`Bureau of Economic Research
`
`Volume Title: The Microstructure of Foreign Exchange Markets
`
`Volume Author/Editor: Jeffrey A. Frankel, Giampaolo Galli, Alberto
`Giovannini, editors
`
`Volume Publisher: University of Chicago Press
`
`Volume ISBN: 0-226-26000-3
`
`Volume URL: http://www.nber.org/books/fran96-1
`
`Conference Date: July 1-2, 1994
`
`Publication Date: January 1996
`
`Chapter Title: One Day in June 1993: A Study of the Working of the
`Reuters 2000-2 Electronic Foreign Exchange Trading System
`
`Chapter Author: Charles Goodhart, Takatoshi Ito, Richard Payne
`
`Chapter URL: http://www.nber.org/chapters/c11364
`
`Chapter pages in book: (p. 107 - 182)
`
`GAIN CAPITAL - EXHIBIT 1016
`
`
`
`One Day in June 1993: A Study
`of the Working of the Reuters
`2000-2 Electronic Foreign
`Exchange Trading System
`
`Charles Goodhart, Takatoshi Ito, and Richard Payne
`
`4.1
`
`Introduction
`
`This is a study of foreign exchange dealers' behavior as revealed in the
`working of Reuters 2000-2, a recently developed electronic foreign exchange
`trading system. It was launched in 1992 with twenty-three subscriber sites in
`two countries and by September 1993 had more than 230 dealing sites in
`twenty-eight cities in seventeen countries (Blitz 1993). The working of the
`system is described in more detail in section 4.2. This dealing system 2000-2
`(henceforward termed D2000-2) is, however, still at the developing rather than
`a mature stage, and the snapshot that we have of its operations on one day—
`
`Charles Goodhart is the Norman Sosnow Professor of Banking and Finance and deputy director
`of the Financial Markets Group at the London School of Economics. Takatoshi Ito is professor
`of economics at Hitotsubashi University and senior advisor of the Research Department at the
`International Monetary Fund. Richard Payne is a Ph.D. student at the London School of Econom-
`ics and a research assistant at the Financial Markets Group.
`This lengthy empirical exercise was conducted in a number of stages. After one of the authors,
`C. Goodhart, had obtained the original videotapes from Reuters, to whom we are most grateful,
`the data on the tapes were transcribed onto paper by two of the authors' wives, Mrs. Goodhart and
`Mrs. Ito, assisted by Yoko Miyao, a painstaking task beyond and above the normal requirements
`of matrimony. The data were then sorted and organized by T. Ito and R. Payne, separately in the
`United States and the United Kingdom. The graphic appendix is entirely Ito's work. The descrip-
`tive material in sections 4.1 and 4.2 was mostly written by Goodhart. The comparison of D2000-2
`and FXFX in section 4.3 had input from all authors, but mostly Goodhart and Payne. The compara-
`ble FXFX data were obtained from Olsen and Associates, to whom we are most grateful. Only the
`first three sections were ready in time for the July Perugia conference, so this is all that our discus-
`sants, to whom we are most grateful, then had before them. Section 4.4, completed thereafter, was
`entirely the work of Goodhart and Payne, with Payne responsible for the econometrics, apart from
`table 4.16 by Ito. Charles Goodhart and Richard Payne wish to thank the Economic and Social
`Research Council for financial support. Takatoshi Ito thanks Charles Kramer for technical assis-
`tance in producing the graphic appendix.
`
`107
`
`
`
`108
`
`Charles Goodhart, Takatoshi Ito, and Richard Payne
`
`16 June 1993—may have become outdated and obsolete by the time that this
`is published.1
`Reuters has become subject to competition in this marketplace, from Minex
`and from the Electronic Broking Service (EBS). The former was established
`in April 1993 by Japanese institutions and, according to Blitz (1993), is "much
`used in Asia," although, as of September 1993, it did not reveal the number of
`trades crossed or terminals used. EBS was founded on Wednesday, 21 Septem-
`ber 1993. It cost, again according to Blitz, around £40 million to launch and
`has been backed by a dozen leading banks in foreign exchange—such as Citi-
`bank and Chase Manhattan—who formed a consortium with Quotron, an elec-
`tronic information screen competitor with Reuters.
