BIS Working Papers

No 405 Information Flows in Dark Markets: Dissecting Customer Currency Trades

by Lukas Menkhoff, Lucio Sarno, Maik Schmeling and Andreas Schrimpf

Monetary and Economic Department

March 2013

JEL classification: F31, G12, G15. Keywords: Order Flow, Foreign Exchange Risk Premia, Heterogeneous Information, Carry Trades, Hedge Funds.

BIS Working Papers are written by members of the Monetary and Economic Department of the Bank for International Settlements, and from time to time by other economists, and are published by the Bank. The papers are on subjects of topical interest and are technical in character. The views expressed in them are those of their authors and not necessarily the views of the BIS.

This publication is available on the BIS website (www.bis.org).

Bank for International Settlements 2013. All rights reserved. Brief excerpts may be reproduced or translated provided the source is stated.

ISSN 1020-0959 (print) ISBN 1682-7678 (online)

Information Flows in Dark Markets: Dissecting Customer Currency Trades

Lukas Menkho Lucio Sarno Maik Schmeling Andreas Schrimpf

This version: March 5, 2013

Abstract We study the information in order ows of dierent customer segments in the worlds largest over-the-counter market, the foreign exchange market. The analysis draws on a unique dataset covering a broad cross-section of currency pairs and distinguishing trades by key types of foreign exchange end-users. We nd that order ows are highly informative about future exchange rates and provide signicant economic value for the few large dealers who have access to these ows. Moreover, customer groups systematically engage in risk sharing with each other and dier markedly in their predictive ability, trading styles, and risk exposure.

JEL Classication: F31, G12, G15. Keywords: Order Flow, Foreign Exchange Risk Premia, Heterogeneous Information, Carry Trades, Hedge Funds.
We would like to thank an anonymous Referee, Alessandro Beber, Claudio Borio, Geir Bjnnes, Michael Brandt, Steve Cecchetti, Jacob Gyntelberg, Hendrik Hakenes, Campbell Harvey, Joel Hasbrouck, Terrence Hendershott, Sren Hvidkjr, Gur Huberman, Alex Kostakis, Jeremy Large, Albert Menkveld, Roel Oomen, Richard Payne, Alberto Plazzi, Lasse Pedersen, Tarun Ramadorai, Jesper Rangvid, Paul S oderlind, Christian Upper, Adrien Verdelhan, Michel van der Wel, as well as participants at several conferences, workshops and seminars for helpful comments and suggestions. We are very grateful to Gareth Berry, Georey Kendrick and UBS for providing us with the proprietary data used in this study, and for numerous conversations on the institutional details of foreign exchange trading at UBS. Sarno acknowledges nancial support from the Economic and Social Research Council (No. RES-062-23-2340) and Menkho and Schmeling gratefully acknowledge nancial support by the German Research Foundation (DFG). The views expressed in this paper are those of the authors and do not necessarily reect those of the Bank for International Settlements. Kiel Institute for the World Economy and Department of Economics, Leibniz Universit at Hannover, K onigsworther Platz 1, 30167 Hannover, Germany, Tel: +49 511 7624552, Email: menkho@gif.unihannover.de. Cass Business School and Centre for Economic Policy Research (CEPR). Corresponding author: Faculty of Finance, Cass Business School, City University London, 106 Bunhill Row, London EC1Y 8TZ, UK, Tel: +44 20 7040 8772, Fax: +44 20 7040 8881, Email: lucio.sarno@city.ac.uk. Faculty of Finance, Cass Business School, City University London, 106 Bunhill Row, London EC1Y 8TZ, UK, Email: maik.schmeling.1@city.ac.uk. Bank for International Settlements and CREATES, Centralbahnplatz 2, 4002 Basel, Switzerland. Tel: +41 61 280 8942. Email: andreas.schrimpf@bis.org.

The foreign exchange (FX) market is the largest nancial market in the world with a daily trading volume of about four trillion U.S. dollars (BIS, 2010). Also, the FX market is largely organized as an over-the-counter (OTC) market, meaning that there is no centralized exchange and that market participants can have only partial knowledge about trades of other market participants and available liquidity in dierent market segments. Hence, despite its size and sophistication, the FX market is fairly opaque and decentralized because of its market structure. Adding to this lack of transparency, various trading platforms have been introduced and market concentration has risen dramatically over the last decade with a handful of large dealers nowadays controlling the lions share of FX market turnover (see, e.g., King, Osler, and Rime, 2012). The FX market can thus be characterized as a fairly dark market.1 This paper addresses several related questions that arise in this opaque market setting. First, do large dealers have an informational advantage from seeing a large portion of customer trades, that is, do customer trades carry economic value for the dealer? Answering this question is relevant for regulators and useful for understanding the implications of the observed shift in market concentration. Second, how does risk sharing take place in the FX market? Do customers systematically trade in opposite directions to each other or is their trading positively correlated and unloaded onto dealers (as in, e.g., Lyons, 1997)? Answering these questions is highly relevant to provide a better understanding of the working of the FX market and, more generally, the functioning of OTC markets. Third, what characterizes dierent customer groups FX trading, e.g., do they speculate on trends or are they contrarian investors? In which way are they exposed to or do they hedge against market risk? Answering these questions allows for a better grasp of what ultimately drives the demand for currencies from dierent types of end-users and enhances the knowledge about the ecology of the worlds largest nancial market. We empirically tackle these questions by means of a unique data set covering more than ten years of daily end-user order ow for up to fteen currencies from one of the top FX
Due (2012) provides a general overview of how opaqueness and market structure impact price discovery and trading in dark markets, that is, OTC markets.
1

dealers, UBS. The data are disaggregated into four dierent groups of nancial (asset managers and hedge funds) and non-nancial (corporate and private clients) end-users of foreign exchange. We therefore cover the trading behavior of various segments of end-users that are quite heterogenous in their motives of market participation, informedness and sophistication. Putting these data to work, we nd that: (i) Order ow by end-users is highly informative for future exchange rate changes and carries substantial economic value for the dealer observing these ows; (ii) there is clear evidence that dierent end-user segments actively share risks with each other; and (iii) end-user groups follow very heterogeneous trading styles and strategies and dier in their exposures to risk and hedge factors. This heterogeneity across players is crucial for risk sharing and helps explain the vast dierences in the predictive content of ows across end-user segments that we document in this paper. To gauge the impact of order ow on currency excess returns, we rely on a simple portfolio approach. This multi-currency framework allows for a straightforward measurement of the economic value of the predictive content of order ow and is a pure out-of-sample approach in that it only conditions on past information. Specically, we sort currencies into portfolios to obtain a cross-section of currency excess returns, which mimics the returns to customer trading behavior and incorporates the information contained in (lagged) ows.2 The information contained in customer trades is highly valuable from an economic perspective: We nd that currencies with the highest lagged total order ows (that is, the strongest net buying pressure across all customer groups against the U.S. dollar) outperform currencies with the lowest lagged ows (that is, the strongest net selling pressure across all customer groups against the U.S. dollar) by about 10% per annum (p.a.). For portfolios based on disaggregated customer order ow, this spread in excess returns is even more striking. A zero-cost long-short portfolio that mimics asset managers trading behavior yields an average excess return of 15% p.a., while conditioning on hedge funds ows leads to a spread of about 10% p.a. Flows by corporate customers basically generate no spread in returns, whereas private customers ows even lead to a highly negative spread (about -14% p.a.). In sum, we nd that order ow contains signicant economic value for a dealer with
2

Lustig and Verdelhan (2007) were the rst to build cross-sections of currency portfolios.

