Journal of International Economics 61 (2003) 307329 www.elsevier.

com / locate / econbase

Informed trade in spot foreign exchange markets: an empirical investigation

Richard Payne*
Department of Accounting and Finance and Financial Markets Group, London School of Economics, Houghton Street, London WC2 A 2 AE, UK Received 31 August 2002; received in revised form 18 September 2002; accepted 5 December 2002

Abstract This paper presents new evidence on information asymmetries in inter-dealer FX markets. We employ a new USD/ DEM data set covering the activities of multiple dealers over one trading week. We utilise and extend the VAR structure introduced in Hasbrouck [J. Finance 46(1) (1991) 179] to quantify the permanent effects of trades on quotes and show that asymmetric information accounts for around 60% of average bid-ask spreads. Further, 40% of all permanent price variation is shown to be due to transaction-related information. Finally, we uncover strong time-of-day effects in the information carried by trades that are related to the supply of liquidity to D2000-2; at times when liquidity supply is high, individual trades have small permanent effects on quotes but the proportion of permanent quote variation explained by overall trading activity is relatively high. In periods of low liquidity supply the converse is trueindividual trades have large permanent price effects but aggregate trading activity contributes little to permanent quote evolution. 2003 Elsevier B.V. All rights reserved.
Keywords: Exchange rates; Market microstructure; Asymmetric information JEL classication: C22; F31; G15

1. Introduction Prior to the 1990s, analysis of the causes of exchange rate movements was a
*Tel.: 144-20-7955-7893; fax: 144-20-7242-1006. E-mail address: r.payne@lse.ac.uk (R. Payne). 0022-1996 / 03 / $ see front matter 2003 Elsevier B.V. All rights reserved. doi:10.1016 / S0022-1996(03)00003-5

308

R. Payne / Journal of International Economics 61 (2003) 307329

eld that was rmly in the hands of macroeconomists. Exchange rate models, based on the goods and asset market approaches were set out and tested using low frequency (e.g. monthly or quarterly) data on exchange rates and macroeconomic fundamentals. However, these tests most often revealed that the fundamentals were less important for exchange rate determination than models predicted. The explanatory power of macroeconomic data for exchange rates was poor and the forecasting power of regressions based on fundamentals was less good than that of a simple random walk. A classic reference along these lines is Meese and Rogoff (1983). This failure has led, in the last decade or so, to increasing attention being paid to models of FX market activity and exchange rate determination based on market microstructure analysis. This large and growing literature places the process by which currencies are actually exchanged in centre stage and focusses on the impact of heterogeneities in the trading population for prices and traded quantities. A key source of heterogeneity in standard microstructure nance is informationalsome agents are assumed to be better informed about future asset prices than others (Glosten and Milgrom, 1985). On an empirical level, such informational asymmetries have several important implications. First, faced with the possibility of trading with a better-informed individual, uninformed liquidity suppliers widen the bid-ask spreads that they charge. This allows them to recoup the losses inicted upon them by insiders from uninformed individuals. Second, and more importantly in the current context, transaction activity carries information and thus trades permanently alter prices. Episodes in which aggressive traders tend to be buying a given currency will lead to its price rising while the converse is true during episodes of aggressive sales. This second prediction is vitalit opens a channel through which transaction activity in FX markets might play a role in exchange rate determination, a feature that is entirely absent from standard macroeconomic exchange rate models. The current study seeks to assess the importance of this channel. Recent empirical papers that also focus on the explanatory power of currency trading activity for exchange rate changes include Lyons (1995) and Yao (1998). Both of these studies use data from single FX dealers to demonstrate that spreads contain an asymmetric information component. Lyons (1996) extends his prior work by examining the role of time in the relationship between trades and quotes. He nds that trades occurring in periods when the market is active convey less information than those consummated when the market is quiet. This is interpreted as consistent with his hot potato hypothesis by which high interdealer volumes are generated more by inventory rebalancing than exploitation of private information. Most recently, studies by Evans and Lyons (2001) and Evans (2001) provide strong evidence for an information content to inter-dealer FX order ow using 4 months of data on direct (i.e. non-brokered) FX trading activity. Theoretical models that focus on the information contained in inter-dealer spot FX trading activity can be found in Lyons (1995) and Perraudin and Vitale (1996).

R. Payne / Journal of International Economics 61 (2003) 307329

309

In these models dealers receive private signals of future exchange rate evolution from their customer (i.e. non-dealer) order ow.1,2 On an institutional level, this is possible as customerdealer trade is entirely opaque (i.e. dealer B cannot observe the customer order ow arriving at dealer A and vice versa). The customer order ow arriving at dealers may be informative for a number of reasons. First, and most blatantly, a given dealer might have an intervening central bank as a customer and thus may learn about future interest rates from the central banks order. However, such occurrences are likely to be rare and hence this channel is not ideal for arguing that customer orders contain information in general.3 A second rationale for the information content of customer order ow comes from arguing that information regarding future exchange rate fundamentals, for example trade balances, is dispersed among individual customers. A given dealer observes the trading behaviour of a group of individual customers and, thus, from their aggregate trading activity receives a signal regarding future fundamentals which can be exploited in inter-dealer trade. Note that in both of the prior examples, customer order ow is informative about exchange rate fundamentalsi.e. interest rates or trade balances. Another class of models, see Evans and Lyons (2001), generates customer trading activity that forecasts future prices because it is informative about risk premia. These portfolio shifts or portfolio balance models generate permanent price shifts due to risk-aversion. Assume that the non-dealer segment of the market experiences a portfolio shift that requires currency trade. Further assume that there is no expectation that this trade will be reversed such that the aggregate inventory imbalance (and risk) foisted upon the market is permanent and undiversiable. Assuming risk averse agents operating in the FX markets then delivers a permanent change in prices due to a permanent change in risk premia. Dealers are informed in this setting as they see a signal of the aggregate portfolio shift through

Throughout this paper we will repeatedly refer to trades, volumes and order ow. Volume has its usual denition as the sum of all Dollar quantities traded in a given interval. However, each individual trade may be given a direction according to whether the aggressor (the agent demanding liquidity) is a buyer or a seller. Order ow is the difference between buyer-initiated and seller-initiated trading activity. It can be measured in Dollar terms (i.e. as the volume of buyer-initiated activity less the volume of seller-initiated) or simply in terms of the number of trades (i.e. as the number of buyer- less seller-initiated trades). 2 Another interesting recent paper by Chakrabarti (2000) explores the possibility that dealers also learn from the quotes of other dealers. Whilst this is interesting, it is not possible to empirically evaluate this proposition with our data and thus we leave it to one side. 3 Peiers (1997) examines USD/ DEM dealer quoting around Bundesbank intervention events, nding that Deutsche Bank is a price leader at these times. Given that Bundesbank intervention operations are carried out via Deutsche Bank, these results lend credence to the notion that observing central bank trades conveys price-relevant information. Corroborating evidence is provided by Naranjo and Nimalendran (2000), who demonstrate that the adverse selection component of FX spreads increases around central bank intervention events.