`In September 1993, Bob Etherington, Reuters' international marketing man-
`ager, would not reveal his dealing system's current volume levels, although
`Blitz (1993) did report that the "system has reached [its] initial target of 1,000
`trades a day, each for a minimum 1 million units of currency dealt."2 As noted,
`Minex was not then disclosing the number of trades, and EBS had not started
`but was going to invite dealers "to trade in standard amounts of $5 million in
`Dm/$ and £5 million in £/Dm."
`Such electronic dealing systems (as contrasted with informational pages
`supplying indicative bid-ask quotes, such as the Reuters FXFX page) are still
`in their early stages and are highly competitive. Moreover, they may have an
`important future: "Roughly 60 per cent of deals in the currency market are
`now done by traders in two banks—or counterparties—who call one another
`up directly. The remainder of deals are done through brokers, who bring to-
`gether diverse buyers and sellers.. .. But they [the banks] complain that the
`commissions charged for broking a deal are very high. Automated brokerage
`terminals do the same job as humans at a reduced cost... . The banks are
`attracted by the reduced cost of commission. But they fear that 2000-2 will
`help monopolize the market in electronic dealing systems. Mr. Bartko [chair-
`man of the EBS partnership] admits that this is one of the principal motives
`for this week's launch of EBS" (Blitz 1993).
`Electronic trading systems have been in use for rather longer in other finan-
`cial markets, notably in standardized futures and options markets. Instinet and
`Globex are two such that Reuters has again been developing. A useful taxon-
`omy of the modus operandi of such electronic trading systems has been pro-
`vided by Domowitz (1990, 1993).
`
`1. Readers wanting more up-to-date information should refer directly to Reuters Limited, 85
`Fleet Street, London EC4P 4AJ, United Kingdom.
`2. The total amount thus traded is large in absolute amount but small relative to reported daily
`turnover in this market of some $900 billion or more. We find it hard to relate the data reported
`above to the BIS (1993) report in their 1992 survey that, "in the United States and the United
`Kingdom, the share of deals going through such [automated dealing] systems in April 1992 was
`32 and 24% respectively" (table 1, p. 21, and p. 24). Probably definitions of automated dealing
`systems would have been somewhat wider, including Reuters D2000-1 as well as D2000-2, but,
`even so, the above percentage seems surprisingly high.
`
`
`
`109
`
`A Study of the Reuters D2000-2 Dealing System
`
`Under these circumstances, details of the workings of such systems remain
`commercially sensitive. The database that we have studied, a videotape of
`all the entries over D2000-2 for almost exactly seven hours for the deutsche
`mark/dollar, and some sixteen minutes less for five other bilateral exchange
`rates, shown on the D2000-2 screen during European business hours on
`16 June 1993 (from 08:31:50 to 15:30:00 British Standard Time [BST], i.e.,
`GMT + 1 ), remains the copyright of Reuters.3 Anyone wishing to use these
`data should refer to Reuters, not to us. We should like to emphasize that this
`videotape did not include, and we have not been given any access to, any infor-
`mation regarding the identity of any of the parties involved in trading; all the
`trades observed by us remain anonymous. Indeed, it is not possible for any
`observer, even in Reuters itself, to identify which are the individual banks us-
`ing the system.
`Readers should keep in mind the shortcomings of these data. They represent
`a short snapshot of conditions in a rapidly changing market over a year ago.
`Trading undertaken over such electronic trading systems may well be, as dis-
`cussed further below, not representative of the market as a whole; trading activ-
`ity on D2000-2 on 16 June 1993 may have differed in some respects signifi-
`cantly from that in surrounding days and weeks; the volume and characteristics
`of electronic trading (over Reuters) in June 1993 may well be quite different
`from that now since over a year has passed.
`Given these disclaimers, why should anyone bother to read on? Despite
`these shortcomings, there are, however, several reasons why this study pro-
`vides new insights in the literature of high-frequency exchange rate behavior.
