjusted estimation of correlation described by Equation 10 also retains the desirable characteristicsof the original, standard linear correlation coefficient; it is scale free, invariant and completely different estimations are directly comparable. In addition, this alternative method is only slightly more complicated to implementthan the standard linear correlation coefficient and can easily be implemented on any personal computer. As will be applied later, this statistical correlation measure also fits easily intothe framework of autocorrelation analysis. Given a time series with a copyof itself but with different of correlations&,it can be correlated

102 Mark ~ u n d i nMichel , ~ a c o r oand ~ a~ l r iA. c Miiller ~

time lags(t)between the two,as shown in Equation 13: n

for t >0 and where n

14 For the discussions which now follow, we estimate correlation using 10, unless otherthe covolatility adjusted method described by Equation wise stated and always with m = 6 and a = 0.5 (see Equation 8). Any subsequent useof the comonly recognized linear correlation coefficient (Equation S) will be referredto as the "standard" method. ethod with Monte Carlo andFinancial Data

Various tests were performed with the adjusted correlation esthator in order to test its effect on data gaps and on time series of differing frequencies. Monte Carlo data were used to illustrate the covolatility adjusted correlation method's effectiveness on sections of missing data. Two separate (mcorrelated), normally distributed, random data sets, Ai, Bi, were produced with zero mean, withstandard deviation CT = 0.01 and length m = 10,000. A third distribution, Ci, was then formed as a 15. linear combinationof the previous two, according to relation 15

where the constantk is selected suchthat 0 5 k 5 1. In this way, the new distribution C i was formed with a controllable correlation to oneof the original distributions,Ai. The pseudo return distribution, C i , wasthenusedtocreate a pseudo price distribution, Pi, with starting value P1 = 10 and length m 1 = 10,001. Subsequentpricesweregeneratedbythe return distribution, Ci, according to relation16.

+

16 Repeated sections of fixed length equal to 50 price values were then deleted in the pricedistribution Pi and replaced with prices linearlyinterpolated from the previousand next actual prices.The distance between

Correlationof ~ i g ~ - F r e ~ u Financial ency Time Series 103

these induced data gaps wasalso of length equal to 50, creating an alternating series of original data patches following by data gaps filled with linearly interpolated prices. Finally, a return distribution, Q, was created from this altered price distribution containing periodic patches of linearly interpolated data as prescribed by Equation 4. Equation 10 was then used to estimate the correlation between one of the original return distributions, A i , and the manipulated return distribution, .Q, given various values of the constant multiplier k. Results are shown in 2. comparison tothe standard linear correlation calculation in Table Comparison of columns two @(A, C)) and four @(A,L))) shows that the 10 well covolatility adjusted correlation estimator described by Equation approximates the original, coarse, standard linear correlation between distributions A and C before missing data patches were induced and replaced with linearly interpolated values. Any small deviations which exist are within the bounds dictated by statistical error (-2%) for these tests. This simple example illustratesofone the original design goals of the covolatility adjusted linear correlation estimator; correlation is statistically measured when data exists and the calculation isnot updated when data does not exist. Tests were also performed to exemplify the effect of the covolatility adjusted correlation estimator on time series of differing frequencies. return values were High-frequency USD/DEM3 minute interval, absolute produced in both normal, physical time and through the de-seasonalizing time scale transformation Mime. The absolute value of USD/DEM returns were used since they are known to have autocorrelations of greater magnitude than with actual returns. The USD/DEM foreign exchange rate is one Table 2 Monte Carlo comparison of the standard linear correlation between two related pseudo return dist~butionsand the covolatility adjusted linear correlation method described in the text.

0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

0.00 0.12 0.23 0.38 0.52 0.69 0.83 0.92 0.97 0.99 1.oo

0.00 0.10 0.15 0.28 0.40 0.51 0.62 0.67 0.72 0.74 0.74

0.00 0.12 0.22 0.38 0.51 0.69 0.82 0.91 0.95 0.97 0.99

104 Mark L u ~ d i Mich~l ~, Dacorogna and ~ l ~ iA. c Muller h

of the more active (see Table 1) but is also characterized by large intra-day and intra-week activity fluc~ations(Dacorogna et al, 1993). Autocorrelations (Equation 5) of the absolute returns of this exchange rate were performed in order to demonstrate the effect that correlation e s t ~ a t o r s have on time series whose activity at any given time differs. First, estim tion using the standard method of correlation was performed on a time series withl 8 minutes between data points, where linear interpolation to closest tick values was used to forma homogeneous time series fromm unevenly spaced tick-by-tick data set. In addition, the standard correlation method was applied to a time series of 18 minute @-time DE^ return values. Finally, correlation estimation using the alternative metho describedby Equation 10 was performed on the homogeneous times seri with 3 minute data intervals and with m = 6 (Equation 8), therefore the final granularity of all correlation estimations was the same. Results of these estimations are shown in Figure 1. A total data periodof six months was used, ranging from January 1,1996 to July 1,1996.

I

0

10

20

lag (hours)

30

The triangles referto the standard autocorrelation method applied to the return time series. Darkened circles referto covolatility adjusted autocorrelation estimation. Crosses refer to the standard autocorr~lation estimations but applied to~ S D return ~ values E ~ measured inMime. The sample period used was January1,1996 to July 1,1996.

Autocorrelation of the absolute valueof ~ S D / returns ~ E ~ as a function of lag time.

The covolatility adjusted correlation values (darkened circles) are significantly less than autocorrelation values using the standard method of estimation (triangles). We ascribe this difference to the fact that periods of higher combined or coactivity are given more weight in the correlation calculation than periods of lower coactivity. This amounts to the same as expanding the time scale during periods of higher activity and compressing the time scale during periods of lower activity. The excep24 to hours where thetwo methods tion to this difference is at a lag equal result in largely the same correlation values. In addition, it is noted that the covolatility adjusted correlation estimations largely reproduce correlation estimations calculated with the standard method but with data measured of activity sinceit in +time. The 6 time scale is itself a statistical measure of inactivity while expanding periods of is designed to compress periods higher activity, thus the two results are largely the same. The exception again occurs at the peak in covolatility adjusted correlation at a lag of 24 hours. The6 time scale is also designed to remove intra-day and intraweek seasonalities. For correlation studies this is not always desirable and we find the covolatility adjusted correlation estimation tobe a more suitable method for many applications.h addition, the simplicityof this methodology lendsitself to wider use.

two time series, the implication is When correlation is calculated between that this quantity does not vary itself over time. For the case of financial time series this is seldom the case, although variance of the correlation coefficient over time can sometimes be small. This issue is critical for portfolio pricing and risk management where hedging techniques can become worthless when they are most needed; during periods known as correlation "breakdown", or relatively rapid change. As demonstrated in Boyer et al, (1997),detection of correlation breakdown. or other structural breaks by splittingareturndistribution into anumber of quantiles can yield misleading results. We used high-frequency data to estimate correlations literally as a function of time for a number of different financial of change which can time seriesin an effort to better understand the level occur. These high-frequency correlation estimations are contrasted with lower frequency, lower time resolution estimates for the same sample periods. The "memory" that correlation coefficients have for their past values was also estimated for a numberof examples using a simple and

106 Mark Lundin, Michel ~ u c o r o and ~ a ~ l r i c hA. M ~ l l e r

appropriate parameterization. Such estimations can be applied to long term correlation forecasting which is required, for example, in order to price or hedge financial options involving multiple assets (Gibson and Boyer, 1997). ~ o r r e l a ~ o n ~ a ~ aOver t i o Time ns