310

R. Payne / Journal of International Economics 61 (2003) 307329

their customer order ow, before the size of the shift becomes public. Again, dealers can exploit their signals in trade in the inter-dealer market. As stated above, for the current paper the key empirical implication of the preceding models is that trading activity in the inter-dealer market will have a permanent affect on interdealer quotes. This permanent effect is generated by the information carried by inter-dealer trades (regardless of whether this information is about future fundamentals or future risk premia). We test for this permanent price effect using the bivariate VAR model for quote revisions and signed trades introduced in Hasbrouck (1991a). Further, whilst earlier studies have demonstrated that at least some FX trades carry information, none has computed the aggregate impact of such information. Use of the variance decomposition presented in Hasbrouck (1991b) allows us to calculate the proportion of all information entering the quotation process via order ow and hence address this issue. Finally, we examine variations in the information content of trades with the pace of the market. Theoretical contributions such as Admati and Peiderer (1988) and Foster and Viswanathan (1993) predict correlations in the intra-day variation of transactions costs, volume and the intensity of informed trade which we evaluate via a time-of-day subsample analysis of the D2000-2 data. We rene the time-of-day analysis by modifying the basic VAR structure to allow for dependence of the parameters on D2000-2 liquidity measures. One of the main innovations of this work is the use of a new data set on inter-dealer USD/ DEM trades, drawn from an electronic brokerage called D20002. Whilst earlier work in this area has employed data based on the operations of single dealers, the D2000-2 data reect the interactions of multiple traders. As such, these data provide broader coverage of inter-dealer activity. Further, the D2000-2 data can be used to construct proxies for the liquidity of the FX market as a whole. D2000-2 operates as a closed, electronic order book and every limit and market order entered onto the system are available from the data. Liquidity and depth measures can be constructed from the limit order data and used as conditioning variables in the analysis of the effects of private information. Our main results are as follows. First, our estimates imply that over 60% of the D2000-2 spread can be thought of as compensation for informed trade. An unexpected market buy, for example, leads to an upwards equilibrium quote revision of 1 pip (i.e. DM 0.0001) on average. The 60% gure derives from comparing this 1 pip permanent price impact to one half of the average bid-ask spread on D2000-2. Further, we estimate that around 40% of all information entering the quotation process does so through order ow, a gure which is comparable in magnitude to equivalent measures from equity market studies. The 40% number quoted above is important in that it quanties the relative impact of trade-related versus non-trade-related information on exchange rate evolution. However, given that the data we study here cover only one of several venues for inter-dealer FX trade we should clarify its interpretation. Our result does not imply that 40% of the information relevant to long-run USD/ DEM

R. Payne / Journal of International Economics 61 (2003) 307329

311

determination is revealed on D2000-2. The number is derived from the correlation between D2000-2 order ow and permanent USD/ DEM movements. If we were to assume, quite reasonably, that all inter-dealer FX venues shared the same permanent price component, and further that all order ows were perfectly correlated, then we would get the same value for the size of the trade correlated component from each venue. Thus in this case 40% would be a market-wide gure. We would suggest that, in the current context, the size of the trade correlated component should be interpreted along similar lines. However, there is the possibility that informed FX trade is more prevalent on other (less transparent) trading venues. In this case, our 40% number might be interpreted as a lower bound on the impact of private information in FX markets. Finally, impact of informed trade on D2000-2 quotes is shown to depend strongly on the level of market activity. When the D2000-2 order book is relatively thin and volume low (i.e. from the late GMT afternoon to the early GMT morning,) an unexpected trade has a much larger permanent effect on quotes than in peak trading periods. This result is consistent with the empirical analysis of Lyons (1996) and the theoretical predictions of Admati and Peiderer (1988). A more direct examination of the relationship between informed trade and market conditions is obtained from some extended VAR results. These demonstrate that the permanent impact of informed trade on quotes and the supply of limit orders to D2000-2 are negatively and non-linearly related. However, it should be noted that while we show that individual trades have larger permanent price effects in periods of lower trading activity or liquidity supply, the variance decompositions tell us that the aggregate contribution of transaction activity to permanent exchange rate determination is low at these times. Thus these periods should not be considered those in which price discovery is particularly intensive. The rest of the paper is set out as follows. Section 2 introduces the basic features of the D2000-2 data set and Section 3 details the empirical methodology employed in the current study. Section 4 presents the empirical ndings from the VAR estimations. Finally, Section 5 concludes and presents ideas for further work.

2. The dealing system and the data Around 80% of all spot FX trade is inter-dealer, with the remainder consisting of trade between dealers and non-nancial customers or dealers and non-dealer nancial customers. Until fairly recently, all inter-dealer trade, both direct and brokered, was carried out over the telephone. Dealers would call one another to request price information and consummate trades or, alternatively, might call human brokers, also known as voice brokers, to express their trading interests. One implication of this was that, aside from the triennial BIS surveys of FX market activity, no consolidated source of FX trade information existed. Further, the order ow information available to dealers themselves was limited. Indicative quote

312

R. Payne / Journal of International Economics 61 (2003) 307329

information was available via a number of screen based systems and reports of voice brokered trades were broadcast over intercom systems, but there were no indications of market wide order ow. In the early 1990s, however, a shift away from telephone based trade occurred with the introduction of several electronic trading systems. Reuters opened a network called D2000-1, facilitating direct, bilateral inter-dealer trade. Evans (2001) and Evans and Lyons (2001) analyse transactions data from this network. A further two electronic broking systems came into beingthat run by the EBS consortium and the Reuters D2000-2 system. These systems have grown quickly, driving the voice brokered portion of trade down considerably. In June 1997 EBS claimed to handle 37% of all brokered trade in London, with the D2000-2 market share commonly assumed to be similar.4 Based on this gure and on BIS survey information, one might estimate that the share of spot inter-dealer USD/ DEM volume passing through D2000-2 in 1997 was around 10%. The data used in this study cover all USD/ DEM trade on D2000-2 over the week 6th10th October 1997. Around 30,000 transactions occurred during this time with total volume approaching $60bn.