`First, until now there have been virtually no continuous time-series data avail-
`able at all on actual trades, prices, and volumes in the foreign exchange mar-
`ket.4 The 60 percent or so of deals done directly by two bank counterparties
`over the telephone remain, naturally, private information. There has been little
`use made of data on foreign exchange transactions intermediated by specialist
`interbank brokers, no doubt partly because of commercial and confidentiality
`sensitivities. The only studies currently known to us making use of such data
`are by Lyons (1995, chap. 5 in this volume). Data of any kind on the character-
`istics and continuous time-series behavior of actual trading transactions on the
`foreign exchange market are, therefore, still rare.5 Second, there have been so
`few data on transactions in the foreign exchange market that almost all the
`
`3. We are most grateful to Reuters in general and to Mr. Etherington in particular for allowing
`us to record the quantitative details reported below.
`4. There is, of course, the survey of foreign exchange business that has now been undertaken
`three times at three-year intervals in April 1986, 1989, and 1992 by central banks under the aegis
`of the Bank for International Settlements (BIS), but this does not provide time-series data. The
`volumes reported are aggregates for the month of April.
`5. We have little doubt that such data will become more plentiful and easily available in the
`future. But for the time being at least they have rarity value. Also, as electronic trading systems
`mature, it should be of historical interest to observe how they looked and operated in the early
`stages of their development.
`
`
`
`110
`
`Charles Goodhart, Takatoshi Ito, and Richard Payne
`
`studies on this market have used data on bilateral currency exchange rates that
`emanate from the indicative bid-ask prices shown on electronic screens by the
`specialist information providers, for example, Reuters, Telerate, Knight Rid-
`der, and Quotron. There has, naturally, been some concern whether the high-
`frequency characteristics of such indicative quotes, for example, the negative
`auto-correlation and the fact that the size of the spread clusters at certain con-
`ventional values, are representative of the characteristics of firm (committed)
`bid-ask quotes at the touch. The touch, a term more commonly used in the
`United Kingdom than in the United States, is defined as the difference between
`the best (highest) bid and the lowest ask on offer, where these are (usually)
`input by different banks. Lyons, for example, expressed such concerns when
`he wrote, "Some of the shortcomings of the indicative quotes include the fol-
`lowing. First, they are not transactable prices. Second, while it is true that the
`indicated spreads usually bracket actual quoted spreads in the interbank mar-
`ket, they are typically two to three times as wide. . .. Third, the indications are
`less likely to bracket true spreads when volatility is highest since there are
`limits to how frequently the indications can change. And finally, my experience
`sitting next to dealers at major banks indicates that they pay no attention at all
`to the current indication; rather, dealers garner most of their high-frequency
`market information from signals transmitted via intercoms connected to inter-
`dealer brokers [see Lyons 1993]. In reality, the main purpose of the indicative
`quotes is to provide non dealer participants with a gauge of where the inter-
`dealer market is trading" (1995, pp. 331-32; see also Flood 1994, esp. n. 6,
`p. 154).
`Do, for example, the frequency and volatility of the indicative quotes pro-
`vide a reasonable proxy for the same characteristics both in the committed bid-
`ask quotes and in the associated transactions in the electronic trading systems?
`We provide an initial answer to such questions in section 4.3, where we seek
`to compare characteristics of the FXFX time series6 with those of the D2000-
`2 data for the overlapping seven hours. As described in more detail in section
`4.3, the D2000-2 series was not time-stamped, and our study of this relation is
`conditional on the assumptions and techniques used to match these two
`series temporally.
`Subject to that condition, and to anticipate some of our main findings in
`section 4.3, the averages of the bid-ask in both series (FXFX and D2000-2) are
`almost identical. A graph of the time path for the deutsche mark/dollar from
`the two sources looks like one line (see figure 4.1). Thus, the time path of the
`indicative quotes can, on this evidence, be taken as a very good and close
`proxy for that in the underlying firm series. Nevertheless, some of the charac-
`teristics of the bid-ask series, for example, the pattern of autocorrelation, are
`somewhat different. Even so, both series indicate a somewhat similar GARCH
`
`6. We obtained the accompanying FXFX data series from Dr. M. Dacorogna of Olsen and Asso-
`ciates in Zurich.
`
`
`
`I ll
`
`A Study of the Reuters D2000-2 Dealing System
`
`Tim (Mcondi)
`
`Fig. 4.1 Average of bid-ask for FXFX and 1)2000-2 data: deutsche mark/dollar
`
`pattern. As would be expected, the two series are cointegrated, with the indica-
`tive series responding more to deviations from the equilibrium (i.e., a larger
`and more significant negative coefficient on the error correction mechanism).