Thegeneralstability of correlationcoefficientswereexamined using variouscorrelationcalculationintervalsanddatafrequencies.This involved examination of a fixed historical data set of horizon or period T , The temporal set of returns (r(ti)) was then divided into N subsets of equalsizeintime ( T / N ) fromwhichcorrelationcoefficientsare computedaccordingtoEquation 5. This wasperformedonourdata set ranging from January 7, 1990 to January 5, 1997. Four values of N were selected, while the total period, T, always remained constant. A number of returns, n, were then obtained via linear interpolation from inside the period T / N , The frequency (or resolution) of data involved in each correlation calculation, f = (n x N ) / T , was adjusted throughout the four sets in order to maintain nearly uniform statistics, as shown in Table 3. Results from these calculations are shown in Figures 2 to 7, where correlations versus time are displayed as circular data points and dashe lines above and below zero correlation 95 are percent confidence intervals assumingnormallydistributedrandomdistributions.Theconfidence limits are slightly non-uniform due to small sample-to-sam~levariations in statistics. Although exactly matching homogeneous data grids were used for each sample, the correlation calculated is not updatedif a weight, mi, from Equation8 is equal to zero. Statistics were increased by one inall other cases. Table 3 The various calculation periods and frequencies selected in order to maintain nearlyuniform statistical confidence for the study of linear correlation coefficient variation over time. The total sampling period, T,was from January7,1990 to January5,1997. Co~relation calculation period T/N

days

365days 384 128 days 32days 7

Data Pequency (quotes/da~) f = (n x N ) / T

1 3 12

N u ~of~ e ~95% Con~dence points l i ~ i ~ n 1.96/fi 365 384

0.10 0.10 0.10

Co~ela~ion of High-Frequency Financ~lTime Series 107

Correlation coefficient mean values and variances are given for each pair of the four calculation frequencies in of financial instruments and for each Table 4. Having virtually the same statistical significance all correlation for of observations calculations shownin Figures 2-7, one can make a number about correlation stability. The highly correlated ~ S D / D E ~ - ~ S D / ~ L ~ FX returnsshown in Figure 2 appearlargelyconstant over thetotal sample period of seven years. As the sub period width for correlation calculation decreases(and the numberof correlation calculations inside the total period increases) more structure becomes apparent. This additional structure is reflected in a changing variance(see Table 4).

-

USD/DEM - USD/NLG

USD/DEM USD/NLG

.......*.*~..

. . * . . . . . . . . a

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

0

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

1996 1994 1992 1990

1996 1994 1992 1990 Time (year)

Time (year)

USD/DE~ USD/NLG

USDlDEM USD/NLG

-

-

9. P

2. iij .g-

g ~

*

.............................................................. .............................................................

P

p?

Ln.

t

_.

1996 1994 1992 1990

1996 1994 1992 1990 Time (year)

Time (year)

T!N = (365 days, 128 days, 32 days and 7 days), for the FX return pair TJSDDEMUSD/NLG.The dashed lines above and below zero correlation are 95% confidenceintervals assuming normal, random distributions.

Figure 2 Linearcorrelationcoefficientscalculatedusingincreasingly small subintervals.

108 mar^ ~ u n ~ iMichel n , ~acorognaand ~ l r iA. c~ ~ l l e r USD/DEM USDlGBP *

" " " " " " " " " " " " " " " " " " "

" " " " " " " " " " _ _ _ _ L _ _ _ _ _ _ _ _ _ _ _

1996199419921990

1996199419921990 Time (year)

Time (year)

USD/DEM USDlGBP

USD/DEM - USD/GBP

-

"""""""""""""".""-""""..,l *

",-,e

q." ,*,,.. ,. ," . " "...,... .-..-..,& *. ' ..

, . , *- . ...".....-,""

" . a " . ,

"11"

.".".".,, six months depending on the instruments involved).

Previous authors have observed a dramatic decrease in correlation as data frequency enters the intra-hour level, forboth stock (Epps, 1979) and FX returns (Guillaumeet al, 1994, Low et al, 1996). We follow the suggestion as the Epps effect after of Low et al. (1996)by referring to this phenomenon

118 Mark Lundin, Michel Dacorogna and ~ l r i c hA.Milller

the first identifiable author (Epps, 1979) to thoroughly document it. In this discussion, the Epps effect is characterized and investigated for a number of foreign exchange rates, stock indices and implied forward interest rate pairs through the examinationof the same seven years of high-frequency return values as has been described previously. High-frequency returns for the pairs of financial instruments given in of equally spaced time series Table 1were used as the basis for formation with varying data time intervals ordata frequencies. Linear interpolation between points was used in order to insure precise time alignment of data from one time series to another. 1,377 equally spaced time series were assembled in this way;500 time series were formedwith data time intervals varying from1data point per 5 minutes of time to 1data point per 2,500 minutes, this with an interval of 5 minutes, Another 877 time series were formed to have data time intervals ranging from one data per 2,530 minutes to onedata per 28,810 minutes (20 days), going bysteps of 30 minutes of time interval. Various calculations were performed with the times series, including the calculation of their variances and the covariances and correlations between different financial return time serieswith the samedata time intervals. Diminution of correlationswhen performing calculations with higher frequency (intra-day) data can be observed in data frequency increases (or Figure 10. A rapid decline to zero is noted as data time interval decreases).This is better viewed in Figure 11,where the same data is shown with a logarithmic horizontal scale andwhere the data point farthest to the left (highest data frequency or smallest shown interval between data points) corresponds to correlations calculated using linearl interpolated, homogeneous time series with five minutes between data points. Table 6 gives the minimum and maximum values for the linear correlation coefficient data shown in Figure 10. Also given are the time intervals at which maximums occurred.We noted several problemswith taking the maximum valueof a given correlation vs. time interval distribution. Noneof the carelation values reach true a stabilization value, even at (> 10 days). This can be the result of correlations very large time intervals rising asymptotically unity to with data intervals approaching infinity but is certainly also a resultof the inevitable increasedspread in correlation data arising fromthe gradual loss of statistics moving toward greater data intervals. Inan attempt to more accurately characterize Epps effect drops, Table 6 also reports the arithmetic mean value of correlation calculations for correlation values whose time interval was between one and two days. These calculations involved 224 data points for each of the correlation pairs. The correlation value of relevance for a particular situation depends entirely on the time horizon one is interested in; there is no best c value forthe general case, only for a particular case. However, this mean

Correlationof ~ig~-Frequency Financial Time Series 119

-

-

USDlJPY JPYIDEM

USDlDEM USDlNLG

0

5 10 15 Data interval (days)

20

0

I

-

5 10 15 Data interval (days)

20

I

-

0

5 10 15 Data interval (days)

20

IM 3-6month-DEM 9-12 monl

DJlA Amex

9

20

USDIFRF -USDlGBP

USDlDEM USDlGBP

O

5 10 15 Data interval (days)

I

T-

i.

u?

0

c

.-

m 0

CI

g o

_ .

S

u?

0

I

9 . l

O

5 10 15 Data interval (days)

20

'

0

5 10 15 Data interval (days)

For all calculations the total sampling period remainedconstant (from January 9, 1990 to January 7,19971, causing the 95% confidence intervals to be small at high data frequencies and larger as data time interval decreases. Rapid declinesin correlation at higher data frequencies are noted in all cases.

Figure 10 Linear correlation coefficients calculated for six example correlation pairs as a function of return data time interval.

120 Mark ~ u y z ~ iMich~l n, ~acorogna and LIlrich A. M ~ ~ l e r

-

U S D l D ~USDlNLG ~

*

..*

0.1 1.0 10.0 Data interval (days)

U S D l D E ~- USDlGBP *

*

-

USDlJPY JPYlDEM

011

l:o

10.0

Data interval (days)

USDIFRF - USDlGBP

.e.*

0.1 1.0 10.0 Data interval (days)

DJlA -Arnex

0.1 1.0 10.0 Data interval (days)

0.1 1.0 10.0 Data interval (days)

DEM 3-6r n o n t ~ ” D E9-12 ~ month

0.1 1.0 10.0 Data interval (days)

This is the same data as Figure in 10 but shown with logarithmic horizontal axes. Rapid declines in correlation are noted in all cases for higher return frequencies.

ure 11 Linear correlation coefficients calculated for six example correlation pairsas a function of date time interval (inverse frequency) used in the calculation.

Table 6 Minimum and maximum values for the linear correlationcoefficient data shown s and in Figures 10 and 11. Also given are the time intervals at whichm a x i m ~ occurred the mean values of correlations for data intervals between one andtwo days. In addition, and the data intervals at which correlation the time intervals at which mean values occurred coefficients rose to at least 90%of the mean value-a nominal sta~ilizationpoint- are reported. The sampling period was from January 9,1990 to January7,1997.

~ ~ s t r ~ ~ e n t Min. pair corr. USD/DEM-USD/GBP USD/DEM-USD/NLG USD/FRF-USD/ITL USD/~G-~SD/F~ USD/FRF-USD/GBP USD/JPY-DEM/JPY USD/GBP-USD/GBP DJIA-hex DIEM 3-6M-DEM 9-12M

0.55 0.78 0.49 0.69 0.48 0.34 0.23 0.00 0.40

mean Max.