access to such information. Hence, the trend towards more market concentration observed in FX markets over recent years clearly benets large nancial institutions acting as dealers and potentially trading on this information in the inter-dealer market. These informational advantages of dealers are further enhanced by the non-anonymous nature of transactions in OTC markets, as trades by dierent categories of customers convey fundamentally dierent information for price movements. What drives the predictive content in ows? We investigate three main channels. First, order ow could be related to the processing of information by market participants via the process of price discovery. According to this view, order ow acts as the key vehicle that impounds views about (economic) fundamentals into exchange rates.3 If order ow contains private information, its eect on exchange rates is likely to be persistent. Second, there could be a price pressure (liquidity) eect due to downward-sloping demand curves (e.g., Froot and Ramadorai, 2005). If a mechanism like this is at play, we are likely to observe a positive correlation between ows and prices for some limited time, followed by a subsequent reversal as prices revert to fundamental values.4 Third, we consider the possibility that order ow is linked to returns due to the dierent risk sharing motives and risk exposures of market participants. For example, order ow could reect portfolio rebalancing of investors tilting their portfolios towards currencies that command a higher risk premium. Related to this, risk sharing could lead to the observed predictability pattern if non-nancial customers are primarily concerned about laying o currency risk and implicitly paying an insurance premium, whereas institutional investors are willing to take on that risk. Discriminating between alternative explanations for the predictive content of order ow, we nd clear dierences across the four segments of end-users. Asset managers ows are associated with permanent shifts in future exchange rates, suggesting that their order ow is
3 See, e.g., Payne (2003), Love and Payne (2008), Evans and Lyons (2002a, 2007, 2008), Evans (2010), and Rime, Sarno, and Sojli (2010). Other papers relate order ow in a structural way to volatility (Berger, Chaboud, and Hjalmarsson, 2009) or directly to exchange rate fundamentals (Chinn and Moore, 2011). 4 Several studies explore the underlying mechanism for the impact of order ow and discuss the evidence in terms of information versus liquidity eects (e.g. Berger, Chaboud, Chernenko, Howorka, and Wright, 2008; Fan and Lyons, 2003; Marsh and ORourke, 2005; Osler, Mende, and Menkho, 2011; Menkho and Schmeling, 2010; Phylaktis and Chen, 2010; Moore and Payne, 2011; Ito, Lyons, and Melvin, 1998).

related to superior processing of fundamental information.5 In contrast, hedge funds ows are merely associated with transitory exchange rate movements, that is, the impact of their trades on future exchange rates is far less persistent. This result is more in line with shortterm liquidity eects but not with fundamental information processing. Corporate customers and private clients ows, however, seem to reect largely uninformed trading. Our results also point to a substantial heterogeneity across customers in their trading styles and risk exposures, giving rise to dierent motives for risk sharing. First, we nd that the trades of various end-user groups react quite dierently to past returns. Asset managers tend to be trend-followers (positive feedback traders) with regard to past currency returns. By contrast, private clients tend to be contrarians (negative feedback traders). The latter nding squares well with recent ndings for equity markets by Kaniel, Saar, and Titman (2008) who show that individual equity investors behave as contrarians, eectively providing liquidity for institutional investors. Dierent from their results, however, private clients do not directly benet from serving as (implicit) counterparties of nancial customers in FX markets. Second, the ows of most customer groups are negatively correlated over short to intermediate horizons, suggesting that dierent groups of end-users in FX markets engage in active risk sharing among each other. It is thus not just via the inter-dealer market that risk is shared in FX markets, as documented by Lyons (1997), but a signicant proportion of risk is shared among end-users in the customer-dealer segment. Third, we nd substantial heterogeneity in the exposure to risk and hedge factors across customer segments. Asset managers trading does not leave them exposed adversely to systematic risk, which suggests that the information in their ows is not due to risk taking but likely reects superior information. Hedge funds, by contrast, are signicantly exposed to systematic risk such as volatility, liquidity, and credit risk. This lends credence to the view that hedge funds earn positive returns in FX markets by eectively providing liquidity and selling insurance to other market participants. For non-nancial customers there is some evidence of hedging but it is not strong enough to fully explain their negative forecast performance arising from
This information processing can come in dierent ways, e.g., a more accurate and/or faster interpretation of macroeconomic news releases, and better forecasting of market fundamentals such as liquidity and hedging demands of other market participants.
5

poor short-term market timing. Our paper is related to prior work on the microstructure approach to exchange rates (e.g., Evans and Lyons, 2002a,b), which suggests that order ow is crucial for understanding how information is incorporated into exchange rates. It is well known from the literature that order ow is positively associated with contemporaneous returns in basically all asset classes; see, e.g., Hasbrouck (1991a,b) for stock markets, and Brandt and Kavajecz (2004) for U.S. bonds. This is a stylized fact which also holds in FX markets, as shown by Evans and Lyons (2002a) and many subsequent studies. There is less clear evidence, however, on whether order ow predicts exchange rates. A few papers have shown that FX order ow contains information about future currency returns but tend to disagree on the source of this predictive power (e.g., Evans and Lyons, 2005; Froot and Ramadorai, 2005; Rime, Sarno, and Sojli, 2010).6 Some other papers fail to nd robust predictive power of exchange rates by order ow in the rst place (see, e.g., Sager and Taylor, 2008). Our work is also related to a dierent strand of recent literature that analyzes the returns to currency portfolios by investigating the predictive power of currency characteristics, such as carry or lagged returns, and the role of risk premia in currency markets.7 Overall, we contribute to the literature in the following ways. We are the rst to show that order ow forecasts currency returns in an out-of-sample forecasting setting by forming order ow portfolios. This multi-currency investment approach provides an intuitive measure of the economic value of order ow for the few large dealers observing these ows. This seems important as earlier papers either did not consider out-of-sample forecasting at all or relied on purely statistical performance measures derived from time-series forecasts of a limited number of currency pairs (e.g., Evans and Lyons, 2005, who study the DEM/USD and JPY/USD crosses). Time-series forecasts are aected by trends in exchange rates, most notably the U.S. dollar. Our portfolio procedure, by contrast, studies exchange rate preThere is also evidence that marketwide private information extracted from equity order ow is useful for forecasting currency returns (Albuquerque, de Francisco, and Marques, 2008). 7 Lustig and Verdelhan (2007), Farhi, Fraiberger, Gabaix, Ranciere, and Verdelhan (2009), Ang and Chen (2010), Burnside, Eichenbaum, Kleshchelski, and Rebelo (2011), Lustig, Roussanov, and Verdelhan (2011), Barroso and Santa-Clara (2011) and Menkho, Sarno, Schmeling, and Schrimpf (2012a,b) all build currency portfolios to study return predictability and/or currency risk exposure.
6

dictability in dollar-neutral long-short portfolios, and it does so out-of-sample over very long time spans compared to the extant FX microstructure literature. Moreover, we are the rst to test whether risk exposure drives the information in customer order ows. We show how dierent key FX market players trade, e.g., to which extent they rely on trend-following or behave as contrarians, and in which ways they are exposed to systematic risk. We nd strong evidence of heterogeneity in the exposures and trading behavior across dierent groups of market participants. These ndings indicate that there is signicant risk sharing between nancial and non-nancial customers as well as between dierent groups of nancial customers (leveraged versus real money managers). Taken together, these results have implications for our general understanding of information ows in dark markets and how large dealers in OTC markets benet from observing a large proportion of the order ow. These results also add to our general understanding of how risk is shared in nancial markets due to dierent motives for trade and trading styles across end-user segments. The rest of the paper is structured as follows. Section I describes our data, Section II presents empirical results on the predictive power of order ow, Section III empirically investigates alternative underlying reasons for why order ow forecasts FX excess returns, and Section IV presents results of robustness tests. Section V concludes.

I.