`By contrast, the characteristics of the spreads in the FXFX as compared with
`the touch in D2000-2 are markedly different. The spreads in the FXFX series
`show clustering among a small number of standard values (e.g., 5, 7, and 10
`pips for the deutsche mark/dollar), whereas the spreads at the touch show no
`such signs of clustering.
`After examining the relations between the quote series and associated
`spreads of FXFX and D2000-2 in section 4.3, we turn in section 4.4 to a more
`detailed study of the characteristics of D2000-2, in particular, the interaction
`between quotes and transactions in that data set. This long section has five
`subsections. First, in section 4.4.1, we examine the statistical characteristics of
`the transaction price series in D2000-2. Whereas for both D2000-2 and FXFX
`the quote series incorporate a first-order negative moving average, the transac-
`tion price data appear to follow a random walk. Our most interesting finding
`is that the series of runs of deals, sequences of trades at the bid and the ask, is
`not normally distributed but contains some very long consecutive sequences,
`another fat-tailed distribution.
`Second, in section 4.4.2, we examine the interrelations between the available
`data series, using nine main series from D2000-2, all of which, apart from the
`spread, can be separately obtained for the bid and the ask. These are the fre-
`quency of transactions (deals), their size, and whether such transactions ex-
`hausted the quantity currently quoted; then the frequency of quote revision,
`the change in the quoted prices, and the quantity quoted; and two measures of
`volatility, the absolute change in the quote and the standard deviation of the
`quotes. Our main finding is that there is a two-way interrelation between the
`
`
`
`112
`
`Charles Goodhart, Takatoshi Ito, and Richard Payne
`
`frequency of quote revisions and the, frequency of deals and that, when a deal
`exhausts the quantity on offer, this then affects (with one-way causality) a
`nexus of relations between volatility, spreads, and quote revisions. We also
`conduct similar companion studies on the (temporally associated) FXFX data
`using a smaller subset of data series (since we have no data on transaction
`characteristics or on posted quantities from FXFX), but these have less inter-
`esting results.
`Our finding that there is a strong two-way relation between the frequency of
`quote revisions and that of transactions within a period is, we believe, new,
`although the underlying cause, that both derive from the arrival of "news," is
`theoretically straightforward. Most studies of transactions in other asset mar-
`kets (e.g., the New York Stock Exchange [NYSE]), have used data series cali-
`brated in transaction (tick) time, with the result that one cannot then infer
`calendar-time frequency. Otherwise, with relatively low-frequency transac-
`tions on the NYSE, so many of the observations would exhibit zero change.
`With much higher-frequency transactions on foreign exchange markets, it
`seemed to us worthwhile to explore the form of these relations in both clock
`time and transaction time, although we feel that much remains to be done in
`clarifying the appropriate econometric usage in this field.
`Next, in section 4.4.3, we examine the ARCH (autoregressive conditional
`heteroskedasticity) characteristics of the quote series, in particular to discover
`whether their GARCH characteristics would be affected by the addition of
`transactions data. In this case, unlike most of the other main results in section
`4.4.2, the results did appear sensitive to whether the exercise was run in clock
`time or tick time.
`Largely because much more data have been made available for the equity
`market, especially the NYSE, and its associated derivative markets, there has
`been much more empirical work on those markets than for the foreign ex-
`change market. Moreover, the two markets are quite dissimilar in format and
`microstructure, as nicely described in Bessembinder (1994). Nevertheless, de-
`spite the comparatively very small size of our data set, its coverage of transac-
`tions as well as quotes brings it somewhat nearer to the richer data sets avail-
`able on equity markets. In particular, our study here, examining the interaction
`between trades at the bid and ask and price quote revisions, has some features
`in common with that of Hasbrouck's (1991) study of such effects in the NYSE.
`So we then replicate his study as closely as we can, using our own data set and
`adding some variations of our own.
`We draw the conclusions of these exercises undertaken earlier in section 4.4
`together in the final part, section 4.4.5. Throughout this work, the caveat that
`our data set lasts for only seven hours, a possibly atypical period, must always
`be kept in mind, despite the comparatively large number of data points. It is in
`this sense a very small sample. All our findings, both positive and negative,
`must be treated with caution.