Max.

cum

poi~t (duys,s)

0.86 1.00 0.86 0.99 0.86 0.62 0.75 0.86 0.90

7.2 14.0 12.0 16.9 7.2 19.4 17.0 13.3 19.2

90%

c o r ~ e ~ a t i o ~ of (1-2 days) ~ e 0.79 0.99 0.79 0.97 0.80 0.48 0.45 0.77 0.82

0.71 0.89 0.71 0.87 0.72 0.43 0.41 0.69 0.74

value is considered as a less arbitrary (though still of correlation maximum (a true maximum value o being difficult to estimate in most cases). The data ti correlations reach90 percent of the means are also S estimation for the correlation stabilization point h it c m be uniformly applied to all cases butit does misleading sta~ilizationintervals as would be point at which correlations drop to 90 percent o value. We conclude from this data that even c highly correlated in the long term, become m o intra-hour data frequency range. The authors a h ~ p o t ~ e sof i sheterogeneous markets w their perception of the market, have differin under different insti~tionalconstraints. Thi volatilities of different time horizons to beha marketsareindeedcomposed of heter time horizons of interest, then the Epps mations couldbe interpreted as a horizon cut-off level.As t decrease below this horizon cut-off, fewer and fewer actor to take actions rapidly enough to result in instruments. The data in Table6 were examined of Eppseffectdrops in correlationand time series involved. Standard correlat

90%

point u ( ~~i ~ ~ t e s , s ) 10 15 25 25 30 30 170 320 340

122 Mark Lundin, Michel ~ a c o r o ~ and n a ~ l r i cA. h M~ller

~ta~ilization vs. Tick ~ctivity

2 4 6 I/sqrt(Al *A2) (minutes/tick)

8

Correlation stabilization points as a function of the inverse square rootof the productof tick activity rates.

individual activities involved for the instrument pairs described 6.in T ean business day activities were used for this and were taken from Table 1. The greater of the two activities for each pair was estimated to have -0.59 (standard) correlation with the 90 percent point of 1-2 day mean values (Table 6, column 4). The same quantity estimated between 90 percent points was the lesser of the two activities in each pair and the -0.65. These values are s i ~ i f i c a nto t 92 percent and95 percent confidence levels respectively, assuming a normal random distribution. Therefore we conclude that, to a reasonable level of confidence, both of the activities play a substantial role in the Epps effect drop in correlations and that th are inversely related. This can be seen graphicallyin Figure 12 where the point of 90 percent correlation drops are plotted versus the inverse squ root of the product of activities. The data points towards a stabilization point of zero in thetimeaxiswhentheactivities of bothtimeseries are infinite. However, at very low activities, a plateau in the correlation stabilization point appears to exist at a data interval of 300 to 400 minutes. This would lead to an indication that the Epps effect does not play a substantial role in attenuating correlation values beyond five six to hour data intervals. This even if the instru~entsinvolved are very inactive (r

g 20% 3

10%

5% 0%

Figure 18 Frequency distributionof extreme times for the FTSE.

considered as a local market and the latter as a global and continuous market. Volume of transactions reachesits peak on the opening of both LIFFE and CME markets. For the FTSE, investors want to adjust their positions following overnight movements in the US and Japanese stock markets. For currency contracts, the peakof volatility during the first 20 minutes is mainly due to the intense market activity caused by the openingof the Treasury bond markets and more significantly the releaseof US figures. Both factors affect the UK stock market far less. Volume grows towards the close of the LIFFE market becausethis is the last opportunity for market participants to hedge their bookor trigger an overnight position.'This is less the case on the CME markets. Currency exposure canbe still hedged after the close of the market through the 24 hours OTC market. Indeed, currency markets can be fairly thin on the close of Chicago. Bid/Ask spreads widen as the liquidity decreases. Consequently, hedging might be better achieved through the cash market either a couple of hours later on the opening of Tokyo or during the European day time. This might indeed explain the relative lack of interest of currency playersin the close of Chicago.

146 ~ r n ~ nAcar u eand ~ Robert l'oflel

It is well known that deviations with the normal assumption are num such as skewness, kurtosis and heteroskedasticity. It is however unknown to what extent such deviations might change the timing of extreme prices. This study has shown that the impactof the drift on the timingof intraday extrema is negligible. Heteroskedasticity seems to affect timing of extremes more significantly. New highs and lows are more likely to occu during volatile periods when volumes are large. New stylized facts of extreme clustering have been discovered for the CMEcurrencycontracts. A strikingquestionwould be: doesvolume cause extreme prices or vice-versa? Both hypotheses couldbe defended. One may argue that volume on CME is higher at the opening of the market because of the ex-ante higher probability of observing extreme prices. On the other hand,it could be said that rising volume may cause increased trends in futures contracts and therefore increase the probabi of observing extreme prices. Clearly, more theoretical and empirical work is needed to gain insight on the relationship between volume, volatility and extreme price.

NOTE 1. The data traded on the 24 hours Globex market has not been included in this study because of its too short history.

BIBLIOGRAPHY Acar, E.,Lequeux, P.and Ritz, S. (1996),"Timing the Highs and Lows of the Day", Life E ~ u i~roducts t~ Review, 2nd Quarter. t s , Models", Academic Press.,SanDiego. Duffie, D.(1988),"Security ~ a r ~ Stoc~astic to Probability Theory and Its Applications,', John Feller, MT. (1951), "An Intro~uctio~ Wiley & Sons. Ferger, D. (1995),"The Joint Distribution of the Running Maximum and itsLocation of D-valued Markov Processes",~ o u r n of a ~Applied Probabi~ity(34,8424345. Garman, M. and Mass, M.(1980),"On the Estimation of Security Price Volatilities from Historical Data", ~ o u r n of a ~~usiness(53),67-78.

Highs and Lows 147

Karpoff, J.M. (1987),"TheRelationshipbetweenPriceChanges and Trading Volume: A Survey", ~ o ~ r nofa Financial l and ~ ~ a n ~ Analysis i ~ a ~ (3), i ~169-176 e Muller, U.A.,Dacorogna, MM.,Olsen, R.B., Pictet, O.V., Schwarz, M. and Morgenegg,C.(1990),"Statistical Study of ForeignExchangeRates,Empirical Evidence of a Price Change Scaling Law and Intra-day Analysis", ~ o ~ r ofn ~ l ~ a n ~ i and n g Finance (14), 1189-1208.

This Page Intentionally Left Blank

This Page Intentionally Left Blank

Owain ap Gwilym, uckle and Stephen Thomas ~ e ~ a r t ~ e n t o f ~ a n aUniversity g e ~ e n t lof S o u t h ~ ~ ~ t o n l ~uro~ean ~usiness ~an Schoo~l a g e ~University ent of~ a l e s l UK,c o ~ e s ~ o n ~author ing

INTRODUCTION

Research interest in financial market microstructure has increased recently asnumeroustransactiondatasetsfordifferentmarketshavebecome available. Empirical analysisof high frequency financial market data has yielded a number of interesting statistical regularities that have proved challenging to economic theorists. Much of this evidence involves US focusing on the intraday behaviourof several aspects of market activity, such as bid-ask spreads, returns, returns volatility, traded volume and price reversals. This chapter examines the behaviour of these variables using data on UK futures and options contracts traded on the Lon International Financial Futures and Options Exchange (LIFFE) and investigates the impactof UK and US macroeconomic announceme~ts. LIFFE has grown considerably in recent years and is currently the second largest futures and options exchange in the world, having traded over 100 million contracts in the first half of 1997. The exchangeis strate~ically important in terms of m a i n t a ~ g London's position as an inte~ational financial centre. However, level the of published research into the contracts traded at the exchange does not begin to reflect this elevated PO especiallygiventhemorecompletepricedataavailablefrom compared to theUS futures exchanges. This chapter extends our understanding of intraday empirical larities in a number of ways. It offers further insight into the in behaviourobservedinthe US markets by usingEuropeandata;it *

~inancial~~~~~~sTick by Tick Edited byPierre Lequeux. 0 1999 John Wiley & Sons Ltd

compares intraday aspects of numerous futures and options contracts; and finds some notable parallels and divergences which require further explanation. The key features of high opening volatility, so prevalent h US studies of futures markets and "explained" by removin~the effects of scheduled macroeconomicanno~cements,remains in our data. This latter feature is important since many of the theories regarding trading volume andr e t u ~volatility s predictthe U-shape mentioned above, with market closure being an important factor in such models. Further, US studies cannot examine the reaction of equity futures contracts to the main US ~acroecon~mic anno~cements at 08:30 EST since the market opens for trading after the a ~ o u n c e ~ etime. nt We first discuss the relevant theoretical literature and then describe the datasets andme tho do lo^ used. "he results of empirical work on various futures and options contracts are presented at p 158, and this chapter is concluded atp 186.