Data

We employ a unique dataset based on daily customer order ows for up to 15 currency pairs over a sample period from January 2, 2001 to May 27, 2011, for a total of 2,664 trading days. In contrast to much of the earlier literature, we employ order ow from the enduser segment of the FX market and not from the inter-dealer market. This is important since microstructure models suggest that the information in ows stems from trading with customers and not from inter-dealer trading (e.g. Evans and Lyons, 2002a). Order ows in our sample are measured as net buying pressure against the U.S. dollar (USD), that is, the U.S. dollar volume of buyer-initiated minus seller-initiated trades of a currency against the 6

USD. The data cover all trades of customers (end-users) with UBS during our sample period. A positive number indicates net buying pressure in the foreign currency relative to the USD. Order ow therefore does not measure trading volume but net buying (or selling) pressure, as mentioned above. Our order ow data are available both in aggregated form and at a higher level of granularity allowing for a dierentiation across end-user groups. Aggregate order ow. Aggregate order ows, that is, aggregated across customers (regardless of their type), are available for the following 15 currencies: Australia (AUD), Brazil (BRL), Canada (CAD), the Euro (EUR), Hong Kong (HKD), Japan (JPY), Sweden (SEK), Mexico (MXN), New Zealand (NZD), Norway (NOK), Singapore (SGD), South Africa (ZAR), South Korea (KRW), Switzerland (CHF), and the United Kingdom (GBP). In the following, we refer to these ows as total ows since they are aggregated across all customers of UBS. A natural question is whether ows by customers of UBS are generally representative of end-user currency demands in the FX market. While this question cannot be answered without knowledge of the customer ows of all other dealers, there are good reasons to believe that the ows employed in our paper are highly correlated with a large portion of end-user order ows. First, UBS is among the largest dealers in the FX market and their average market share (according to the Euromoney FX Survey) over our sample period amounts to about 13%. Over most of our sample period, UBS was ranked among the top three of all FX dealers (with Deutsche Bank, Barclays, and Citi usually being the closest competitors). Thus, UBS clearly is one of the most important FX dealers with a signicant portion of the market solely on its own.8 Second, a handful of top dealers in the FX market account for more than 50% of total market share (e.g., King, Osler, and Rime, 2012) and all of these large dealers essentially have access to the same set of large customers. Hence, it seems very likely that UBS ows are highly correlated with ows observed at, e.g., Deutsche Bank, Barclays, Citi, or JP Morgan, which in turn implies that our order ows are representative of the top end of customer trading in the FX market.
Note that most UBS FX customers are in fact big players and include other banks, many large asset management rms and hedge funds, and a large fraction of wealthy private clients. According to the Euromoney survey, UBS has a particularly high market share in FX business with nancial customers (banks, real money and leveraged funds).
8

Table I about here Table I shows descriptive statistics for total ows (in billion USD). Daily order ows are largest on average for EUR, JPY, and CHF. The pair with the largest average imbalance (in absolute value) between buyer- and seller-initiated trading volume is the EUR/USD, where customers (on a net basis) sold on average 63 million EUR against USD per trading day over our sample period. Hence, average order ows are fairly small relative to gross daily trading volume in FX markets.9 Flows are fairly volatile, however, which means that order ow imbalances can frequently be very large. Daily ows tend to be positively autocorrelated, but the degree of autocorrelation is very small albeit sometimes statistically signicant. There is also a clear pattern in standard deviations. Major currencies, such as the EUR, CHF, JPY, GBP, have much larger variation in order ows and, hence, a larger absolute size of order ows compared to other currencies and especially emerging markets. This is intuitive as there is much more trading in major currencies, but it also suggests that one cannot easily compare order ows across currencies and that some form of standardization is needed to make sensible comparisons.10 We take this into account in our empirical analysis below. Finally, aggregate order ows display a high kurtosis (especially the British pound), which is largely driven by some days with extremely high (in absolute value) order ows. Eliminating these few outliers does not change our results reported below. Disaggregated order ow. We also obtain order ows disaggregated by customer groups for the same sample period, albeit only for a subset of nine major currencies.11 There are four customer groups for which ows are available: Asset Managers (AM), Hedge Funds (HF), Corporate Clients (CC), and Private Clients (PC). The segment of asset managers comprises real money investors, such as mutual funds and pension funds. Highly leveraged traders and short-term oriented asset managers not included in the asset managers segment are classied
To provide a benchmark, daily gross spot turnover in the Euro/USD pair in April 2010 amounted to USD 469 billion according to the most recent FX triennial survey (BIS, 2010). These (gross) gures for both the customer-dealer segment and the inter-dealer market are based on data collected from about 4,000 reporting dealers worldwide. 10 In addition, the volatility of ows also varies over time and ows tend to become increasingly volatile towards the end of the sample. Also for this reason, some form of standardization is necessary. 11 The nine currencies are: AUD, CAD, EUR, JPY, SEK, NZD, NOK, CHF, and GBP.
9

as hedge funds. Hedge funds are unregulated entities, whereas asset managers are regulated. The corporate segment includes non-nancial corporations that import or export products and services around the world or have an international supply chain. Corporates also include the treasury units of large non-nancial corporations, with the exception of those pursuing a highly leveraged investment strategy, which are classied by UBS as hedge funds. The last segment, private clients, includes wealthy clients with investable liquid assets in excess of 3 million U.S. dollars. Private clients trade primarily for nancial reasons and with their own money. Hence, there is substantial heterogeneity in the motives for market participation by these four customer types, and the groups are likely to dier considerably in the degree of informedness and sophistication. One of the key features of OTC markets, that is, the nonanonymous nature of transactions can thus further enhance the informational advantages of dealers in dark markets. The order ow data are assembled as follows. Each transaction booked in the UBS execution system at any of its world-wide oces is tagged with a client type. At the end of each business day, global transactions are aggregated for each customer group. Order ow is measured as the dierence between the dollar value of purchase and sale orders for foreign currency initiated by a particular UBS customer group. The transaction is recorded with a positive sign if the initiator of the transaction (the non-quoting counterparty) is buying foreign currency and vice versa.12 Summary statistics for the disaggregated order ow data are reported in Table A.1 of the Internet Appendix. Exchange rate returns and excess returns. For our empirical analysis below, we complement these order ow data with daily spot exchange and forward rates from Reuters (available from Datastream). We denote log changes of spot exchange rates as exchange rate
Our data are raw order ow data with ltering limited to the most obvious cases. For instance, data are adjusted for large merger and acquisition deals which are announced well in advance. Cross-border mergers and acquisitions involve large purchases of foreign currency by the acquiring company to pay the cash component of the deal. These transactions are generally well-publicized and thus are anticipated by market participants. Finally, FX reserve managers, UBS proprietary traders and small banks not participating in the inter-dealer market are excluded from the data. Flows from FX reserve managers are stripped out due to condentiality issues, ows from proprietary traders because they trade with UBS own money, while small banks represent small customers less concerned about the FX market.
12

returns st+1 = st+1 st , (1)

where lowercase letters refer to logs and all exchange rates are quoted as the USD price of foreign currency, so that positive exchange rate returns correspond to an appreciation of the foreign currency. Hence, a positive correlation of order ows and exchange rate returns means that net buying pressure in the foreign currency (against the USD) is associated with an appreciation of the foreign currency (against the USD) and vice versa. We also compute currency excess returns which account for the interest rate dierential in a foreign currency position. Hence, currency excess returns rx are given by

rxt+1 = st+1 st + (it it ),

(2)

where i denotes the foreign interest rate and it denotes the U.S. interest rate. Since we are working at the daily frequency in our main analysis, we need to obtain daily interest rates for all 15 countries (plus the U.S. interest rate). However, since one-day interest rates are not directly available for all countries in our sample, we employ information in forward rates to infer interest rate dierentials. Interest rate dierentials for horizon k are commonly approximated by ik,t ik,t st fk,t where fk,t denotes the log forward rate for horizon k of a given currency.13

II.