`
`
`
`113
`
`A Study of the Reuters D2000-2 Dealing System
`
`4.2 The Characteristics of D2000-2
`
`Automated brokerage terminals do the same job as humans but at a reduced
`cost. A bank dealer who is a member of one of these electronic systems can
`enter her buy and/or sell price into them. Reuters D2000-2 and EBS show only
`the touch, the highest bid, and the lowest ask; these will normally, but not
`necessarily, be entered by different banks. This is different from the indicative
`foreign exchange pages (e.g., FXFX), which show the latest update of the bid
`and ask entered by a single identified bank. On all the electronic trading sys-
`tems, the identity of the inputting bank is not shown. The quantity that the
`inputting bank is prepared to trade is also shown on D2000-2. This was then
`shown as integers of $1 million, and in some bilateral cases DM 1 million,
`from 1-5 and entered as M (medium) for a sum between $6 and $10 million
`and L (large) for sums above $10 million.7 More than one bank may input the
`same best bid (ask) price, in which case the quantity shown is the sum of that
`offered by these banks. The limit orders, that is, those below the (best) bid and
`above the (best) ask, and their associated firm quantities are entered and stored
`in these systems but are not revealed over D2000-2 and EBS. Such reserve
`limit orders are shown on Minex.
`Another bank dealer and member of the trading system can then "hit" either
`the bid or the ask by typing instructions on his own machine. The first check
`is prudential. Banks in such systems may want to restrict the amount of dealing
`with certain other counterparties (in some cases refusing to deal at all with
`some counterparties). The computer first checks whether the deal is pruden-
`tially acceptable to both parties (who remain at this stage anonymous). If not,
`the deal is refused and the "hitter" so informed. We have no information as to
`how often this might happen, but we surmise that it might be fairly rare. As-
`suming that the "hit" is accepted and that several banks are offering the same
`best price, their offers are met on the basis of the time of entry, first in first out.
`When a new deal is made, the new transaction price enters on the right-hand
`column of the screen,8 and there must be an associated change in the quantity
`of the bid (ask), depending on which is hit,9 and also in the price offered if the
`size of the deal exhausts the quantity offered at the previous price. In such
`cases, the bid price must move downwards if there was an exhaustive deal at
`the bid, and the ask price upward following an exhaustive deal at the ask, or
`indicate that there are no remaining limit bids (asks) in the systems, that is, no
`quote shown.10 Note that, in an automatic system like this, a deal must be made
`
`7. This classification has since been changed.
`8. When a new deal has been made, the new transactions price initially for a few seconds shows
`purple, rather than the standard black, on the screen in order to alert traders to this.
`9. When the deal is completed, both banks, the hitter and the quoter, will be sent details regard-
`ing to whom and where to make the payment, which is then settled in the standard fashion. So, ex
`post facto, the identity of the counterparty becomes revealed.
`10. Unhappily, we had a few cases in our data where this directional constraint did not hold.
`While this could be due to new bid-ask inputs occurring at exactly the same moment, several of
`
`
`
`114
`
`Charles Goodhart, Takatoshi Ito, and Richard Payne
`
`at either the posted bid or ask and cannot be made at an interior price between
`them, as can happen with nonautomated human dealers, which can cause prob-
`lems in empirical studies. This has been a particular problem for empirical
`studies of the NYSE (see, e.g., Petersen and Fialkowski 1994; and Lee and
`Ready 1991).
`D2000-2 allowed traders to deal in some fifteen major bilateral exchange
`rates at the time of our exercise. The number and range of currencies covered
`have been changing over time, as is no doubt the case for EBS and Minex as
`well. The screen for D2000-2 is not big enough to show all fifteen at once, and
`in any case such a large number of separate rates might be distracting. So the
`dealer on D2000-2 can call up to six bilateral exchange rate onto the screen at
`any one time.
`All this may be made somewhat easier to follow by seeing an example of
`what a dealer would see when looking at her screen. This is shown in table 4.1.