2.1. The D2000 -2 dealing system

D2000-2 is an electronic order driven system. Liquidity is supplied to the system via limit buy and sell orders and is drained from the system in two ways; rst, through market buy and sell orders and direct crossing of limit orders and, second, through voluntary cancellation of limit orders.5 Transaction consummation is governed by rules of price and time priority subject to one proviso. Participants in the system must bilaterally negotiate credit lines if they wish to trade. Thus trades sometimes execute outside the inside spread due to lack of credit lines between limit and market order submitters. A subscriber to D2000-2 sees the following items on the trading display. First there is an indication of the most competitive limit buy and sell prices in the system along with the quantities available at those prices. No information on subsidiary limit orders (i.e. buy / sell orders with prices below / above the extant best price) is displayed. The screen also indicates the last transactions which occurred in a given currency pair, detailing price and volume. The above information is simultaneously available for multiple exchange rates. From our

4 At the current time, that is mid-2002, the market shares of D2000-2 and EBS are certainly not similar. Reuters has the greater market share in Sterling-related exchange rates but EBS appears to have achieved market dominance in most others, including the key EuroDollar rate. 5 By direct crossing of limit orders we mean the situation in which the D2000-2 order book contains a limit buy with price greater than or equal to that of a limit sell. In this case the system automatically transacts the overlapping quantity. Thus, we can treat the entry of limit orders that result in crosses as similar to market order entries.

R. Payne / Journal of International Economics 61 (2003) 307329

313

USD/ DEM data set, one can reconstruct all of the information available to subscribers. Additionally, the data set contains information on every limit order entered on D2000-2.

2.2. The data set

The raw D2000-2 data feed consists of just over 130,000 data lines. Each line contains 10 elds detailing the type of event to which it refers, timestamps with a one hundredth of a second accuracy, price and quantities. Just over 100,000 of these data lines are limit order entries (with timestamps for entry and exit times, a buy / sell indicator, quantity available, quantity traded and price.) The rest of the data consists of market order entries. The market order lines give the quantity transacted and price, a timestamp and whether buyer or seller initiated. From the raw data we construct an event time data set including the following variables; the midquote (i.e. the average of best limit buy and limit sell prices); a signed transaction indicator variable; signed transaction quantity (in $m); the inside spread; aggregate buy and sell order book size and the number of buy and sell limit orders outstanding.6 The rst panel of Table 1 presents descriptive statistics for the market activity variables for seven non-overlapping subsamples of all trading days. The rst six of these subsamples consist of observations from 2 hour segments of each day, covering the period from 6 to 18 GMT. The seventh subsample represents data from GMT overnight periods, 18 to 6 GMT. The main feature of these statistics is that all series have strong repetitive intra-day patterns. The number of limit orders outstanding and aggregate size on the book broadly follow an inverted U-shape across the GMT trading day. Transaction frequency and volume data show a similar pattern, aside from a lull in activity in the period from 10 to 12 GMT. D2000-2 liquidity, as measured by the percentage spread, follows the inverse pattern to transaction activity. It is particularly noticeable that D2000-2 is extremely illiquid between the hours of 18 and 6 GMT. Spreads are very high, order book size limited and in our data this interval accounts for less than 5% of transaction activity. During the complementary portion of the trading day the spread is very small (with a modal value of one pip) and the D2000-2 order book is very deep. Panel (b) of Table 1 contains summary statistics for the percentage change in the midquote for our seven subsamples. Again, the effects of the intra-day in

6 We dene an event as a revision in the best limit buy or sell price or the occurrence of a transaction. The transaction indicator is signed positive when the aggressor (i.e. the market order trader) is buying and negative when the aggressor is selling. The transaction indicator is zero when there is a revision in either of the best limit prices without an accompanying trade. Transaction indicators were constructed both including and excluding crosses. When included, crosses are signed by treating the latest entering limit order as the aggressor. Limit buy (sell) order book size is dened as the aggregate quantity, in $m, outstanding across all buy (sell) limit orders.

314

Table 1 Summary statistics for D2000-2 order book and return data Panel (a): Order book statistics Sample Obs. 6 to 8 8 to 10 10 to 12 12 to 14 14 to 16 16 to 18 18 to 6 GMT GMT GMT GMT GMT GMT GMT 10816 11587 10446 17339 12088 3333 4938 R. Payne / Journal of International Economics 61 (2003) 307329 Bids 31.14 46.63 47.74 50.22 41.65 20.04 9.33 Offers 31.25 47.02 42.15 46.77 33.31 14.46 7.43 Qb 55.24 91.39 99.84 95.31 86.59 54.89 25.45 Qo 58.68 89.82 79.29 80.79 57.37 22.04 15.06 s 0.015 0.011 0.017 0.014 0.018 0.042 0.076 Deal 4921 5751 4654 7519 4101 671 766 Vol. 8742 11036 8600 13888 7302 1080 1155 Size 1.78 1.92 1.85 1.85 1.78 1.61 1.51

Panel (b): Return statistics Sample Mean 6 to 8 8 to 10 10 to 12 12 to 14 14 to 16 16 to 18 18 to 6 GMT GMT GMT GMT GMT GMT GMT 0.00002 20.00000 20.00009 0.00001 0.00004 20.00013 0.00001

Var. 0.00006 0.00003 0.00013 0.00005 0.00007 0.00036 0.00230

Skew 0.40 0.10 20.38 0.04 20.05 0.11 0.05

Kurtosis 92.5 34.2 127.2 55.9 180.9 18.9 34.5

r1
20.38 20.35 20.39 20.34 20.35 20.42 20.50

Q(5) 1619.39 1459.78 1632.87 2086.08 1492.81 601.19 1267.84

Q 2 (5) 1224.92 1254.48 1131.33 1910.44 1066.20 1646.27 1814.57

Note: panel (a) of the table gives basic statistics for order book data. Columns headed Bids and Offers give the average number of orders outstanding across a given sample. The next two columns give average aggregate quantity outstanding on both sides of the order book in $m. s is the average percentage spread in a given subsample. The nal three columns give the total number of transactions, total volume traded in $m and mean transaction size in $m in each subsample. Panel (b) gives summary statistics for midquote returns for the time-of-day-based subsamples. The rst four columns of the table give the rst four sample moments of the return series. The fth column gives the rst order return autocorrelation. Column six presents fth order BoxLjung statistics for returns and the nal column gives fth order BoxLjung statistics for squared residual returns, where residual returns are created by ltering an MA(1) from actual returns. For the tests reported in the nal two columns, the asymptotic 5% and 1% critical values are 11.07 and 15.09 respectively.