The Value of Information in Customer Flows

A. Portfolios Conditioning on Aggregate Order Flow We rely on a portfolio approach, mimicking the returns to customer FX trading by conditioning on lagged order ow. This provides a straightforward and intuitive assessment of the
This approximation is exact if covered interest rate parity (CIP) holds, which tends to be the case at daily or even shorter horizons in normal times (Akram, Rime, and Sarno, 2008). There have been violations of this no-arbitrage relation over the recent nancial crisis. As we show below, the results in this paper are entirely driven by changes in spot rates, whereas interest rate dierentials only play a negligible role. Thus, the results do not depend on whether CIP holds or not.
13

10

economic value of order ow in predicting currency excess returns. As a benchmark test, we rst sort currencies into portfolios based on (lagged) total order ows for each currency. Specically, we sort currencies into ve portfolios (P1 , P2 , ..., P5 ) depending on their total order ow on day t and compute portfolio excess returns (or spot exchange rate changes) for the following day. In this basic setup, portfolios are rebalanced at the end of each trading day. Note that these portfolios are computed from the viewpoint of a U.S. investor as each individual portfolio consists of a short position in USD and a long position in a basket of foreign currencies. Taking the return dierence between any two portfolios Pj Pi thus gives the return of a portfolio short in the basket of foreign currencies in Pi and long in the basket of currencies in Pj , so that the USD component cancels out and the long-short portfolio is dollar-neutral by construction. Standardizing order ows. Before sorting currencies into portfolios, we need to make sure that order ows are comparable across currencies. As the absolute size of order ows diers across currencies (as shown above in Table I) it is not sensible to sort currencies based on raw order ows. To allow for meaningful cross-currency comparisons, it is necessary to standardize ows. We do this by dividing ows by their standard deviation to remove the dierence in absolute order ow sizes across currencies xR j,t = xj,t (xj,t59;t ) , (3)

where xR j,t denotes order ow standardized over a rolling window and xj,t denotes the raw order ow. In our baseline results, we compute the standard deviation of ows via a rolling scheme over a 60-day rolling window. Robustness tests based on alternative approaches to standardize ows are reported in a separate Internet Appendix.14 Portfolio excess returns. Table II shows average annualized excess returns for order ow
In these robustness exercises, we also report results with longer rolling windows of up to three years as well as for an expanding window. Furthermore, we provide tests where we standardize both with respect to volatility as well as the mean. Finally, we also consider a standardization scheme based on gross FX turnover data for dierent currencies drawing on data from the BIS FX triennial survey. These tests, reported in the separate Internet Appendix to conserve space, show that our results are not sensitive with regard to the way ows are standardized.
14

11

portfolios (P1 , P2 , ..., P5 ), where P1 contains the three currencies with the lowest lagged standardized order ow and P5 contains the three currencies with the highest lagged standardized order ow. Hence, P5 can be thought of as a portfolio of currencies with the highest buying pressure, whereas P1 refers to a portfolio with the strongest selling pressure. Column Av. shows average returns across all currencies in the cross-section and column BMS denotes a portfolio which is long in P5 and short in P1 (Buying Minus Selling pressure). We report returns for the full sample period from January 2001 to May 2011.15 To get started, Panel A of Table II reports results for the sample of all 15 markets (T15) as well as for the sub-sample of 9 developed markets (T9); for the T9 sub-sample we only form four portfolios rather than ve to ensure we always have two currencies in the corner portfolios. We observe a strong increase in average excess returns as we move from the portfolio of currencies with low buying pressure P1 to the one with high buying pressure P5 (or P4 for the T9 sample). The spread in excess returns between the high buying pressure and the low buying pressure portfolio, that is, the excess return of the BMS portfolio, is economically large (10.31% and 12.43% p.a., respectively) and statistically highly signicant. Similarly, the Sharpe Ratios (p.a.) of the two BMS portfolios of 1.26 and 1.45 are large and also point toward high economic signicance. Thus, order ows carry signicant information for future currency excess returns, as captured by our dollar-neutral out-of-sample trading strategy which only conditions on real-time information. These results demonstrate the economic value for the owner of this (private) information, that is, the few large FX dealer banks which observe a signicant share of end-user order ow and are able to trade on this information in the inter-dealer market.

Table II about here Table A.3 in the Internet Appendix shows results for the other standardization schemes and for sub-samples. We nd that our results are equally strong in various sub-periods. Likewise, Table A.4 in the Internet Appendix shows the same exercise for exchange rate
Sub-sample tests for a pre-crisis subperiod from January 2001 to June 2007, and a crisis/post-crisis subperiod from July 2007 to May 2011 are reported in the Internet Appendix.
15

12

changes instead of excess returns. Results in that table clearly show that the patterns in average spot exchange rate changes across portfolios are at least as pronounced as for average excess returns or, if anything, even more pronounced. Hence, order ow is informative about future spot rates and not about interest rate dierentials. Tests for return monotonicity. The last three columns MR and Up in Table II report tests for return monotonicity (Patton and Timmermann, 2010), that is, whether there is a signicantly increasing or decreasing pattern of average excess returns when moving from the portfolio of low buying pressure (P1 ) to the one with high buying pressure (P5 ).16 These tests go beyond the standard t-test of a zero BMS portfolio return since they take into account the whole cross-sectional pattern. This is interesting since one would intuitively expect an increasing pattern of average portfolio excess returns when moving from P1 to P5 if order ow is truly informative about future excess returns. This prediction is signicantly borne out in the data for both the T15 and T9 sample of countries and for both the MR and Up test. Hence, there is strong evidence for a signicant relationship between order ow and future excess returns. Excess returns over time. Finally, we plot cumulative excess returns for the T15 and T9 BMS portfolios in the upper left and right panel of Figure 1. As can be seen, excess returns are quite striking and stable for most of the sample period, although somewhat more volatile at the beginning and towards the end of the sample.

Figure 1 about here

The MR statistic tests for a monotonically increasing return pattern, whereas the Up (Down) test is somewhat less restrictive and simply tests for a generally increasing (decreasing) pattern without requiring monotonicity in average portfolio returns. Specically, the MR test requires that the return pattern is monotonically increasing P1 < P2 < ... < P5 and formulates the null hypothesis as H0 : 0 and the alternative hypothesis as Ha : mini=1,...,4 i > 0, where is a vector of dierences in adjacent average portfolio excess returns (P2 P1 , P3 P2 , P4 P3 , P5 P4 ) and i is element i of this vector. The Up test formulates the null hypothesis of a at pattern H0 : = 0 and the alternative hypothesis as 4 Ha : n=1 | i |1{ i > 0} > 0, so that the test is less restrictive and also takes into account the size and magnitude of deviations from a at return pattern. The Down test follows in an analogous way.
16

13

B. Portfolios Conditioning on Disaggregated Order Flow If superior information processing or genuine forecasting ability drive our results above, one expects clear dierences in the forecasting power of dierent customers ows, depending on the groups characteristics (see, e.g., Fan and Lyons, 2003; Evans and Lyons, 2007, among others). Specically, one would expect to see superior information in ows of nancial customers, given that non-nancial players do not specialize in FX trading as their core activity. To investigate this, we now build portfolios based on our disaggregated data for customer ows. We closely follow the earlier approach with the exception that we only build four portfolios (rather than ve) here since we only have disaggregated ows for nine currencies and want to have a minimum of two currencies per portfolio. Table II, Panel B, reports results for the four customer groups: Asset Managers (AM), Hedge Funds (HF), Corporate Clients (CC), and Private Clients (PC). Results are clearcut. Asset managers net buying or selling pressure of currencies is the most informative about subsequent exchange rate behavior. Conditioning on asset managers ows generates a cross-sectional spread in excess returns of 15% p.a., followed by hedge funds with a spread of about 10%. In stark contrast, corporate clients and private clients ows actually generate a negative spread in portfolio excess returns of about 4% and 14%, respectively.17 The results point towards substantial dierences in the customers predictive information and provide a quantitative summary of the value of this information in economic terms. The latter is underscored by the large spread in (annualized) Sharpe Ratios of BMS portfolios across customer groups. The asset managers BMS portfolio yields a Sharpe Ratio of 1.79, whereas the private clients BMS portfolio has a Sharpe Ratio of -1.55.18 As above, we also present p-values for tests of return monotonicity. Since order ow of
Table A.5 shows results for spot rate changes instead of excess returns, which display no qualitative dierences. 18 Table A.6 in the Internet Appendix also shows that excess returns to the BMS portfolios based on dierent customers ows are not highly correlated. Hence, the information contained in the dierent ows appears to stem from dierent sources. In practice, this also means that BMS portfolios could be combined to obtain even higher Sharpe Ratios. For example, a combined portfolio long in the asset managers BMS portfolio and short in the private clients BMS portfolio yields an annualized Sharpe Ratio of 2.19, which is substantially higher than the individual Sharpe Ratios.
17