This section considers the theories which aim to predict the intra-day behaviour of key market variables, and examines someof the published empirical evidence.

A number of empirical re~ularities have been identified using US data. It is now well established that stock returns and the variance of returns (across days) followa familiar U-shaped intraday pattern (Wood et al. 1985, Harris 1986). Returnson stock indices tend to be positive near both the open and close of trading,thoughHarris (1986) findsnegativereturnsnearthe Monday open. Jain and Joh (1988) also find U-shaped intraday patterns i SpSP500 index. Ekman (1992) extends hourly returns and volatility for the the analysis to the S&P500 index futures contracts,exam in in^ 15-mi~ute returns, returns volatility, the number of recorded transactions, percentage price reversals, and the ~utocorrelationof l-minute returns. The intra-day patterns in volatility and volume are approximately U-shaped, while the returns are similar in shape to those observedin NYSE index returns by Harris (1986). However, Ederington and Lee (1993) find an L-shape p

in intra-day volatility for US Treasury Rond, Eurodollar ~eutschmar and futures markets, with a peak in volatility early in the day followed b reasonably flat volatility for the restof the day. However, they find that this peak is not at the open of the market but10 minutes later, coincidin with the releaseof macroeconomic a~ouncements. Thereare a number of theoriesseektoexplaintheseintrapatterns,usuallyinvolvingtheinteractioprivatelyinformedtraders, liquidity traders and market makers. Admati~ and ~ e i ~ e(1988) r e r explain theintra-daypatternswithinme-theoreticmodelinvolving infor~e tradersandliquiditytradersoseactivitiesleadtoconcentrationsin volume and volatility atthe open and close of trading. In a similar vein, Foster and Viswanathan (1990) further develop the work of Admati an Pfleiderer (1988) to provide a game-~eoreticmodel to explain patterl~s in returns volatility both across weekdays and intra-day. In this model, liquidity traders can delay trades, and hence choose times when there islessprivateinformationandchoosesecuritiesforwhichthereis more efficient public informationpro~uction.In particular, if w e e ~ e n ~ s feature little publicinfor~ation, li~uidity traders will prefer not to trad on Mondays.

Reversals are an indicator of the level of autocorrelation in ret~rnsd The intra-day pattern for autocorrelation in returns is explained by and ~ i l ~(1985) r o in~ terms of infor~ationtradin existence of bid-ask spreads will lead to n tive serial correlation betwee successivepricechangesaspricesoscilbetweenbidandaskprice As thelevel of informationtradingincreases,thebid-askspread wi ains from in forme increase as marketmakersattempttocapture traders.Alargerspreadincreasesthepossibili of tradestakingpla between the bid and ask prices thus reducing the bid-ask bounce an effect hence increasin~the autocorrelation coefficient ( t o ~ a r d zero). s andWood(1990)findU-shapedintra-daypatternsin l-minute inde returns autocorrelation and attempt to disting~ishwhether this is due to non-synchronous tradin~or orm mat ion arrival. They favour the latter as returns variance also has a similar pattern (which they considera for information arrival). Gosnell (1995) examines the intra-day patterns of thepercentage of transactionpricereversals,which is aproxyfor transac~onsreturns autocorrelations. This rises quickly for NUSE stocks early in the day and then stabilizes, unlike the U-shaped findings for the autocorrelation of returns in McInish and Wood (1990). Gosnell (1995)

154 Owain ap Gwilym, Mike ~uckle and S~~~~ ~ k o ~ a s

examines the price change reversals pattern for individual US stocks, and finds that intra-day reversal patterns vary systematically between stocks, across each day, and betweendays. In particular,there is a lower concentration of reversals during the opening 90 minutes of trading than during the rest of the day. Evidence on intra-day transactions price change con~uationsand reversalsis important when studying the impact of new formation on security prices, andthe pattern of such variablesaround an eventisinformative about the speed of price adjustment to new infor~ation. reads

It is convenient to consider three main approaches to explaining the intra-day behaviour of bid-ask spreads (e.g. see O’Hara (1995), and Chan, and Johnson (1995)). Thefirst approach considers the institutional re of a market,in particular the degree of market powerof particiants; a single ”specialist” will have more market power than competing market-makers. Theother approachesfocus onthe role of inventories and information asymmetries.

~ t r ~ c t ~There r e . is a clear theoretical and empirical distinction etween the intra-day behaviour of bid-ask spreads in markets occupied y a monopolistic specialist (e.g. NUSE) versus those with competing ~ a r ~ e t - m a k e(e.g. r s CBOE and National Associationof Securities Dealers Automated Quotation (NASDAQ)). Brock and Kleidon (1992), ~ c ~ s and Wood (1992)’ Lee et al. (1993) and Chan, Chung and Johnson (1995) show that the spreads in the NYSE follow a U-shapedpattern throughout the day, whereas Chan, Christie and Schultz (1995) find that spreads in the NASDAQ market are relatively stable through the day but narrow si~nificantlynear the close. Chan, Chung and Johnson (1995) find that spreads for actively traded CBOE options on stock decline sharply after the open andthen level off. The differencein the intra-daybehaviour of spreads can be explained by he differentstructure of the markets.Brock and Kleidon (1992)develop a ode1 where a single specialist has monopolistic power and is faced with fairly inelastic transactions demand at the open and close of trading ue to the overnight accumulation of information prior to opening and he imediacy of the non-trading period after the close. The specialist can price discriminate during these crucial periods of inelastic demand nd can extract monopolistic profits,particularly if he has access to any opening order imbalances. Price volatility willthus ensue (see Stoll and

The rntra-

ay ~ e ~ a v i oofu rKey Market Variables 155

Whaley (1990)). Consequently, with a market served by a single specialist we would expect high price volatility, volume and bid-ask spreads near the open and close of trading. In contrast, in markets with co~peting market makers we would not expect to observe the U-shaped intra-day pattern if the specialist’s monopolistic power is an important influence on such behaviour. odels. Bid-ask spreads exist as a reward to market-makers for bearing the risk of holding inventories. Amihud and Mendelson (1980) develop a model for specialists whereby spreads are widened as inventory imbalances accumulate. Leeet al. (1993) find evidence linking spreads to inventory control costs;in particular, higher trading volume is associated with wide spreads. Consequently, the high volumes observed at the open and close of many markets (e.g. NYSE) would be accompanied by wide spreads.Hence,foraspecialiststructure,suchas NYSE, aU-shaped intraday patternof spreads is predicted as a single market maker may be forced to accumulate unwanted inventories during peak trading volumes, whereas competing market makers will be (individually) less likely to accumulatesuchpositions.Further,Chan,ChungandJohnson (1995) suggestthatspecialistsandcompetingmarketmakersmaydifferin their ability to manage imbalancesby using their bid and ask quotes; in maintaining a fair and orderly market specialists cannot execute orders on only one side of the spread, in contrast to competing market makers who can set bid and ask quotes to attract trades on only one side of the spread (and thus enhance their ability to avoid unwanted inventories). Inventory based models thus suggest that specialists will widen spreads during periodsof high volume,i.e. at the open and close. This theory does not explain the occurrence of high volumes at these times; for we thisturn to the information models. odels: the Adverse Selection Problem. A numberof recent contributions to the literature concentrate on information asymmetries, (e.g. Glosten and Milgrom (1985), AdmatiPfleiderer and (1988), and Foster and Viswanathan(1990,1994)). In these models, the following agents are activeinthemarket:marketmakers,informedtraders,andliquidity traders. In some models, a distinction is made between liquidity traders who are forced to trade at a given time of day regardless of cost and discretionary liquidity traders. The market maker is at an informational disadvantage relative to informed traders, and this is termed the adverse selection problem. Spreads must therefore be wide enough to ensure that thegainsfromtradingwiththeuninformedagentsexceedthelosses associated with trading with the informed agents. Admati and Pfleiderer (1988) predict narrow spreads when volume is high and prices are more

volatile, while Viswanathan F (1990) predict narrow spreads when volume is prices are less volatile. However, in model a f strategic trading between two asymmetrically informed traders, Foste nd Viswanathan (1994) predict high volume, high variance and wide preads near the open.