R. Payne / Journal of International Economics 61 (2003) 307329

315

D2000-2 activity are apparent. Both return dependence and volatility are inversely related to traded volume. An information ow model of trading would suggest the opposite relationship. Our explanation for the observed correlation is that it is due to D2000-2 liquidity variation: while in peak trading periods more information is impounded into midquotes and midquotes change more frequently, in less active periods the thinness of the D2000-2 order book causes measured return volatility to be high around trades or limit order cancellations. The gures in Table 1 imply that D2000-2 only operates effectively during European and North American trading hours. Hence, in the empirical analysis presented in Section 4, observations from the 18 to 6 GMT period are omitted.

3. Empirical methodology As discussed in the previous section, D2000-2 is a multi-lateral order driven system. This implies that the empirical models used in Lyons (1995) and Yao (1998) are inapplicable here as they are based on an underlying quote-driven, single-dealer structure. Instead, we employ the reduced form VAR in trades and quote revisions developed in Hasbrouck (1991a) and Hasbrouck (1991b). This framework is not predicated on any particular underlying microstructure model and has been used in the analysis of entirely order driven markets in de Jong et al. (1995) and Hamao and Hasbrouck (1995). Information-based trade is identied via a positive and signicant long-run response of quotations to transaction activity, in line with the theoretical argument presented in Section 1. Further, the variance decomposition presented in Hasbrouck (1991b) allows one to evaluate the amount of information entering the FX quotation process that is trade-related and hence the contribution of information-based trade to price discovery. There are two main assumptions which underlie the application of this framework to the current data set. The rst of these is that informed agents exploit their advantage through the use of market orders rather than limit orders. Non-informed agents, on the other hand, submit either market or limit orders to D2000-2 depending on their desire for execution speed. This implies that private information can only inuence prices through unexpected trading activity. The assumption may be justied by noting that an informed agent submitting a limit order is noisily advertising his beliefs and hence possibly eroding his advantage. Moreover, using limit orders to exploit (short-lived) information advantages is risky due to the possibility of non-execution. Finally, the analysis of informed order placement strategy presented in Harris (1998) suggests that informed agents should prefer to trade via market order in fast paced, liquid markets. The spot FX markets certainly exhibit these features lending credence to our assumption. The fact that everyone trading on D2000-2 is a dealer makes the distinction between informed and uninformed agents less clear than in other examples. In the Introduction we argued that customer order ows were the source of information

316

R. Payne / Journal of International Economics 61 (2003) 307329

asymmetries between dealers. Thus one way to discriminate between informed and non-informed dealers is to appeal to variation in the size and quality and size of dealer customer bases.7 It should also be borne in mind that there is nothing here that requires the identity of informed traders to be the same over time. At certain points in time, dealers with unusually good customer order ows may act as informed agents in the inter-dealer market (and, via the preceding arguments, trade by market order) whilst at other times those same dealers are essentially uninformed (and supply liquidity to D2000-2). The second required assumption is that public information is immediately reected in quotes. If this was not the case, traders observing information releases could form protable trading strategies which would generate correlation between trades and subsequent quote revisions in the absence of private information. However, given the levels of liquidity and competition in FX markets this assumption would seem to be reasonable. A nal point that should be discussed here is the attractiveness of D2000-2 as a venue for informed trade. As D2000-2 competes for order ow with several other trading venues we should ask where informed dealers are most likely to trade. Given the microstructural similarities between D2000-2 and EBS, this discussion should focus on the implications of differences between electronically brokered trading and direct trading (whether electronic or over the telephone). The former offers trading opportunities that are pre-trade anonymous but which are broadcast to the rest of the market. The latter offers trading opportunities that are nonanonymous but which remain private to the two counterparties. In our eyes, then, both avenues for trade have an advantagethe pre-trade anonymity of electronic brokers versus the post-trade opacity of direct trading. Thus, on a theoretical level, the venue of choice for informed trade is not clear cut and we would argue that there is no reason to believe that informed traders would always avoid trading on D2000-2.

3.1. The VAR model

Denote the percentage change in the midquote by r t and let x t represent a vector of transaction characteristics, where t is an eventtime observation counter. The basic VAR formulation used in our empirical work is as follows: rt 5

i51

O a r 1O b x
P P i t 2i i50

i t 2i

1 1t

(1)

Citibank, as perhaps the biggest player in customerdealer FX trade, might be thought of as an informed participant in the inter-dealer market whilst a smaller bank with a much smaller customer base might be considered uninformed. Along the same lines, an interesting recent paper that links effects of a dealers trades on the market to that dealers reputation is Massa and Simonov (2001).

R. Payne / Journal of International Economics 61 (2003) 307329

317

xt 5

i51

O g r 1O d x
P P i t 2i i 51

i t 2i

1 2t

(2)

where the errors are to be zero mean and mutually and serially uncorrelated at all leads and lags. Further, we dene Var(1t ) 5 E( 2 ) 5 s 2 and Var(2t ) 5 E( 2 ) 5 1t 2t V. Eqs. (1) and (2) form a general model of the dynamics of trades and quotes and the interactions between these variables. Note that the VAR is not entirely standard as the contemporaneous realisation of x t enters the return equation. Hence trades logically precede quote revisions. This ensures that the innovations to the two equations are uncorrelated and identies the VAR. Given the two assumptions detailed above, the innovations in the VAR can be interpreted as follows. The innovation to the return equation reects transitory quote variation and the effects of public information on quotes. The innovation to the trade equation represents unpredictable transaction activity and hence the possibility of information-based trade. Using this classication, the effects of private information on quotations are easily retrieved. First, we invert the VAR to retrieve the VMA representation:

SD S

rt a(L) 5 xt c(L)

b(L) d(L)

DS D
1t 2t

(3)

where, for example, a(L) 5 a 0 1 a 1 L 1 a 2 L 2 ? ? ? 1 a k L k . Given the lack of correlation between the innovations, the coefcients in the VMA lag polynomials are precisely the impulse responses implied by the VAR. The coefcient a k , for example, is the effect of a unit return shock on the midquote return at a k period horizon. The effects of private information are revealed in the b(L) polynomial. The possibility of informed trade implies that quotes respond permanently to trade innovations and hence b(1) (i.e. o `50 b i ) should be positive and signicant. i Estimation of the VAR and calculation of the implied impulse response functions will hence allow us to evaluate the existence of information-related trade on D2000-2. In the empirical analysis reported in Section 4, the order of the VAR was chosen via application of the Schwarz information criterion. The VAR equations were estimated by OLS and are reported with heteroskedasticity robust standard errors. The VMA representation was calculated by simulation.