14

corporate and private customers negatively forecasts returns, we modify the MR test in these cases to test for a monotonically decreasing pattern. Results from these tests corroborate the simple t-tests for the BMS portfolios. There is a monotonically increasing pattern in average excess returns for portfolios based on asset managers and hedge funds ows which is highly signicant. By contrast, we nd a monotonically decreasing pattern in average excess returns for portfolios based on private customers ows, and marginally signicant evidence for a decreasing pattern in portfolios based on corporate ows. Hence, it is not the case that all order ow is equal in terms of its information content for exchange rates. Instead, nancial customers ows (asset managers and hedge funds) account for the positive relation between lagged ows and future exchange rate returns uncovered in the previous section. Flows by corporates are more or less uninformative, whereas private clients ows even forecast returns in the wrong direction. Using total end-user order ow, which is likely to be dominated by nancial customers due to their higher trading volume19 , masks these dierences and might even lead to wrong inference about the link between ows and returns. In a nutshell, what matters for the relation between end-user order ows and future returns is disaggregated data since the information content of ows for future returns varies markedly across customer groups. The middle and lower panel of Figure 1 shows cumulative returns for all four customer groups. It can directly be seen that returns are very dierent across customer groups, even when comparing, for example, asset managers and hedge funds. Both groups BMS portfolios generate signicant excess returns but returns for hedge funds are much more volatile than those of asset managers. Hence, we will investigate possible sources of these dierent behaviors of returns below.

C. Marginal Predictive Content of Flows at Longer Horizons Our analysis so far has been concerned with the relation between order ows and returns over the subsequent trading day. An interesting question, however, is whether the information
19 This is especially true for the order ow employed in this paper since UBS is one of the largest dealers in FX and has a high proportion of nancial customers (relative to corporate clients).

15

contained in order ow quickly decays or whether it is useful for forecasting returns over more than one trading day. We examine the marginal predictive content of ows by forming portfolios as in the analysis above, but we now allow for a longer lag between the order ow signal and portfolio formation. Table III contains the results for dierent lags of 0, 1, 2, . . . , 9 days. To be more specic, a lag of 0 days means that ows of trading day t are used to predict returns of day t + 1 (and thus reproduces BMS returns from Tables II above), whereas a lag of, e.g., 2 days means that ows of day t are used to forecast returns of trading day t + 3.

Table III about here Results in Table III show that order ow appears to be most informative for the rst two to three days after portfolio formation and that the information in ows becomes insignicant afterwards. Hence, the information contained in daily ows is fairly short-lived and is impounded relatively quickly into exchange rates, especially considering that the order ows employed here are private information that is available to only a small number of large FX dealers and that could not realistically be incorporated into prices without some lag. This nding is in contrast to, e.g., Evans and Lyons (2005) who study a shorter and smaller sample and nd that times-series predictability of returns by order ow increases at longer horizons when judged from statistical metrics of forecast evaluation. This contrast in results also highlights the importance to assess the predictive power of order ow using measures of economic value as opposed to purely statistical ones, as statistical evidence of exchange rate predictability in itself does not guarantee that an investor can earn prots from a trading strategy that exploits this predictability.

D. Order Flow vs. Carry and Momentum To further learn about the predictive content of customer order ow for future FX returns, we run panel regressions, which allows us to control for other possible determinants of currency excess returns as well as cross-sectional and time xed-eects. For example, it could be the 16

case that asset managers and hedge funds order ow mimicking portfolios simply reproduce a carry trade (Burnside, Eichenbaum, Kleshchelski, and Rebelo, 2011; Lustig, Roussanov, and Verdelhan, 2011; Menkho, Sarno, Schmeling, and Schrimpf, 2012a) or that their order ow just picks up momentum eects in currency returns (Menkho, Sarno, Schmeling, and Schrimpf, 2012b). Specically, we run panel regressions of the general form rxj,t+1 = c OFtc + 1 (ij,t it ) + 2 rxj,t + 3 rxj,t60;t1 + j,t+1 (4)

where j (1, ..., N ) indexes currencies, rx denotes currency excess returns, OF c denotes order ow of customer group c, (ij,t it ) denotes interest rate dierentials (carry), and rxt and rxt60;t1 denotes lagged excess returns over the prior trading day and the average over the past 60 trading days, respectively.20 The error term is given by j,t+1 = et+1 + uj +
j,t+1

and

thus captures time and cross-sectional xed-eects (we also report results without xed-eects below). Standard errors are clustered by currency pair. Note that these panel regressions employ non-standardized order ows and are based on individual currency returns and not on portfolio returns. Results are shown in Table IV and corroborate our ndings based on our portfolio approach above, namely that order ows of nancials positively predict future excess returns, whereas ows by non-nancial end-users negatively forecast returns. Based on specication (vi) in Table IV, the coecients on lagged order ow imply that a positive order ow of USD 1 billion forecasts a four basis point (b.p.) higher excess return on the following day for asset managers ows, a one b.p. higher return for hedge funds, a minus one b.p. lower return for corporates, and a minus two b.p. lower return for private clients. The magnitude of these eects seems reasonable given the deep liquidity of the FX market.

Table IV about here More important, however, is the fact that the predictive relation between lagged order
20

Using other windows of less or more than 60 trading days does not yield qualitatively dierent results.

17

ow and future FX excess returns remains very strong when controlling for two common predictors of returns in FX markets, interest rate dierentials and (short-term) momentum. Carry shows up with a positive sign, that is, high interest rate currencies deliver high excess returns on average in line with the large literature on the forward discount bias (e.g., Fama, 1984). Interest rate dierentials, however, do not drive out the information contained in order ows and they become insignicant once we include cross-sectional xed-eects in the regression. In our panel regressions, lagged currency returns do not have consistent predictive content beyond order ow and carry, which is in line with recent evidence in Menkho, Sarno, Schmeling, and Schrimpf (2012b), who show that FX momentum strategies are not protable for major exchange rates over the last decade.

III.

What Drives the Predictive Power of Flows?

A. Permanent vs. Transitory Forecast Power of Flows To better understand the driving forces behind our results above, we next investigate whether order ow forecasts returns because it signals permanent shifts in spot exchange rates or whether it merely forecasts temporary movements which are eventually reversed after some time. The question whether order ow has a permanent or transitory eect in prices is a central theme in the earlier microstructure literature (see Hasbrouck, 1991a,b). A transitory movement is interpreted as suggesting that order ow eects are merely due to short-term liquidity or price pressure eects which eventually die out, whereas a permanent movement in spot rates would indicate that order ow conveys information about fundamentals.21 More specically, a permanent price impact would most probably indicate that order ow is related to changes in expectations about fundamentals given the daily frequency we are working on. This question is relevant for our analysis since we nd substantial heterogeneity with
One strand of literature argues that order ow is the conduit by which information about fundamentals is impounded in prices and therefore has permanent eect on exchange rates (e.g., Evans and Lyons, 2002a; Brandt and Kavajecz, 2004; Evans and Lyons, 2007, 2008). Another strand of the literature suggests that order ow matters due to downward sloping demand curves or illiquidity and, hence, that order ow only has a transitory impact on prices (e.g., Froot and Ramadorai, 2005).
21

18

regard to the forecasting power of dierent customer groups order ows. Therefore it is particularly interesting to nd out if all (or some) customers ows signal information relevant for permanent changes in FX rates or whether some customer groups order ow simply exerts price pressure and liquidity eects. To this end, we apply our portfolio sorts framework as above but now track cumulative exchange rate returns to BMS portfolios for overlapping periods of 30 trading days after portfolio formation. This approach yields a direct estimate of how spot rates move after experiencing intensive buying or selling pressure by customers. Figure 2 illustrates the persistence of the predictive content of order ow. The solid lines show the cumulative excess returns (in basis points), whereas the shaded areas show 95% condence intervals based on a moving-block bootstrap with 1,000 repetitions. Total ows for all 15 currencies (T15) forecast a permanent change in spot rates which is statistically signicantly dierent from zero. Exchange rates with the highest net buying (selling) pressure appreciate (depreciate) against the USD for approximately three days. Currency returns on the BMS portfolios increase by about 15 basis points over this period, and afterwards the eect of the order ow signal levels out. Importantly, these ndings suggest that order ow conveys information and its impact on exchange rates is not reversed.