`Note, in particular, that not all the cells have entries. There are periods, espe-
`cially in the less actively traded bilateral exchange rates, when no bank is mak-
`ing a firm offer. A bilateral currency can have a firm bid (ask) exhibited without
`there being any corresponding ask (bid) on the screen, as in this example for
`the deutsche mark/French franc exchange rate; so there is no observed spread
`at such times. Any bid-ask price must be associated with an accompanying
`quantity offered (and vice versa). As electronic trading becomes more popular,
`such gaps in prices may be expected to become fewer. Note also that the repre-
`sentation of the bilateral exchange rate in the left-hand column is the reverse
`of what would be normally expected, that is, row 1 would in normal usage be
`described as the number of yen per dollar. (We thank a discussant for noticing
`this.) The reason, we understand, for this ordering is that all the volumes are
`denominated in units of 1 million of the first currency shown. Henceforth,
`however, we will revert to the standard representation of the bilateral rates.
`D2000-2 runs throughout the whole day during the week, apart from a short
`break from 2300 GMT to 0100 GMT. On 16 June 1993, a Reuters employee
`started to videotape the bilateral deutsche mark/dollar exchange rate at approx-
`imately 0830 hours BST. This is the dominant and most active of all exchange
`rates (see, e.g., Goodhart and Demos 1990, 1991a, 1991b). About sixteen
`minutes, thirty seconds later, he also put the additional five bilateral exchange
`rates that were shown in table 4.1 up onto the screen.11
`
`these cases probably arise from mistakes in transcribing the videotape (see section 4.2). When we
`had identified these few errors, we removed them from the data set.
`11. Reuters had decided to videotape a day (seven hours) of the working of D2000-2 for their
`own purposes. We do not know why their operator chose these other five bilateral exchange rates.
`There is some autocorrelation in volatility and activity in differing rates from day to day, and
`maybe the operator felt that these would provide either more interest or a better representation
`than the other nine available. But, basically, we do not know, just as we do not know how the
`characteristics of the observations in this seven-hour snapshot compared with the same hours on
`other days, or with other hours on the same day, or with other bilateral rates at the same time.
`
`
`
`115
`
`A Study of the Reuters D2000-2 Dealing System
`
`Table 4.1
`
`D2000-2:
`
`Screen at 10:17:40 on 16 June 1993
`
`Currency
`
`USD/JPY
`DEM/JPY
`USD/CHF
`DEM/CHF
`USD/DEM
`DEM/FRF
`
`Bid
`
`Ask
`
`Quantity
`Columns
`
`Blank
`Columns
`
`106.16
`/
`1.4672
`0.8925
`1.6439
`3.3633
`
`106.25
`/
`1.4679
`0.8933
`1.6443
`/
`
`2X 1
`II
`42
`32
`21
`M/
`
`X X
`X X
`X X
`X X
`X X
`X X
`
`Latest
`Price
`
`106.26
`64.59
`1.4676
`0.8929
`1.6443
`3.3634
`
`Note: USD = U.S. dollar; JPY = Japanese yen; DEM = deutsche mark; CHF = Swiss franc;
`FRF = French franc.
`
`It is this videotape, initially filmed for its own purposes, that Reuters was
`kind enough to let us observe, subject to confidentiality commitments. There
`are four Betacam tapes, which ran virtually continuously, subject to a future
`minor qualification, from 0832 BST to 1530 BST (on 16 June 1993). The
`screen does not show the clock time, and the entries are not time-stamped, but
`a time elapse (time passed since the start of videotaping) was entered onto
`the tape.12
`As might be expected, when the commitments made on screen are firm and
`deals are made at those prices, the original data are, as far as we can judge,
`remarkably accurate. We ended with only a couple of data points that we felt
`must be in error. This compares with errors that occur about once in every four
`hundred entries over FXFX (see Pictet et al. 1994, table 5). By contrast, we
`are conscious that there will be a number of transcribing errors. In particular,
`whether because of the need to copy the tapes or for some other reason, the
`final digit of the five-digit (in one case four-digit) number was often hard to
`decipher. In particular, it was difficult to distinguish zero from eight when these
`were faint on the videotape.13
`In one respect, fortunately, the data are self-checking. When a deal occurs,
`the transaction price in the right-hand column has to be the same as the prior
`(i.e., within seconds earlier) bid, or ask, that was hit and must change the quan-
`tity offered at that prior price, and also the price itself, should the quantity be
`fully taken up. The two series (i.e., of transactions prices, on the one hand, and
`bid-ask prices and their associated quantities, on the other) were transcribed at
`
`12. We were working at Harvard University when we sought to take the details of the tape,
`every entry, from the video onto paper and then back onto electronic diskette. Since no Betacam
`video machines were available in the United States, the tapes were first copied onto S-VHS, and
`the entries on the S-VHS tapes were viewed over a special video player, with adjustable speeds,
`forward and backward, pause, etc.