ons. Early intra-day 2)) involveddetailed andpricereversals forvariousfinancial instr~ments,buttherewas little discussion of competingtheories of theintradaypatterns.However,the distinc~on ebveenspecialistand m ~ l t i ~market le makerinstitutionalstructures asprovidedonepossiblewfimprovingourabilitytodiscriminate compare (1995) Johnson competin and etween the intra~daybehaviour of volume, volati~ityand spreads for a sample of USE stocks and their associated options traded on GBOE. the The former IS a specialist market while the latter involves multiple market makers; further, information hethe her public or private) affecting a stock will also ~ n ~ ~the e as~ociated ~ c e call and put options. ~ o n s e ~ u e n t lthe y , authors canattempttodistinshbetweenthe idluence of marketstructure, adverseselectionandentorycontrolfortheintra-daypatterns.They co~cludethathigheruncertaintyattheopening of both marketscan xplainthewidespreaobserved,butthatinformationcannotthen xplain wide stock narrow option spreads the at close. The narro~ option spre re likely to arise from differing a market tructure. iC

ere is a growing literature relating i~tra-daypatterns in futures data o theimpact of schedudmacroeconomicnews a ~ o ~ c e m e n tFor s. xarnple, Ederington Lee a (1993) find that these announcements observedtime-of-dayandday-of-theterest rate and foreign currency futures arkets. In particula that returns volatility early the inday, ound in studies SUC (1992), does not occur at the opening of rading if more fine egated data is used (at 5-min~teinterccur during the third 5-minute interval with macroeconomic ~ o ~ c e m e n t s Removing the effect of these announcets is found to remove most of this volatility.

The Intra-day~ e h a ~ of Key i ~ ~~f l rr k E~t f l r i a ~ l e157 s

A number of papers document the volatility of UK futures prices around UK and US macroeconomica~ouncements.Becker et al. (1993,1995) find thatreturnsvolatilityintheFTSE100andLongGiltfuturesmarkets responds to the releaseof economic data in theUK and US. This chapter extends this to cover intra-day returns, volume and reversals.

The futures data used consists of every quote, transaction and associated volume for the UK stock index (FTSElOO), 3-month interest rate (Short Sterling), and government bond (Long Gilt) futures contracts. The data is potentially far more informative than that used in many previous U studies, because "time and sales" data from US futures exchange Chicago Mercantile Exchange (CME) and Chicago Board of Trade (C only contain bid and ask quotes if the bid quote exceeds iforthe ask quote is below the previously recorded transactionprice. Also, trades are only recorded if they involvea change in price from the last trade. The futures contracts examined are primarily traded by followed by evening screen trading sessions. The floor tradin Long Gilt futures market was08:30-16:15 GMT until July31 opening was brought forward to08:OO GMT. Floor trading times for FTSE100 and Short Sterling are 08:35-16:lO GMT and 08:05-16~05G respectively. The analysis focuses on the most he ily tradedcontracts,For the Long Gilt and Short Sterling contracts, trad ends to be concentrated inthenearestexpirymonthuntiltheturn e expirymonth, e. Marchcontractstrademostheavilyuntiltheend of Februarywhen trading switches to the June contract. For FTSElOO the contract, trading is concentrated in the front month contract right up to the expiry date. Index options are traded by open outcry at LIFFE from 08:35-16:lO GMT. "he optionsdatasampleconsists of alltradesandquotesfor American-style FTSE100 index options at LIFFEfrom January 4, 1993 to March 31,1994. The database contains the time of the quote, exercise style, expiry month, exercise price, call or put, matched bid and ask quotes, and the current level of the underlying asset. Only at-the-money and n e a r - ~ a ~ r icontracts ty are considered here. Intradaybehaviour is examined by partition in^ thetradingdayas appropriate for the frequency of the data and calculating mean values for each interval across the days in the sample. In order to examine the statistical significance of any observed intra-day patterns, regressionsof

158 Owain ap Gwilywz,Mike Buckle and Stephen ~homas

the following form are estimated: k

1

t=l

where Xi,i,n is the value taken by the variable of interest i (i = return, returnvolatility,tradedvolume,pricereversalor bid-ask spread)for theinterval j onday n, and (0,l) dummyvariablesareincludedas necessary for the k intervals of the day and/or days of the week under ~vestigation.A positive(negative)coefficientonone of thedummy variables indicates higher (lower) levels of the variableof interest during that interval (day) than the average across the rest of the day (week). The ~stimationuses Hansen's (1982) Generalized Method of Moments (GMM) to ensure robustness to returns autocorrelation and heteroscedastic errors, 0th of which are common in this type of data. Since the system is just identified, theGMM estimates are identical to those from Ordinary Least quares (OLS) although their standard errors differ. This methodology of is commonlyused in the US literatureontheintra-daybehaviour financial markets, e.g. Chan, Christie and Schultz (1995) and Chan, Chung and Johnson (1995).

E~PIRICAL EVIDENCE

Theempiricalevidenceispresentedasfollows.Thebehaviour of returns,volatility,volumeandreversalsfortheFTSE100andShort terlingfuturescontracts is examined at p159.Thefollowingsection resents evidence on intra-day volatility and volume in the Long Gilt futures market. Both sections consider the impact of scheduled UK and US macroeconomic announcements The following nine scheduled UK ~acroeconomic announcements are included: the Public Sector Borrowi equirement, Labour Market Statistics, Retail Price Index, Retail Sales, Gross Domestic Product, Balance of Payments, Confederation of British Industry Surveyof Industrial Trends, Producer Price Index, and Moneta Statistics. These have been chosen largely to reflect our a priori view of which a~ouncementsare most likely to move the futures markets. Thes US datawhichisreleasedat variablesbroadlycorrespondwiththe 0 ~ : 3EST 0 as documented in Ederington and Lee (1993), and these are the US announcements we consider. The intra-day behaviourof bid-ask spreads, returns and volatility for FTSE100 stock indexis discussed options at p 175.

The Intra-day Behaviourof Key Market Vflriables 159

our of the FTSE100 Stock Index and Short Futures ont tracts

Formoredetailontheanalysisinthissub-section,seeBuckle et al. (1998).Transactionsdataisusedforthe period fromNovember 1, 1992 to October 31, 1993. Returns, returns volatility, traded volume and percentage of price reversals are examined over 5-minute intervals. All the UK anno~cementsconsidered in this section are released at 11:30 GMT.132In addition to these regular announce~ents,we also examined theeffect of interestratechangesonthemarkets.Duringthesample period there were two UK base rate changes on November12,1992 and January 26,1993, both of which were announced at 09:50 GMT. Further, we examine how market closure and different days of the week impact on the variablesof interest. Returns and~olatility for FTSEIOO and Short Sterling Futures ~ont~acts. Returns are calculated as the logarithm of the last transactions price in the current 5-minute interval minus the logarithm of the last transactions price in the preceding interval. The results for the 5-minute mean returns for the Short Sterling and FTSElOO contracts are reported in Figures 1 and 2. The theoretical literature generally has little to say about intra-day returns behaviour. Figures 1 and 2 show very noisy patterns in returns with no clear intra-day behaviour emerging. Mean returns at the open and close are negative for Short Sterling while positive for the FTSE100 contract. Positive and relatively large returns are found at the timing of interest rate changes (09:55 GMT), but no clear pattern occurs at times of UK nor US macroeconomic anno~cements(11:35 and13:35 GMT). Table 1 presents the resultsof regression Equation1 for returns, which areconsistentwiththeobservationsfromtheFigures. No significant coefficients are reported for the intra-day effects for either contract, though a large positive coefficient exists on the dummy for interest rate changes. In termsof days of the week, the only significant coefficient is for Mondays for Short Sterling and it is negative. This indicates that returns on the Short Sterling contract tend tobe lower on Mondays than the mean across the rest of the week. Returns volatilityis initially calculated as the standard deviation of log returns across days for a given interval and the intra-day patterns are presented in Figures 3 and 4. It is observed from these figures that both the Short Sterling and FTSElOO returns are highly variable at the open but less so at the close. However, for US bond, interest rate, and exchange rate futures contracts, Ederington and Lee (1993) find that it is the third 5-minute intervalof the day (when US announcements are released) that is associated with high volatility rather than the opening.

m 0 0

0

g

E;

z

0 0

0

g

0 0 0

m

0

0 0 0

O

m

160 Owain a p Gwilym, Mike ~ ~ c kand l e S t ~ T ~ o ~e z u~ s

0 0 0

z

T-

0

0 0

E

0 0

(U

0 0 0

9 ?