Figure 2 about here This picture changes when looking only at the nine developed currencies. Here, we observe the same increasing pattern initially, followed by a subsequent partial reversal. After approximately 25 30 trading days, about one half of the initial impact of 15 basis points is reversed and the condence interval includes zero. Hence, there is much less evidence that order ow conveys information about fundamentals when only looking at major developed markets. This nding makes sense, however, since the major currency markets are most probably more researched and more ecient than smaller currency markets so that the scope for superior information processing is reduced.22
This may be interpreted in the context of the adaptive markets hypothesis (see e.g. Neely, Weller, and Ulrich, 2009, for an analysis in FX markets).
22

19

As a natural next step, we also examine the same question separately for disaggregated order ows (lower panels of Figure 2). Results are clear-cut. The only end-user group with a statistically signicant permanent price impact is asset managers. Hedge funds trading has a positive but transitory impact in line with an interpretation that they provide liquidity. Corporate clients have no impact at all, and private clients have a transitory negative impact. Given our nding for total ows of the nine major currencies above, it is interesting to see that asset managers ows are indeed associated with permanent spot rate changes. Hence, order ow of asset managers seems to be related to the processing of fundamental information whereas hedge funds order ow corresponds to short-lived information unrelated to fundamental information. Similarly, it seems reasonable that the negative relation between private clients ows and future spot rates eventually dies out over time. These ndings are novel in the literature and suggest that order ows by dierent enduser groups even by the two nancial customer groups embed dierent information for future exchange rates. These dierences can arise either because they are based on dierent mechanisms to process information or because of dierent trading motives and hedging needs. To explore this further, we investigate the drivers of order ow in more detail and shed light on the observed dierences in end-user order ows.

B. Risk Sharing Among Foreign Exchange End-Users The analysis above suggests that asset managers ows are related to the processing of fundamental information that is quickly, but permanently, impounded into prices whereas the other customer groups ows are not. A potential explanation is that risk sharing among market participants drives (part of) our results. For instance, private clients negative BMS returns could be explained by their possible need for hedging FX risk, whereas the positive returns of hedge funds might implicitly reect a compensation for taking on such risks. While these are just examples, a risk sharing story in general implies that we observe customers systematically trading in opposite directions and that their portfolios load on dierent sources of systematic risk. We investigate these issues below.

20

Portfolio returns in event time. We rst provide a more detailed look at the return behavior around portfolio formation dates to better understand dierences in customer groups. Figure 3 shows the average annualized BMS excess return for the ve days prior to portfolio formation (5, 4, ..., 1), the day of portfolio formation 0, and the rst ten days after portfolio formation (1, 2, ..., 10). Shaded areas correspond to 95% condence intervals based on Newey and West (1987) standard errors. Note that these returns, unlike Figure 2, are not cumulative. Figure 3 about here Two results stand out. First, asset managers tend to be trend-followers in that they exert buying (selling) pressure in currencies that recently appreciated (depreciated). Conversely, private clients tend to trade against the trend, that is, they react upon past returns in a contrarian fashion. The pattern for hedge funds and corporates is less clear. Second, formation day returns (day 0) are signicantly dierent from zero for all four customer groups. However, hedge funds (positive) and private clients (negative) have the largest contemporaneous returns in absolute value, indicating that their trading either heavily drives exchange rates or is heavily triggered by returns (e.g., via stop-loss and stop-buy orders). The latter explanation seems more reasonable especially for private clients who do not trade large enough volumes to move prices in FX markets. Overall, these ndings suggest that customer groups trading positions at least partly oset each other, as asset managers and private clients clearly dier in terms of their trendfollowing behavior. This nding is dierent from equity markets where Kaniel, Saar, and Titman (2008) nd that individual investors also tend to be contrarian traders but that they experience subsequent positive returns, presumably due to implicitly providing liquidity to institutional investors. In our data, we nd a similar contrarian behavior of individual investors, but this trading behavior does not yield positive returns on average. Flow correlations over longer horizons. Given these ndings, we next look at the correlation of customer groups ows directly. While there is little contemporaneous correlation in ows, as noted above (see Table A.2 in the Internet Appendix), it is nevertheless interesting 21

to look at ows over longer horizons to nd out if customer groups tend to trade in the same or in opposite directions. For a risk sharing explanation to make sense, we would expect to see negative ow correlations between customer groups at some horizons. Figure 4 plots contemporaneous correlations between ows of all four customer groups for horizons of one to 60 days (using overlapping observations) where the shaded areas correspond to 95% bootstrap condence intervals. For the two nancial customer groups, there is a small and short-term negative ow correlation which turns positive after three days. Hence, asset managers and hedge funds tend to trade in opposite directions over very short horizons but in the same direction over the longer run. Moreover, all correlations between nancial and non-nancial customers are signicantly negative at all horizons while there is no signicant correlation between ows of the non-nancial customer groups. These results are generally in line with a risk sharing story where nancial players trade in the opposite direction of non-nancial market participants. This nding is interesting because the perception in the literature is that risk sharing takes place in the inter-dealer market (see, e.g., Lyons, 1997) where dealers quickly lay o their accumulated inventory from customer orders. However, the high concentration in todays FX market implies that large dealers can match customer trades to a large extent internally, allowing them to manage their inventory more eciently. Given the negative correlation of ows we observe in the data, there clearly seems to be scope for such warehousing of inventory risk (also see King, Osler, and Rime, 2012, on this topic).

Figure 4 about here Drivers of ows. As a natural next step we seek to provide a better understanding of the drivers of end-user order ows and shed light on the source of the negative ow correlations discussed above. First, we examine whether the ows of some customer groups systematically lead the ows of other groups. Second, we study whether customers ows dier in their response to lagged asset returns in other key asset classes. In this context we are interested in the possible eects of portfolio re-balancing on the end-user demand for currencies (Hau and Rey, 2004). To investigate this, we run panel regressions of order ows on lagged ows and further explanatory variables, such as interest rate dierentials (it it ), lagged exchange 22

eq eq rate changes over one and 20 days (st , st1;t20 ), lagged stock returns (rt , rt1;t20 ), and b b lagged bond returns (rt , rt1;t20 )

PC CC HF AM c OFj,t +1 = + AM OFj,t + HF OFj,t + CC OFj,t + P C OFj,t

+1 (ij,t it ) + 2 sj,t + 3 sj,t1;t20

eq eq b b +4 rj,t + 5 rj,t 1;t20 + 6 rj,t + 7 rj,t1;t20 + j,t+1 ,

(5)

where c denotes one of the four customer groups, j denotes currencies/countries, and j,t+1 = et+1 + uj +
j,t+1

includes both cross-sectional and time xed-eects. Standard errors are

clustered by currency pair. We use benchmark 10-year government bonds and country equity indices from Datastream for bond and stock returns. The frequency is daily. Results from these regressions are shown in Table V. For each customer group we report one specication which only includes lagged ows and one which additionally includes interest rate dierentials and lagged returns.23 Looking rst at the specications which only include lagged ows, we nd that the ows of asset managers are signicantly related to the ows by the other groups. These results (akin to simple Granger causality tests) indicate again that asset managers trade in the opposite direction of non-nancial customers. Flows by hedge funds, on the other hand, do not load signicantly on lagged ows of any group, which shows that asset managers and hedge funds show a quite dierent behavior. Corporate ows are positively driven by own lagged ows and those of private clients, whereas ows by private clients are signicantly negatively related to lagged hedge funds ows and signicantly positively autocorrelated. In sum, there is a wealth of interrelationships between customer ows and their lags although it seems overambitious to interpret them in any structural way.