3.2. The variance decomposition

While the VAR model allows us to quantify the information content of a single trade, it does not permit one to assess the overall importance of informed trade in determining the evolution of the exchange rate. Hasbrouck (1991b) presents a variance decomposition for returns, based on the VAR structure above, which permits retrieval of the variance of the permanent component of midquotes, plus

318

R. Payne / Journal of International Economics 61 (2003) 307329

the proportion of permanent variation related to order ow. The overall inuence of informed trade can be estimated with the latter measure. Denoting the (log) of the midquote as qt , the following structure is assumed: qt 5 m t 1 s t (4)

Eq. (4) decomposes the midquote into a random walk component (m t ) and a mean-zero stationary term (s t ). Hence,
2 m t 1 m t21 1 vt , vt | N(0, s w ), E(vt vs ) 5 0 for t s

and lim k ` Et s t1k 5 lim k ` Es t 1k 5 0. From an economic perspective, m t can be thought of as the fundamental or full-information price process. The transitory component, s t , represents that portion of the current midquote generated by any non-information based microstructure effect, e.g. price discreteness, digestion effects or inventory control. The key parameter in the preceding formulation is s 2 which measures variation w in the permanent component. This can be estimated by equating the return representation implied by Eq. (4) (using the fact that r t 5 Dqt ) and the return equation from the VMA. Furthermore, variation in the permanent component due to order ow alone can also be retrieved. These measures are calculated as:

s 5

2 w

swx 5

SO D SO D S O D S O D S O 9D
` ` `

bi V

b i9 1 1 1 bi

ai

s2

(5) (6)

i 50 `

i 50 `

i 51

bi V

i 50

i 50

Standard errors for these two variance estimates can be computed via a residual based bootstrap of the estimated VAR system. An economic interpretation of Eqs. (5) and (6) is as follows. Public information events are incorporated into the exchange rate via the return innovation, 1t . The permanent effect on midquotes of a unit return shock is given by unity (the contemporaneous impact) plus o `51 a i i and hence the variation in the permanent component implied by public information events is given by the second term on the right hand side of Eq. (5). Private information is impounded into the exchange rate via trade innovations with the permanent impact of an unexpected unit trade given by o `50 b i in the case where i x t is scalar. For vector x t , the variation in the permanent component driven by trade innovations is thus the rst term on the right hand side of (5).

3.3. Non-linear effects in the tradequote relationship

While the VAR structure presented in Section 3.1 provides a fairly robust characterisation of the determination of trading activity and quote revisions, it restricts the relationship between these two variables to be invariant to underlying

R. Payne / Journal of International Economics 61 (2003) 307329

319

market activity. Theoretical considerations, however, suggest that the pace of the market will impact upon the response of quotes to trades. Admati and Peiderer (1988) suggest that there should be a negative correlation between the information content of trades and overall volume, for example, and the empirical results of Lyons (1996) might be considered consistent with this. We allow for the possibility that the dynamic relationship between quotes and trades varies with market conditions by using an arranged VAR estimation. Specically, consider a weakly exogenous candidate measure of the pace of the market, z t , and denote by Dt the matrix of time-series data on z t , returns, trades and their respective lags i.e. Dt 5 hz t , r t , r t21 , . . . , r t 2P , x t , x t 21 , . . . , x t 2P j. Now, consider a sort of Dt according to an increasing order sort of z t . Denote the sorted data set by Ds . One way to investigate variation in the VAR parameters with market activity would be to estimate Eqs. (1) and (2) using successively larger numbers of sorted observations via recursive least squares. Evidence of parameter change could then be evaluated via plots of the recursive coefcient estimates or through formal tests for parameter instability. We follow a similar route to the above but we run a series of regressions where the number of observations in each regression is xed and this xed window is moved through the sorted data sample. We use a window size of 2000 observations such that our rst regression uses sorted observations labelled s512000 (i.e. the rst 2000 rows of the sorted data set Ds ), our second regression uses rows 1012100 of Ds , the next regression uses rows 2012200 etc. This technique provides similar information to that outlined in the previous paragraph but the series of estimated parameters is not smoothed by the use of an increasing sample size. Plots of the parameter estimates are used to gain insight into the existence of non-linearities in the VAR structure. The rst lag of the number and aggregate size of outstanding limit orders are employed as z t in our analysis.

4. Results This section presents the results from estimating the VAR models in Section 3.1 and Section 3.3 and the variance decomposition in Section 3.2. We begin with results for the entire D2000-2 trading day data set. Results for time-of-day based subsamples are discussed next. Finally we analyse the effects of market activity on the information content of trades via the arranged VAR models. The basic variables included in our VARs were percentage midquote returns and a signed transaction indicator (which takes the values 21, 0 and 11). In the construction of the transaction indicator only market orders were used, limit order crosses were ignored. Inclusion of crosses in the denition of the transaction indicator changes the results only marginally. A signed volume variable was similarly constructed. The rst few columns in the top row of row of Table 2 give a summary of the relevant VAR parameters estimated using all trading day observations. Some

320 R. Payne / Journal of International Economics 61 (2003) 307329

Table 2 Summary of VAR results for trading day and subsamples Subsample VAR results Lag 6 to 18 GMT 6 to 8 GMT 8 to 10 GMT 10 to 12 GMT 12 to 14 GMT 14 to 16 GMT 16 to 18 GMT 8 5 5 5 4 9 5 SIC 89877 28425 216964 2106 27213 24220 2406 o i bi 0.00668 0.00565 0.00406 0.00758 0.00548 0.00750 0.01853 Variance decompositions

x 2 ( bi )
4178.7 837.8 1060.1 561.5 1344.1 804.6 247.1

QIR 0.00518 0.00502 0.00404 0.00583 0.00490 0.00564 0.01192

Q(10) 36.1 42.4 19.9 70.0 55.6 11.2 56.7

swx /s 2 v
0.41 (0.015) 0.43 (0.030) 0.47 (0.027) 0.32 (0.034) 0.36 (0.020) 0.41 (0.041) 0.35 (0.046)

s 2 /s 2 v r
0.27 (0.009) 0.32 (0.021) 0.42 (0.022) 0.25 (0.019) 0.39 (0.018) 0.23 (0.021) 0.17 (0.014)