N 0 0

z 0

zE

zB

U')

0 0 0

g

0

5;

0

ct 4

E

4

"

0

c!

8 4

N

The rntra-~ayBehaviour of Key Market ~ariables 161

U')

9

0

0 0 0

162 Owain ap Gwilym, Mike Buckle and S t ~ ~Thomas e n

The table presents the results of the following regression: where R],. is the return during interval j on day n, D1 is equal to 1 during the opening interval of the day and 0 otherwise, D2 equals 1 during the interval 09:50-09:55 Gh4T on days of UK interest rate changes and 0 otherwise, D3 equals 1 during the interval 11:30-11:35 GMT on days of scheduled UK macroeconomic announcements and 0 otherwise, D4 equals 1 duringtheinterval13:30-13:35GMTondays of scheduledmajor US macroeconomic ~ o u n c e m e n t sand 0 otherwise, and D5 equals 1 in the closing interval of the day and 0 otherwise. For daysof the week, D6 equals 1 on Mondays and 0 otherwise, D7 equals 1 on Tuesdays and 0 otherwise, D8 equals 1 on Thursdays and 0 otherwise, and D9 equals 1 on by lo6. Fridays and0 otherwise. All coefficients are multiplied -~

Short Sterling Coe~cient ~t-stutistic~ Coe~cient

--0.67(-0.72) -4.52(-0.36) 2026.88( 1.49) 34.65(0.89) 6.69(0.55) -12.30(-1.75) -2.77(-2.08)b 1.47(0.95) -0.88(-0.65) 1.47(1.10)

FTS~ZOO

~t-stutistic~ -2.28(-0.21) 133.23(1.41) 6611.86( 1.59) 106.54(0,75) -49.25(-0.41) 70.58(1.29) 20.47(1.31) --5.82(-0.38) -2.34(-0.16) 7.66(0.50)

There are two further clear spikes in Short Sterling volatility. The first peakishigherthantheopeningvolatilityandoccursattheinterval 9:50-9:55 GMT. It is driven by anunexpectedinterestratechange on January 26, 1993.3 The second occurs over the interval 11:30-11:35 GMT, the timing for UK macroeconomic data releases. For the FTSE100 contract, peaks in volatility at the above times are also apparentbut not as pronounced. There is a third peak at the timing of US macroeconomic announcements (13:30-13:35 GMT), and it therefore appears that these have more impact on the stock index futures volatility than the interest rate futures volatility. In Figures 3 and 4 there is a crude L-shaped pattern in volatility i.e. higher volatility at the open and close than through the day, though less pronouncedat the close. This is in contrast to the U-shaped pattern t been well documented for stock markets, e.g. Harris(19861, Admati and

The Intra-day Behaviourof Key Market Variables 163

0

8

0

8

c-

0

x

8

CO 0 0

a s

0

8

(v

0

8

0

0

8

0

0

U)

164 Owa~n ap Gwilym, Mike Buckle and S t ~ ~ ~~ eo ~~

0

8

0

The lntra- ay ~ehaviourof Key Mnrket VariabZes 165

The table presents the results of the following regression: where AXi,. is the absolute value of the difference between the actual return Ri,. for the 5 minute interval j on day n and the mean return R; for interval j across all the days in the sample. Dt are defined asat Table 1. A11 coefficients multipliedby lo6. Sterling Short

FTSElOO

C o e ~ c i e n(t-stutistic) t

Coe~cient ~t-stu~istic)

46.96(23.38)a 76.59(7.91)” 1969.41(1.45) 153.91(5.15)a 11.71(1.26) 35.00(7.66)a -7.00(-2.28)b -2.74(-0.91) 5.00(1.66) 2.24(0.72)

522.70(37.01)a 672.85(11.80)a 5972.30(1.43) 394.19(4.45)a 147.59(1.69) 138.55(3.95)a -46.19( -2.24)b -45.23(-2.63)b -1.77(-0.10) -0.94(-0.05)

Pfleiderer (1988), Berry and Howe (1994). However, it is similar to Ekman (1992) who finds for the S&P500 futures contract that volatility is three two times higher at the close. times higher than average at the open and Table 2 presents regression tests of volatility based on E~uation1.In contrast to the returns results, some statistically significant patterns emerge which are broadly consistent for both contracts. Both Short Sterling and the FTSE100 contract have significantly higher volatility at the market open (Dl) and close (D5). However, the size of the open coefficient is larger than the close coefficient, confirming the pattern observed in Figures 3 and 4. For the news release dummies, the only one which is statistically significan~is at the 11:30-11:35 GMT UK macroeconomic a ~ o ~ c e m e n t time. The other a~ouncementcoefficients are positive but ~ s i ~ i f i c a n t at the5 percent level. Volatility on Mondays is significantly lower in both markets, and Tuesday volatility is also lower than theofrest the week for the FTSE100 case. edVolume for F ~ S E ~ O and O ShortSterlingFutures on tract§. Figures 5 and 6 report the mean traded volume both in contracts over each interval. Thereis a U-shaped pattern for bothbut, like volatility, it is not

5:

0

3

0

c7

0 0

0 0 N

166 Owain ap Gwilym, Mike Buckle and S t ~ ~ h ~ ~o m ~ s

U)

0 0

g

0

The Intra-day ~ e h a ~ i o uofrKey Market Va~iables 167

symmetric in that volume is higher at the start than theofend the day. The high traded volume at the open is consistent with Berry and Howe (19 who find a positive relationship between public information and trading volume. In Figure 5 for Short Sterling, the main deviation from a relativ smooth pattern is a definite peak at 11:30-11:35 GMT, the timing of UK announcements. Thereis a muchless pronounced peak at this time for th FTSE100 contract in Figure 6. There are no obvious peaks in volume at the times of interest rate changes norUS macroeconomic a~ouncementsfor either contract. Table 3 presents the results of applying Equation1 to the volume data, which are consistent with Figures 5 and 6. Both contracts show similar patterns with significantly higher volume at the open and close and the interval 11:30-11:35 GMT. Volume is not significantly higher of on interest rate changes and is lower atof times US macroeconomic announcements (significantly lower for Short Sterling). Volume in both markets is significantly lower on Mondays than across the of the restweek, which is a common result for stock markets, and is consistent with the predictions of Foster and Viswanathan (1990). This corresponds with significantly low volatility on Mondays as reported in the previous section. Hence the intra-day pattern in both volatility and volume are essentiall similar to each other, with higher levels at the open close, and but higher Table 3 Regression testsof the intra-day behaviourof traded volume in Short Sterling and FTSE100 futures contracts. 9

The table presents the results of the following regression: j on dayn, and De are definedas at Table 1. where Vi,. is traded volume during interval Short Sterling

F~S~200

C o e ~ c i e ~ (~~ - s ~ ~ t i s ~ i c~) o e ~(t-st~tistic~ cie~t

102.97(12.10)a 398.21(19.33)a 310.52(1.29) 356.57(4.88)” -27.43(-2.15)b 77.80(5.94)” -25.35(-2.48)b -10.77(-1.00) 8.79(0.76) -1.17(-0.10)

68.49(17.16)a 205.37(10.38)a 53.59(0.77) 58.01(5.11)” -10.68(-1.36) 82.39(14.68)a -13.70(-2.69)a -3.69(-0.71) -2.47(-0.51) -1.66(-0.32)