Table V about here When including lagged returns as additional regressors, we nd that asset managers trade against the interest rate dierential, whereas corporate customers trade with the interest
Using more than one lag of ows in the regressions generally yields insignicant coecient estimates so we restrict the regressions to include one lag of ows.
23

23

rate dierential. Surprisingly, ows by hedge funds (and private clients) are not aected by the interest dierential suggesting that, on average, carry trading is not a dominant driver of their ows over our sample.24 Results for lagged exchange rates indicate that asset managers are trend-followers (positive feedback traders), whereas private clients can be described as contrarians (negative feedback traders). Asset managers ows also react signicantly positively to lagged equity returns, whereas private clients ows are positively driven by lagged bond returns. Hence, investors tend to increase their position in a currency (against the USD) when the countrys stock market return has been high (asset managers) or when government bond prices went up (private clients). These results do not suggest that order ows are driven by portfolio rebalancing in the sense that investors sell a currency in response to rising equity or bond prices in the country (see, e.g. the mechanism described in Hau and Rey, 2004). However, the results strongly support the notion that ows of dierent groups are to some extent driven by the returns of other asset classes, although the factors that inuence ows clearly dier across end-user groups. Finally, it seems worthwhile mentioning that the estimated overall constant in the panel regression is signicantly negative for hedge funds and corporates but signicantly positive for private clients. Hence, hedge funds and corporates have been net sellers of the U.S. dollar, whereas private clients have been net buyers of U.S. dollar. Given the large current account decit of the U.S. over the last decade, the negative coecient for corporate clients is not surprising. Also, the strong inow of foreign savings into U.S. capital markets over the sample period makes sense of the positive coecient for private clients.
While this result may be surprising, it is worth bearing in mind that it relates to the aggregate hedge funds community and on average over the full sample. It is therefore entirely possible, or even likely, that there is variation across hedge funds and across time: For example, it may well be the case that some hedge funds follow carry trade strategies and some follow uncovered interest parity (anti-carry trade) strategies, or that for a particular hedge fund carry trades were implemented for the rst part of the sample and deleveraged during the second part of the sample which is characterized by the recent global crisis. Thus, our result is not intended to detect the individual behavior of hedge funds.
24

24

C. Dierences in Risk Exposures Finally, we investigate if dierences in risk exposures can account for BMS return patterns across FX end-users. A risk channel could explain the observed BMS excess returns if asset managers and hedge funds tilt their portfolios towards risky currencies and earn a risk premium whereas corporate and private clients tilt their portfolios towards safe currencies and, hence, earn low or even negative returns. Since there are many possible sources of systematic risk that might be relevant in our case, we consider an augmented version of the Fung and Hsieh (2002, 2004) multi-factor model as the basis for these risk adjustments. The Fung-Hsieh model has served as the workhorse for understanding risk exposures in the hedge funds literature (see, e.g. Patton and Ramadorai, 2013). The model relies on various U.S. equity-market and bond-market factors and also includes the returns to trend-following strategies to capture exposure to non-linear optionlike payos that are quite typical of hedge funds. The trend-following factors are constructed from portfolios of lookback straddles in various asset classes. We modify the model to make it amenable to an analysis focused on the FX market and to allow for conditional exposures (e.g. Ferson and Schadt, 1996; Patton and Ramadorai, 2013). The regression which serves as the basis of these tests takes the following form
K J

rxp;t = +
k=1

k F k ;t +
j =1

j rm;t zj ;t1 + t .

(6)

The set of factors Ft includes the excess return on the U.S. equity market (rm ), the change in the yield spread of U.S. long-term bonds (T S ), and changes in credit spreads (DF ). It further includes returns on portfolios of lookback straddles for FX futures and interest rate futures, denoted by P T F SF X and P T F SIR respectively. We augment this sub-set of factors from Fung and Hsieh (2004) by additional factors that are intended to capture FXrelated risk. We include the Dollar risk factor (DOL) and the carry factor (HM LF X ) by Lustig, Roussanov, and Verdelhan (2011) as well as a factor-mimicking portfolio of global FX volatility (V OLF X ) (Menkho, Sarno, Schmeling, and Schrimpf, 2012a). Following Patton

25

and Ramadorai (2013), we also allow for conditional risk exposures by interacting the equity market risk factor rm;t with lagged conditioning variables zj ;t1 . We consider (a) changes in the TED spread (Brunnermeier, Nagel, and Pedersen, 2009), (b) changes in the VIX (Whaley, 2000), and (c) the change in the 3-month T-Bill rate. To keep the analysis tractable and to avoid overtting, we perform model selection of the space of risk factors. Ideally, we want to explore the same set of factors for each of the customer segments to be able to compare the exposures across customers and learn about dierences which might explain the variation in BMS excess returns. However, as nancial and non-nancial customers are likely to be very dierent, we focus on asset managers versus hedge funds in the rst set of results and private clients versus corporates in the second set of results. More specically, we perform model selection over a two-equation seemingly unrelated regression (SUR) for asset managers and hedge funds BMS returns, and a separate model selection for a SUR for corporate and private clients. Results for asset managers and hedge funds are shown in Table VI. Panel A shows results for linear models, whereas Panel B allows for conditional market exposures. We report the four best performing models with a maximum of three factors included in the regression. The best linear model in Panel A picks global FX volatility (V OLF X ) as the single factor. Other model specications which also perform well tend to incorporate the trend-following factors as well as term spread and default spread changes. Interestingly, when comparing asset managers and hedge funds exposures to these factors, we nd that the signs are always opposite. While asset managers BMS returns load positively on FX volatility shocks, trendfollowing factors, and changes in the default spread, hedge funds load negatively on these factors. Hence, asset managers FX trading positions tend to perform well in periods of market-wide stress and when there are large returns to trend-following (which happens to be in volatile periods, when markets trend more). Hedge funds FX trading positions, however, are adversely exposed to systematic risk and market distress. These results are quite striking as they indicate that asset managers show a very dierent FX trading behavior and exposure

26

to systematic risk than hedge funds.25 Table VI about here Allowing for conditional exposures by adding interaction terms of market returns (rm ) with lagged changes in TED spreads and the VIX (Table VI, Panel B) leaves the main factors chosen largely unchanged but tends to improve the model t. The results reported in Panel B corroborate the previous results. The equity market exposure of asset managers tends to decrease when the lagged TED spread and VIX increase, and vice versa for hedge funds exposures. This is further evidence that the trading by asset managers and hedge funds is very dierent and that their FX trading positions are exposed dierently to market stress. It should be noted, though, that the alphas of asset managers and hedge funds are signicantly dierent from zero and still quite large, that is, exposure to risk does not drive the information in ows for excess returns to zero.26 Table VII about here We repeat the analysis above for corporate and private clients BMS portfolios as well, and results are shown in Table VII. However, as might be expected, risk exposures do not matter as much for non-nancial customers. Still, we nd a negative equity market exposure for both groups (Panel A), which increases (decreases) following increases in the TED spread for private (corporate) clients. Moreover, there is some evidence that the private clients BMS portfolio has positive exposure to changes in credit spreads.