`13. The transcription from video to paper was primarily done by the wives of two of the authors,
`Mrs. Margaret Goodhart and Mrs. Keiko Ito, also with the assistance of Ms. Yoko Miyao, who did
`this extremely complex and difficult exercise in a dedicated, patient, and conscientious fashion,
`and we are most grateful to them. But there will inevitably be some errors in variables.
`
`
`
`116
`
`Charles Goodhart, Takatoshi Ito, and Richard Payne
`
`separate times. By marrying these up14 and reviewing in cases of errors, we
`can both cross-check the accuracy of our transaction data and get some idea of
`the remaining errors in variables for the entries (bid-ask and associated quanti-
`ties offered) where no such cross-check was possible.15
`Turning now to the data themselves, the database divides into two separate
`parts. First, there is the deutsche mark/dollar market. This is the dominant ex-
`change rate in the foreign exchange market overall, and its dominance of the
`electronic market in our snapshot is even more marked. There were 799 bid
`entries and 823 ask entries (note that these entries would usually come from
`separate banks). Quantities offered at the bid were entered on 802 occasions
`and at the ask on 841 occasions. (Note that the quantity offered can, and does,
`change quite frequently without an associated bid-ask price change. Similarly,
`the price can change without the associated quantity being altered; this hap-
`pened on more occasions than we would have expected, perhaps because a
`bank changed the price for a given amount that it wanted to trade.) Although
`we cannot possibly deduce the total number of independently made entries,
`these might conservatively be put at around fifteen hundred in seven hours, or
`two hundred or so per hour. This compares with some thirty-five hundred en-
`tries over FXFX for the deutsche mark/dollar bilateral exchange rate in the
`same hours, about five hundred per hour. Considering that FXFX represents
`almost costless advertising and is the most commonly used indicative foreign
`exchange price screen, this shows just how busy the deutsche mark/dollar mar-
`ket on D2000-2 was during this snapshot.
`The number of deals in the deutsche mark/dollar was also quite large, rela-
`tive to the commercial target, reported in section 4.1, of one thousand per day
`for deals in all fifteen exchange rates. During this snapshot, there were 186
`deals done at the bid and 251 at the ask. Whether this ratio of deals to bid-ask
`entries is high, low, or normal, we cannot tell. We examine whether this ratio
`varied significantly from half hour to half hour over our data period in sec-
`tion 4.3.
`The depth of the deutsche mark/dollar on D2000-2 was fairly good, although
`it can, and no doubt will, improve further. Following a deal that exhausted the
`
`14. There were a couple of cases when we could not marry the two data points, despite several
`reviews. It is this to which we referred earlier as the only examples of probable errors in the
`original data.
`15. Thus, the cross-check revealed that the accuracy of visually timing the exact moment of an
`entry on a screen was to within about plus or minus three seconds. From the adjustments and
`reviews that had to be made to marry the transaction price data with the bid-ask (and associated
`quantity) data, it may well be that the final digit in the remaining data is incorrect about once every
`thirty observations and the penultimate digit incorrect once every one hundred observations. Some
`of our statistical anomalies, e.g., the few zero and negative spreads and the incorrect direction of
`price movement following a deal, need to be seen in that context. Such inevitable human error
`could have been eliminated had the data been available in electronic disk form, but that was not
`on offer. Moreover, there are some advantages in getting to know the raw data thoroughly before
`proceeding to econometric testing.
`
`
`
`117
`
`A Study of the Reuters D2000-2 Dealing System
`
`quantity offered or the removal of a bid-ask price, most of the time there was
`another limit order on the computer at a closely related price. Histograms of
`quantities offered at the bid measured over both frequency and duration of
`entry are shown in figures 4.2 and 4.3. The histograms for the ask are nearly
`identical and have been omitted to save space. From these it can be seen that
`the frequency and length of time during which no bid or ask price is on th