The Intra-day ~ehaviouro f K q Market ~ a r i ~ b l e169 s

at the open than the close. This contrasts with the more symmetric pattern found for these two variables for stock markets is similar but to that found by Ekman (1992) for the S&P500 futures market. The positive correlation between volume and volatility suggested in the above results is consistent withempiricalstudies of thisrelationship,assummarizedinKarpoff (1987). The results are also consistent with Admati and Pfleiderer (1988) with volatility and volume being higher at open and close compared to the rest of the day. "he differing intra-day pattern compared to some stock market evidence may reflect different trading arrangements. One hportant difference is that trading at LIFFE and major US futures markets occurs in pits with a number of traders known as scalpers present. Scalpers trade a small number of contracts, holding positions for very short periods of time (often just a few minutes) and generally do not hold positions overnight. High volatilityat the close in futures markets could therefore be explained by scalpers and day traders closing out their positions before the close. eversals forFTS~lOOand Short Sterling Futures ~ o ~ t r a cA t sprice . reversal is defined as a price change that in is the opposite direction to the A price change in the same direction as the previous previous price change. price changeis termed acontin~ation. For consistency with US studies, the data examined in this section uses price change transactions only, i.e. all non-price change transactions were excluded. Thus, every observation is either a reversal or continua~on. a Reversal price changes are generally a of buy and sell orders that are executed consequence of the random arrival at stationary bid and ask prices. The price change transaction series will then display a sequence of up and down movements between the bid and ask prices. ~ontinuations,on the other hand, are usually associated with information arriving into the market that causes an imbalance between supply and demand for the security at the currentprice level. Traders will then revise their view of the value of the security and if this new value lies outside the current bid-ask spread there will be an imbalance of buy and sell orders until the spread changes to capture the new security value. Under these conditions, we will observe a series of continuation price changes as the price adjusts to its new e q u i l i b r i ~Thus, . a low level of reversals (i.e. a high levelof continuations) is typically a consequenceof new information arriving into the market. For this analysis we calculate the percentage of price change transactions which are reversals using 15-minute intervals across the day. The interval 7 and means of percentage reversals are displayed graphically in Figures 8. Short Sterling reversals vary between 73 percent and100 percent, while those for theFTSE100 contract range between64 percent and 72 percent.

170 Owain ap Gwilym, Mike Buckle and S ~

omas~as ~

" -

n

The lntra- ay ~ehaviouro f K q Market V a r ~ ~ l e171 s

172 Owain ap Gwilym, Mike Buckle and S t ~ ~Te~no ~ a s

The higher level of reversals for Short Sterling may reflect relatively few information events affecting that market. Figure 7 for Short Sterling shows a very noisy plot for reversals, w clear pattern emerging. Reversals are relatively lowat the open, suggestive of information arrival,and relatively high at the close. The lowest point 0950 GMT, the timing of interest in the plot is for the interval following rate changes. There are many intervals with a very high level of reversals, reflecting a predominance of prices bouncing between bidand ask and 8 shows relatively few information events to cause continuations. Figure a different pattern for the FTSE100 contract with reversals lower level. Consistent with Short Sterling, reversals are relatively atlow the openand relatively highat the close. Reversals are also relatively low from 11:30-14:OO GMT, suggestive of information arrival. Table 4 presents regression results from Equation 1 for reversals. For Short Sterling, reversals aresi~ificantlylower than average at the open. Table 4 Regression testsof the intra-day behaviourof price reversals for Short Sterling and FTSE100 futures contracts. REVj,, = U

+

x 9

t=l

PtDt

+ cj,n

The table presents the results of the following regression: where REVj,n is the percentage of price changes which are reversals during the 15-minute interval j on day n. D1 is equal to 1 during the opening interval of the day and 0 otherwise, D2 equals 1 during the interval from 09:50 GMT on days of UK interest rate changes and 0 otherwise, D3 equals 1 during the interval from 1k30 GMT on days of scheduled UK macroeconomic ~ o ~ c e m e nand t s 0 otherwise, D4 equals 1 during the interval from 13:30 GMT on daysof scheduled majorUS macroeconomic announcements and0 otherwise, and D5 equals 1 in the closing intervalof the day and 0 otherwise. For daysof the week,D6 to D9 are defined as at Table 1.

Short Sterling

F~S~~OO

Coe~cient (t-stafistic)

Coe~cient(t-statistic)

0.89(57.09)a -0.06(-2.87)a -0.41(-1.65) -0.07(-1.48) 0.03(0.31) 0.03(1.03) 0.03(1.45) 0.02(0.95) O.Ol(0.32) 0.02(0.77)

0.69(97.45)a -0.02 (-2.16)b -0.40(-1.95) -0.04(--1.57) -0.08(-1.99)b 0.03(2.92)a -0.01(-1.35) O.Ol(1.21) -O.OO(-0.03) -0.01(-1.11)

The Intra-day~ e h a v i oo~ f Key r Market Variables 173

Despite a large negative coefficient on the interest rate dummy, it is only significant at the 10 percent level. All other coefficients on the dummy variables are insignificantly different from zero. For the FTSE100 contract, reversals are again significantly lower at the open. Reversals are lower than average at each announcement time, but only significantly so at 13:35 GMT (US a~ouncements).Further, there is a significantly higher level of reversals at the close. As for Short Sterling, no day of the week effects appear for reversals. The significantly lower level of reversals at the open is suggestive of formation arrival and supports the results of previous sections showing higher volume and volatility at the open. The higher level of reversals at the close (significant for the FTSE100 contract), suggests that higher volume and volatility at the close is not a result of information arrival. This supportsthe hypothesis in the previous section that scalpers and day traders are an important factor in the behaviour of these variables near the closeof futures markets.

In this section, data from the period January 24, 1992 to July 31, 1994 is used to measure volatility and volume at 5-minute intervals. The end of the sample is determined aby change in opening hours at that date. In this sample there are93 intra-day intervals for 614 days (57,102 observations). The analysis covers a period when the release timeUK formacroeconomic a n n o ~ c e m e ~changed ts from 11:30 GMT to 09:30 GMT (as discussed in endnote l), therefore both times need to be examined. For analysis of returns, Volatility, volume and autocorrelation for the German and Italian bond futures contractsat LIFFE, see apGwilymet al. (1996). Figure 9 presents the volatility and volume patterns forthe Long Gilt contract. Notable peaks occur at 11:30-11:35 GMT and 13:30-13:35 GMT. In Table5, the regression results for volatility show a significant coefficient for each of the d u m y variables, indicating that volatility is s i ~ i f i c ~ t l y higher for the opening and closing intervals of the day and for the intervals of UK and US economic data releases. Table 5 also presents the results for Equation 1 for traded volume. Significantly higher traded volume is observed near the open and close, and following US data releases and UK announcements at 09:30GMT, The UK a~ouncementsat 11:30 GMT do not induce significantly higher traded volume in the following five a positive correlation between volatility and minutes. These results suggest volume across the trading day, although macroeconomic a~ouncements seem to havea greater effect on volatility than on volume.

.I.

-r

8

0 0

9 0

(

0

0

0

0

D

0 0 a3

V

)

0

00 8

0 0 U3

0 0

0

3

: 0

0

0 0 d u q a J aqnlosqe ueayU

s

r.

h .

I

174 Owuin up Gwilym, Mike Buckle and S ~ ~ ~~~~s ~ e n

0 0

2

m 0 0

9 0

“ . ,

N

0 0

0

0

% S 0; 0

0 0

0

0

The Intra-day ~ e ~ v i o ofu Key r Market Varia~les 175

The table presents the results of the following regression: j on day n, D1 where Xj,. is the absolute return or traded volume for the 5-minute interval is a dummy variable taking the value1during the first intervalof the day and 0 otherwise, D2 takes the value 1 during the 09:30-09:35 GMT interval and 0 otherwise, D3 takes the value 1 during the 11:30-11:35 GMT interval and 0 otherwise, D4 takes the value 1 during the 13:30-13:35 GMT interval and 0 otherwise, and D5 takes the value 1 during the last lo* in the caseof absolute interval of the day and0 otherwise. Coefficients are multiplied by returns (volatility). " _

Volatilit~ Coeficient t-stat

i5. D1 D2 D3

D4 D5

3.61 3.21 3.73 0.61 0.98 1.54

164.3a ll.la 5.7a 2.l b 4.3a 5.9a

V o l ~ ~ e Coe~cient

303.30 790.03 333.79 19.33 48.48 255.57

t-stat

122.1a 23.1a 7.9a 1.o 2.6a 14.3a

The Intra-day Behaviour of id-ask Spreads, Returns, and Volatility for FTSElOO Stock Index Options

The optionsdata sample used is described at p 157. Formore detail on the analysis in this section, see ap Gwilym et al. (1997). Bid-ask Spreadson FTSElOO IndexOptions. Figure 10 presents the intra-day pattern in mean absoluteand percentage spreads for 15-minute intervals across alldays for at-the-money, near-maturity contracts. Both calls and puts demonstrate higher levels of absolute spreads at the market open which quickly decrease to the level around which they fluctuate for much of the rest of the day. However, there is also a fallin the level of spreads towards the end of the day. Percentage spreads for calls follow the pattern in absolute spreads but this is not the case for puts? This pattern in spreads contrasts with the U-shaped pattern across the day which has been frequently documented in previous research. However, Chan, Christie and Schultz (1995) and Chan, Chung and Johnson (1995) report decreasing spreads nearthecloseinmarkets with competing

176 Owain ap Gwilyrn, Mike Buckle and S ~ ~ horna ~ eas n

U)

m

_. _ .