swx /s 2 r
0.11 0.14 0.20 0.08 0.14 0.10 0.06

Note: the table summarises the VAR results from the entire trading day and for the 6 time-of-day subsamples. The column headed Lag reports the number of lags in the VAR, chosen using the Schwarz information criterion, which is reported in the second column. o i bi is the sum of the asymmetric information coefcients from the VAR. The following column gives a Wald test statistic for the null that all are zero. The asymptotic 5% and 1% critical values for this test are 3.84 and 6.63 respectively. QIR denotes the long-run quote impulse response implied by the VAR and the column headed Q(10) gives 10th order BoxLjung statistics for the VAR residuals. The asymptotic 5% and 1% critical values for this test are 18.31 and 23.21 respectively. swx /s 2 is the proportion of the permanent midquote variance that is v trade-correlated. s 2 /s 2 is the size of the permanent component as a proportion of total return variance. swx /s 2 is the ratio of the trade correlated component to overall v r r return variation. These gures are calculated using the expressions in Eqs. (5) and (6). Numbers in parentheses in the rst two columns are bootstrapped standard errors for the ratios. They are calculated from a standard, residual-based bootstrap of the model using 500 bootstrap replications.

R. Payne / Journal of International Economics 61 (2003) 307329

321

general comments on those parameters not presented are as follows. First, quote returns demonstrate signicant negative autocorrelation at all included lags. The transaction indicator displays strong positive autocorrelation. This indicates runs in buying and selling activity and may be due to dealers splitting large market orders or possibly order ow imbalances induced by informed trade.8 Finally, the effects of lagged quote returns on transaction direction are mixed. Although marginally signicant as a group, these coefcients were generally individually insignicant. The R 2 in the return equation was 0.22 whilst the trade equation R 2 was approximately 0.075. From the current perspective, the key parameters in the VAR formulation are those labelled bi in Eq. (1) i.e. the effects of trades on current and subsequent midquote returns. As shown in row one of Table 2 the sum of these coefcients is positive. Moreover, each individual coefcient is positive and all are statistically signicant. Hence, a market buy tends to increase quotes. Computing the VMA representation and calculating the equilibrium midquote impulse response shows that an unexpected market buy leads to an upward quote revision of around 0.005% on average.9 The average percentage bid-ask spread on D2000-2 is around 0.016%. Expressing the equilibrium price impact of a trade as a fraction of one-half of the bid-ask spread gives an estimate of the contribution of asymmetric information to spreads. In this case, the numbers tell us that around 60% of the spread is compensation for asymmetric information. This 60% gure is very high relative to recent estimates of similar quantities from equities markets. For example, Huang and Stoll (1997) report average asymmetric information components between 10 and 20% for the 20 stocks of the US Major Market Index. At least some of the discrepancy might be explained by fundamental microstructural differences between the NYSE and the electronically brokered segment of the inter-dealer FX market. One would conjecture that both order processing costs and inventory control costs would be larger for an intermediated equity market, such as the NYSE, where trade frequency is low relative to that in FX, for example. The fact that spreads in the Huang and Stoll (1997) data are around 25 basis points on average, while in our inter-dealer FX data the average percentage spread is less than 1 basis point, lends credence to this conjecture. Nonetheless, the fact that the asymmetric information component is so large in our data is noteworthy. It should be noted that we also experimented with the use of trade size variables in the trade description vector (x t ). One such experiment involved the use of signed volume and signed squared volume in addition to trade direction. In a
8 In many other data sets a similar result would be found due to the data recording one transaction in which a market order executes against a number of limit orders as a series of trades, one for each limit order affected. Note that in the D2000-2 data this is not the case: one market order appears as one trade in the data set, regardless of how many limit orders it executes against. 9 This translates to a 1 pip (i.e. DM 0.0001) quote increase approximately.

322

R. Payne / Journal of International Economics 61 (2003) 307329

second set of experiments we examined whether the largest k% of trades (where we varied k from 5 to 25) had return impacts that differed from those of all other trades. In contrast to the results of Yao (1998), the size variables were generally insignicant. This is likely to be due to the fact that there is very little variation in trade size on D2000-2. Over 90% of D2000-2 trades are for less than $5m such that the data may not contain the power needed to identify a size effect. The result presented in the preceding paragraphs is keymarket orders permanently alter D2000-2 quotes and we take this as evidence that they are information-motivated. Of course, given that our study is based on 5 days of data, it might be argued that the effects we uncover are not permanent in that they die out over a horizon longer than 1 week. This is a possibility. However, a number of theoretical and empirical arguments support our viewpoint. First, on a theoretical level, effects of order ow on exchange rates that mean-revert over weeks or months might indicate prot opportunities, violating market efciency. Second, corroborating empirical evidence on the long-run / permanent effect of market order activity on quotes can be drawn from regressions of returns on order ows aggregated over xed calendar time intervals as in Evans and Lyons (2001). We have run a set of such regressions having midquote returns for sampling frequencies ranging from 20 seconds to 1 hour on the left-hand side and aggregate order ow (i.e. the excess of the number / volume of market buys over market sells in an interval) on the right-hand side. We choose 1 hour as the lowest sampling frequency due to the fact that our entire span of data only covers 5 days. Results demonstrate that across all sampling frequencies, the effect of order ow on returns is highly signicant. Moreover, the R 2 of the regression increases as the sampling frequency decreases such that at a 1 second level, order ow explains approximately 70% of all return variation. Thus, we believe that the results presented here, coupled with those in Evans and Lyons (2001), provide clear evidence for the information content of order ow and its longer run effects on the USD/ DEM. The results above conrm the ndings of previous research in that a portion of trade on D2000-2 can be characterised as information-related. A question which has not been directly addressed in the literature, however, is the extent to which such asymmetries alter over the trading day. Microstructure theory relevant to intra-day variations in the intensity of informed trade and liquidity can be found in Admati and Peiderer (1988) and Foster and Viswanathan (1990). The endogenous information acquisition model of Admati and Peiderer (1988), for example, predicts that high volume periods should be characterised by relatively small price impacts from trade. This is due to the clustering of discretionary liquidity trade and increased competition between informed traders in equilibrium. The model of Foster and Viswanathan (1990) has a similar setup except that the information advantage of insiders is assumed to decline over time. We examine these issues by estimating the VAR separately for the six previously dened non-overlapping subsamples of the trading day. The results, are given in

R. Payne / Journal of International Economics 61 (2003) 307329

323

rows 2 to 7 of Table 2, correspond with the predictions made by Admati and Peiderer (1988). First note that Table 1 shows that the segments of the trading day with greatest volume are those with lowest percentage spreads. Admati and Peiderer (1988) construct a batch trading model and, hence, it does not contain a spread. However, given that their framework predicts more liquid markets in high volume intervals this result is consistent with their analysis. Further, Table 2 indicates that the information content of trading activity follows an intra-day pattern which is the inverse of that followed by volume. Thus, in high volume / liquidity periods, the price response to a trade is relatively low. This is consistent

Fig. 1. The response of the midquote to a unit trade innovation. Notes: these plots show the response of the D2000-2 midquote (i.e. the average of the best limit buy and sell prices) to a unit trade innovation. The impulse response functions are calculated via simulation from the VAR estimates summarised in Table 2. Panel (a) presents the impulse response function estimated from all trading day data (i.e. all data falling in the 618 GMT interval. Panel (b) shows the function calculated using data from the 810 GMT subsample only and, similarly, panel (c) shows the impulse responses implied by the VAR estimated using data from the 1618 GMT interval. The solid line in each panel gives the actual impulse response function and the dotted lines are a 95% condence band, estimated using a residual based bootstrap of the VAR model with 500 bootstrap replications. The x-axis values give the number of observations since the trade shock was rst felt.

324

R. Payne / Journal of International Economics 61 (2003) 307329

with the hot-potato hypothesis of Lyons (1996), although it runs counter to the results presented in Dufour and Engle (2000). A further point to note is that, while the size of the asymmetric information effect is inversely related to volume, the signicance of the trade variables is greater in high-volume periods. These results contrast with those of Foster and Viswanathan (1993) who demonstrate a positive relationship between adverse selection costs and trading volume for a sample of NYSE equities and reject the models of Admati and Peiderer (1988) and Foster and Viswanathan (1990). Hence, our results suggest that high volume periods are characterised by a concentration of liquidity trades and increased competition between informed agents, with these effects reducing the price impact of a trade. Fig. 1 graphically illustrates the VAR results, plotting the quote impulse response functions (along with 95% condence intervals) for the entire trading day data set plus the 8 to 10 GMT and 16 to 18 GMT subsamples in panels (a) through (c) respectively. These two subsamples were chosen as they are those with the lowest and highest average spread respectively. The individual panels of Fig. 1 also demonstrate variation in D2000-2 depth across the trading day through variation in the immediate response of the midquote to a trade. In panel (c) this is about ve times as large as the corresponding value from panel (b)and thus peak trading periods are characterised by greater clustering of limit orders around the inside spread. Based on the VAR results, the return variance decompositions are presented in the nal three columns of Table 2. Across the entire trading day, 41% of the permanent return variance is attributable to order ow (with a bootstrapped standard error for this estimate of 0.015). Comparable gures are contained in Hasbrouck (1991b) who reports an average value of 33% for a sample of U.S. equities and de Jong et al. (1995) who analyse a sample of French stocks traded on the Paris Bourse and report an average trade correlated component of 40%. Hence our results imply that the information content of USD/ DEM order ow on D2000-2 is of the same magnitude as that on equities markets. Examination of the last two columns of Table 2 shows that the permanent component accounts for only one quarter of all return variation such that the information contained in order ow contributes one tenth of total return variance.10 Of course, the distinction between the size of trade correlated component, at 40%, and the 10% number from the previous sentence is due to high-frequency, transitory return variation reducing the latter. The variance decompositions for the six trading day subsamples are also presented. While the VAR results showed that the information content of a single trade was inversely related to the level of overall traded volume, the variance
Note that if one performs the permanent / transitory decomposition on data from the 186 GMT period only, the ratio of permanent to total return variation is less than 0.05. Thus, during this interval returns are almost entirely noise rather than signal. It is for this reason that we omit these data from our econometric analysis.
10

R. Payne / Journal of International Economics 61 (2003) 307329

325

decompositions show that the information contained in order ow as a whole is positively related to volume. The ratio of the trade-correlated component to the variance of total permanent quote variation (swx /s 2 ) is greatest during episodes of v high traded volume. Furthermore, the proportion of all return variation explained by order ow information and the size of the permanent component have a positive correlation with volume. Again, these results can be reconciled with the predictions from theoretical models of intraday variation in transaction activity and price formation. Admati and Peiderer (1988) imply that, in high volume intervals, return variance should be high and prices more informative. In Section 2 we discussed the negative correlation between total traded volume and volatility arguing that the lack of order book depth in low trading periods was behind this result. However, the variance decompositions show that both permanent return variation and the information entering prices through order ow increase with volume. Aggregate trading activity hence contributes a greater proportion of all information in peak trading periods in line with Admati and Peiderer (1988). The prior results provide evidence that the information content of a trade and that of aggregate order ow are strongly linked to market activity via time of day. Our nal set of estimations rene this analysis. The relevant theoretical models predict such patterns due to clustering of liquidity trading in equilibrium. Hence, empirical identication of times of concentration in uninformed activity should corroborate the prior results. Given the assumptions on order placement strategy made in Section 3, a concentration of noninformation based trade should be associated with a high level of limit order placement (the informed preferring to trade by market order due to the execution certainty it yields.) Data from such episodes should hence display small price impacts from trades. This is examined using the arranged VAR technique discussed in Section 3.3. Both the rst lag of the number and aggregate size of all limit orders outstanding are used as the variable governing the non-linearities in the VAR (z t ). Fig. 2 presents the moving window OLS estimates of the asymmetric information coefcients from the VAR (i.e. o P bi ) after arranging the data by the i50 aggregate limit order size (panel (a)) and numbers of orders outstanding (panel (b)) respectively. These estimates are from an 8 lag VAR using the 618 GMT D2000-2 sample. The x-axis labels in the panels give the sorted observation number of the nal included data point. Both panels of Fig. 2 strongly suggest that the asymmetric information problem is lower when z t is higher i.e. in periods of high limit order placement. Conversely, when either the number or size of outstanding limits is low, the impact of an unexpected trade on the midquote is large. The relationship is clearly non-linear, however, with the sum of asymmetric information coefcients declining sharply over the rst 15,000 sorted observations and less rapidly thereafter. Nonetheless, the graphs provide evidence consistent with Admati and Peiderer (1988) and the hot-potato hypothesis of Lyons (1996).

326

R. Payne / Journal of International Economics 61 (2003) 307329

Fig. 2. Sum of asymmetric information coefcients from the arranged VAR specications. Notes: these plots show the sum of the asymmetric information coefcients (i.e. o P51 bi ) from arranged VAR i estimations, as discussed in Section 3.3. In panel (a) the data are arranged by aggregate order book size (i.e. the sum across limit buy and sell sides of order quantity outstanding in $m) and in panel (b) the data have been arranged by aggregate number of order outstanding. The values on the x-axis give the number of the last observation included in the subsample upon which the VAR was estimated.