0

U)

ri

m

The Intra-day ~ e h a v i oof~ rKey Market ~ a r i a b ~ 177 e~

1'78 Owainap Gwilym, Mike Buckle andS t ~ h e ~n h o ~ s

market-makers, and suggest that market structure influences the intra-day behaviour of spreads. Regression results for spreads appear in Table 6, where forboth calls and in interval 1,i n s i ~ f i c a n t puts, the absolute spread is significantly positive in interval 2 and significantly negativein the final two intervals of the day. This observationof significantly widerspreads near the marketopen and significantly narrowerspreads near the close is consistentwith Figure 10, The narrower spreads at the close contrasts with the U-shape reported NYSE. While the specialist at the NYSE may use for markets such as the its monopolist position to widen spreads to take advantage of inelastic demand near the close, market makers at LIFFE may quote aggressively on one side of the spread as a means of controlling overnight inventory. For further analysis of the bid-ask spread on FTSElOO index options, see ap Cwilym et al. (1998). Table 6 Regression results for the bid-ask spread on FTSE100 index options. The table presents the results of the following regression: Si,. = a

+

4

~

t

+si,.~

t

t= l

where Si,. is the absolute spread for intervalj on day n, Dl is a dummy variable taking the value 1 during the first 15-minute interval of the day and 0 otherwise, D2 takes the value 1 during the second 15-111inute interval of the day and 0 otherwise, D3 takes the value 1 during the penultimate 15-minute interval of the day and 0 otherwise, and D4 takes the value 1 during the final 15-minute interval of the day and 0 otherwise. A significant positive coefficient on a dummy variable indicates wider spreads than average occur during that interval and a significant negative coefficient indicates narrower spreads than average. Th are 19,181 observations for calls and 17,869 for puts.

CALLS

Ll

D1 (0835-0855)

D2 (0855-0910) D3 (1540-1555) D4 (1555-1610)

PUTS Ll

D1 (0835-0855) D2 (0855-0910) D3 (1540-1555) D4

(1555-1610)

asignificant at1%

3.209 0.326 -0.082 -0.175 -0.235

232.58a 7.72a -1.68 -3.08a -5.01a

3.211 0.365 0.021 -0.190 -0.312

213.46a 7.52a 0.37 -3.21a -6.21a

The Intra-day~ e h a v i oof~ Key r Market Variables 1'79

etums on FT~ElOOIndex ptions. For options, the returns calculation must useprices from contracts with identical maturity, exercise price, and exercise style (American or European for LIFFE index options), and both must be either calls or puts. Even with high-frequency data, this imposes a restriction on the time interval over which returns can be calculated. Tracking observations which satisfy the above conditions at a suitably high frequency requires the use of at-the-money, near-maturi~contract^.^ We conduct analysis of hourly returns (as in Sheikh and Ronn (1994) for US options) calculated from the midpoints of bid and ask quotes. The trading day was split as follows: Interval 1: 08:35-09:10, Intervals 2-43: Hourly from 09:ll to 16:lO. Intervals of equal length back from the close are used, with the remaining time at the open treated as the overnight return. The of that price for an interval is takenas the last observation before the end interval. A further issue is encountered when the exercise price which is at-the-money changes between intervals. A return based on contracts with differing exercise prices is meaningless, thereforeit is also necessary to track the previously at-the-money contractq6 For returns, Figure 11presents our results for calls, and mean returns do not show a consistent pattern across days. The mean overnight return is negative across all days and for each day except Tuesday. A larger negative mean return on Monday open is notable. Figure l 2 presents the results for puts, and a lack of distinct patterns is again observed. However, mean overnight returns are negative and relatively large forofeach the week. day Table 7 presents the coefficients for returns based on Equation 1. The results for calls demonstrate no significant(at the 5 percent level) coefficients on the dummy variables. This indicates that the call returns near the open and close are not significantly different from returns across the rest of the day, whichis consistent with Figure11. For puts, insignificant coefficients appear for intervals 2,7 and 8, but a highly significant negative coefficient is observed for the opening interval of the day. This demonstrates that the consistently negative overnight returns in Figure 12 are significantly lower than those during the rest of the day. The reason for consistently negative overnight put returns is not clear, though we now offer some suggestions. The time of options value naturally decreases as time passes and this be could a contributing factor, as calls also exhibit smaller negative overnight returns for each day except Tuesdays. Due to the derivative natureof options, if the underlying marketis rising call prices will tend to increase while puts will tend to lose value. Over the sample period used, the underlying marketby rose 7.6 percent, which may indicate that the fair expected overnight put return would tend be to negative,The result could also be due to a mispricing in the options market whereby puts lost money because they were systematically overpriced.

180 Owain ap Gwilyrn, Mike Buckle and S ~ e p h ~ nh o ~ a s Mondays 0.04

0.03 0.02 0.01

f

0.00

"-

-0.01

f

-0.02

. " " "

" "

-0.03 -0.04

-0.05

--

Interval

1-Mean

return =

Mean absolute return

l

Wednesdays

0.04

0.16

0.03

0.14

0.02

0.12

55

O.O1 0.00

5

0.10 e! pt 0.08

-0.01

3 5.02

0.06 0.04

-0.04

0.02

0.00

-0.05

3

1

4 return =

-Mean

-

Interval

5

6

7

8

= Mean absolute return

Fridays

0.04

0.16

0.03 0.02

g

0.12

0.01

0.10

3

-0.01

5

0.14 E

0.00

0.08

"0.02

0.06

-0.03

0.04

-0.04

0.02 0.00

-0.05 2

g E

-0.03

2

fj %

3

1

4

-

5

7

6

interval

I-Mean

return = =

Mean absolute return

I

8

2

Tuesdays

0.04

0.16

0.03

0.14

0.02

0.12

E 0.01

g

5

3

0.00

0.08

-0.01

0.04

-0.03

C

3

0.02

-0.04 2

1

4

3

Interval

6

5

8

7

Thursdays

0.04 0.02

0.14 0.12

~

E 0.01

0.10

-$ 0.00

$ E!

er

0.08

-0.01

0.06

0.04

-0.03

2 C

0.02

-0.04 -0.05

0.00

0.16

0.03

5

i5

0.06 2

-0.02

-0.05

i j

0.10 e!

2

1

3

-Mean return

4

5

Interval

0.00

6

8

7

= = = Mean absolute return

All Days

0.16 0.14 E 0.12 i3 0.10

0.08 0.06

1

2

3

4

5

6

7

8

0.00

interval -Meanreturn

-- -

Mean absolute return

1

Figure 11. Meanreturnsandmeanabsolutereturnsforat-the-money, near-maturity American-styleFTSElOO index call options.

e!

i

Owain ap Gwilym,Mike Buckle andS t e p ~ Thomas e~

Mondays 0.04

0.18 0.16

0.02

0.14

0

0.12 0.10

-0.02

E

3

0.08

-0.04

0.06

0.04

-0.06

c

0.02

-0.08 2

0.00 3

1

4

5

--

7

6

8

Interval

I-Mean

return

= Mean absolute return

I

Wednesdavs 0.04

0.16 0.14

0.02

$

0.12

0

E

3

0.10

-0.02

0.08

m

0.06

0.04 -0.06

< 3

0.02

l

-0.08 I 1

2

3

5

4

-

7

6

0.00

8

Interval Mean return = =

Mean absolute return

I

Fridays 0.16

0.14 0.12

0.08 " " " "

0.06

0.00 3

1

4

5

--

7

6

Interval

I-Mean

E

0.10 8

_""""" 2

E

3

return =

Mean absolute return

I

8