We have tested the stability of the VAR coefcients in a number of ways. For example, we have estimated extended versions of Eqs. (1) and (2) where right-hand side variables are interacted with dummies for low, medium and high levels of liquidity supply. Formal tests for parameter stability conrm the visual

R. Payne / Journal of International Economics 61 (2003) 307329

327

evidence in Fig. 2, yielding strong evidence of structural change in the VAR equations related to the level of liquidity supply. Finally, it is clear that these arranged VAR results tell essentially the same story as do our results from time-of-day subsamples of the data. When D2000-2 is busy, price impacts of trading tend to be smaller. We have included them here as we feel that time-of-day effects, in themselves, are uninterestingthe interesting issue is what economic phenomenon generates them. We feel that our result relating the size of information effects to the rate of liquidity supply is a step towards understanding the underlying economics.

5. Conclusion This paper analyses trades and quotes from the electronically brokered segment of the spot USD/ DEM market. Specically, we examine whether this market can be characterised as subject to information asymmetries. Using the technology introduced by Hasbrouck (1991a) and Hasbrouck (1991b) we nd that trades do carry information. Roughly 60% of the bid-offer spread in our data can be related to the asymmetric information problem. Further, around 40% of the variation in the fundamental price is shown to be order ow driven, a proportion which is comparable to those found in studies of equity markets. We uncover clear intra-day variation in the information content of a trade and the total information content of order ow. In line with the theoretical predictions, in high volume / liquidity intervals the information content of a single trade is low while the share of information entering the midquote through order ow is high. A more rened analysis of the relationship between the price impact of trades and market activity is also presented. Similar to the time-of-day analysis, results from an arranged VAR estimation indicate that the asymmetric information coefcients vary systematically with market liquidity. There are several possible extensions to the current study. First, a more careful analysis of the non-linearities in the VAR structure may prove to be fruitful. Parameterising the nonlinearities would likely yield insights into the determination of price impacts from trade. Another interesting area of research would be to examine how transaction activity affects subsequent liquidity supply through analysis of the excess demand and supply schedules for currency. Furthermore, econometric analysis of individual limit order entry and execution would provide stronger evidence on the order placement strategy of dealers and how it reacts to trading activity and quote volatility. Such issues are currently under investigation. Finally, as we have mentioned above, D2000-2 is one of several venues for inter-dealer foreign exchange trade. It would be interesting to evaluate how these venues interact (in terms of trading activity and quoting behaviour). This would give a clearer picture of when and where informed trade was trading price.

328

R. Payne / Journal of International Economics 61 (2003) 307329

Acknowledgements I would like to acknowledge the nancial support of the Financial Markets Group at LSE and the ESRC (UK). I am grateful to Jon Danielsson, Charles Goodhart, Roger Huang, Rich Lyons, Carol Osler, Paolo Vitale, seminar participants at LSE and the City University Business School, three anonymous referees and the Editor, Andrew Rose, for helpful comments. Thanks also to Reuters plc for providing the data employed in this study. All remaining errors are my own.

References
Admati, A., Peiderer, P., 1988. A theory of intra-day patterns: volume and price variability. Review of Financial Studies 1, 340. Chakrabarti, R., 2000. Just another day in the inter-bank foreign exchange market. Journal of Financial Economics 56, 2964. de Jong, F., Nijman, T., Roell, A., 1995. A comparison of the cost of trading French shares on the Paris Bourse and on SEAQ International. European Economic Review 39, 12771301. Dufour, A., Engle, R., 2000. Time and the price impact of a trade. Journal of Finance 55, 24672498. Evans, M., 2001. FX Trading and Exchange Rate Dynamics. Working Paper, Department of Economics, Georgetown University. Evans, M., Lyons, R., 2001. Order ow and exchange rate dynamics. Journal of Political Economy 110, 170180. Foster, D., Viswanathan, S., 1990. A theory of the interday variations in volume, variance and trading costs in securities markets. Review of Financial Studies 3, 593624. Foster, D., Viswanathan, S., 1993. Variations in trading volume, return volatility and trading costs: evidence on recent price formation models. Journal of Finance 48 (1), 187211. Glosten, L., Milgrom, P., 1985. Bid, ask and transaction prices in a specialist market with heterogeneously informed traders. Journal of Financial Economics 14, 71100. Hamao, Y., Hasbrouck, J., 1995. Securities trading in the absence of dealers: trades and quotes on the Tokyo Stock Exchange. Review of Financial Studies 8 (3), 849878. Harris, L., 1998. Optimal dynamic order submission strategies in some stylized trading problems. Financial Markets, Institutions and Instruments 7 (2). Hasbrouck, J., 1991a. Measuring the information content of stock trades. Journal of Finance 46 (1), 179206. Hasbrouck, J., 1991b. The summary informativeness of stock trades: an econometric analysis. Review of Financial Studies 4 (3), 571595. Huang, R., Stoll, H., 1997. The components of the bid-ask spread: a general approach. Review of Financial Studies 10, 9951034. Lyons, R., 1995. Tests of microstructural hypotheses in the foreign exchange market. Journal of Financial Economics 39, 321351. Lyons, R., 1996. Foreign exchange volume: sound and fury signifying nothing? In: Frankel, J., Galli, G., Giovannini, A. (Eds.), The Microstructure of Foreign Exchange Markets. University of Chicago Press. Massa, M., Simonov, A., 2001. Reputation and inter-dealer trading: a microstructure analysis of the treasury bond market. Working Paper, INSEAD. Meese, R., Rogoff, K., 1983. Empirical exchange rate models of the seventies: do they t out of sample? Journal of International Economics 14, 324.

R. Payne / Journal of International Economics 61 (2003) 307329

329

Naranjo, A., Nimalendran, M., 2000. Government intervention and adverse selection costs in foreign exchange markets. Review of Financial Studies 13, 453477. Peiers, B., 1997. Informed traders, intervention, and price leadership: a deeper view of the microstructure of the foreign exchange market. Journal of Finance 52 (4), 15891614. Perraudin, W., Vitale, P., 1996. Interdealer trade and information ows in the foreign exchange market. In: Frankel, J., Galli, G., Giovannini, A. (Eds.), The Microstructure of Foreign Exchange Markets. University of Chicago Press. Yao, J., 1998. Market making in the interbank foreign exchange market. New York University Salomon Center Working Paper [S-98-3.