IV.

Additional Tests and Robustness

We provide extensive robustness checks to all our main results. Below, we rst examine the eect of transaction costs and then turn to a brief discussion of various other robustness tests
25 Additional evidence is provided in the Internet Appendix. Table A.17 summarizes exposures to equity factors, Table A.18 considers FX factors, Table A.19 focuses on the Fung and Hsieh (2002) factors, and Table A.20 reports results for the BMS portfolio based on total ows for completeness. 26 Table A.16 reports pricing errors for the cross-section of order ow portfolios. Specically, we report the Gibbons, Ross, and Shanken (1989) test for the null that the alphas are jointly equal to zero. Corroborating the time-series regressions in Tables VI and VII, the test always rejects the null of zero alphas.

27

for which results can be found in a separate Internet Appendix to conserve space.

A. Transaction Costs Our analysis above is intentionally quiet on questions of exploitability of order ow information for trading strategies or the eects of transaction costs. This is because our data on customer order ow are not available to participants in the broader market and thus cannot form the basis for a trading strategy, except for UBS itself or for one of the other few large dealers with access to similar customer ows. However, an interesting issue is whether owners of this type of private information, that is, large FX dealer banks with a large concentration of informed customers, could potentially employ this information by simply piggy-backing the order ow of their customers.27 To examine this question, we compute net excess returns for BMS portfolios by adjusting for bid-ask spreads.28 We investigate returns to strategies with varying portfolio re-balancing frequencies to balance the eects of transaction costs and using the most recent information. Figure 5 presents the results for re-balancing frequencies from 1 to 10 days. The dashed lines give average excess returns (p.a.) and 95% condence intervals for excess returns before transaction costs to show the eect of dierent re-balancing periods. The solid line and shaded area show average net excess returns (p.a.) and 95% condence intervals when taking transaction costs into account.

Figure 5 about here We nd that exploiting the information in ows should in practice be feasible for a dealer. This holds for both sets of currencies T15 and T9. Average excess returns are signicantly
Obviously the data should be used in respect of clients condentiality and the specic compliance agreements governing customers transactions. 28 The bid-ask spread data available is for quoted spreads and not eective spreads. It is well known that quoted spreads are much higher than eective spreads (Lyons, 2001). We therefore follow earlier work, e.g., Goyal and Saretto (2009), and employ 50% of the quoted bid-ask spread as the actual spread. Even this number seems conservative, though. First, banks with access to this kind of customer order ow data are big dealers and pay very low spreads since they are key market makers. Second, Gilmore and Hayashi (2011) nd in a recent study that transaction costs due to bid-ask spreads are likely to be much lower than our 50% rule. This nding was corroborated by our own conversations with UBS dealers.
27

28

dierent from zero for all re-balancing horizons and economically attractive even for short frequencies. These results clearly demonstrate the potential value of being able to observe order ow by customers, especially the one by informed customers such as asset managers or leveraged funds.

B. Further Robustness Checks We check whether our results are robust to other sensible choices of standardizing ows. First, we check whether standardizing ows over longer horizons of one and three years produce similar results (see Tables A.7 and A.8 in the Internet Appendix). They do. Second, we measure ows relative to total currency trading volume (obtained from the BIS FX triennial surveys).29 Table A.9 shows the results, which also indicate signicant predictability of returns by order ows. Third, we standardize ows by additionally demeaning ows over the rolling window (Table A.10). As a nal step, we form portfolios based on ows for all currencies via the following procedure: We cross-sectionally standardize order ows for day t (we subtract the cross-sectional mean and divide by the standard deviation), rescale these standardized ows to sum to two in absolute value, and then use these as weights to form a portfolio held from day t to t + 1 (Table A.11). Our results remain robust. We also check if order ows forecast returns at longer horizons. To this end, we use an exponential moving average to sum order ows into the past and then use these lower frequency ows to build BMS portfolios which we rebalance every 2, 3, 4, 5, 10, 20, 60 trading days. We report results for two dierent decay parameters (0.25 and 0.75) in the exponential moving average in Table A.12. We nd that predictability dies out fairly quickly and that only asset managers ows have some predictive power over longer horizons of up to one month (20 trading days). To rule out that a simple liquidity story drives our predictability results, we also look at a sub-sample of the four most liquid currency pairs in our sample: EUR/USD, JPY/USD,
We linearly interpolate between the data of the BIS survey to obtain a daily time-series of trading volumes in USD for the nine developed currencies and then use the ratio of customer ows to total trading volumes as our sorting variable.
29

29

GBP/USD, and CHF/USD. Table A.13 reports results for BMS portfolio returns and Figure A.1 shows results for BMS returns in event time (similar to Figure 3 in the main text). We nd that our main results remain qualitatively unchanged. We next explore whether a specic currency is driving the protability of the order ow portfolios. To investigate this issue, we rely on a cross-validation setting in which we form portfolios as before but in each case delete one of the available currencies. For example, we exclude the EUR/USD pair and compute BMS portfolio returns for the remaining 14 (total order ows) or 8 currency pairs (disaggregated order ows). Table A.14 summarizes the results from this exercise. We always nd the same general return pattern, that is, our main ndings do not depend strongly on any particular currency.

V.

Conclusion

This paper empirically addresses three related questions to improve our understanding of the ecology of the worlds largest nancial market, the FX market. First, given that the FX market is fairly opaque with a large concentration of market making in the hands of a few large intermediaries, how valuable is it for dealers to observe a large proportion of the markets order ow? Second, do FX end-users share risks among themselves, or is their trading highly correlated and unloaded to the dealers and the inter-dealer market? Third, how can we understand the trading behavior, trading styles and risk exposures of various key players in FX markets, and how is this linked to risk sharing? We nd that observing customer order ows in a dark market is highly valuable from the dealers perspective. Currency excess returns to portfolios mimicking aggregate customer order ows in real-time are about 10% p.a. and highly signicant. In addition, trading in FX markets (as in other OTC markets) is anonymous, meaning that dealers know the identity of their clients. Incorporating this feature into our setup, we nd excess returns as high as 15% p.a., that is, non-anonymity further increases the informational advantage of dealers. The ows by asset managers have the strongest predictive power for exchange rates, likely reecting the processing of fundamental information. Their ows have a permanent forecast 30

power, whereas ows originating from the other groups only predict transitory changes of exchange rates. All this suggests that dealers have a strong incentive to gain large market shares (besides other reasons such as economies of scale in the provision of trading infrastructure, for example) and to set up trading in a way that reveals end-users identities. These ndings about strong information asymmetries and incentives should be useful to inform policy discussions on the appropriate framework for OTC markets. We also show that the main segments of end-users dier markedly in their trading strategies and hedging demands. Asset managers, for instance, tend to be trend-followers, whereas individual investors behave as contrarians. Hedge funds (on aggregate) do not seem to fall in any of these two categories. Moreover, ows of dierent end-user segments tend to be negatively correlated over longer horizons. These ndings suggest that risk sharing also takes place among end-users and not only via the inter-dealer market as suggested by previous FX microstructure research. Taken together, these results bring some light into one of the main dark nancial markets. Our ndings suggest that the FX market is populated by quite heterogeneous market participants and that we gain valuable economic insights from observing their transactions and learning about their dierent predictive ability, trading motives, trading styles, and risk exposures.

31

References
Akram, Q. Farooq, Dagnn Rime, and Lucio Sarno, 2008, Arbitrage in the foreign exchange market: Turning on the microscope, Journal of International Economics 76, 237253. Albuquerque, Rui, Eva de Francisco, and Luis Marques, 2008, Marketwide private information in stocks: Forecasting currency returns, Journal of Finance 63, 22972343. Ang, Andrew, and Joseph Chen, 2010, Yield curve predictors of foreign exchange returns, Working Paper, Columbia University. Barroso, Pedro, and Pedro Santa-Clara, 2011, Beyond the carry trade: Optimal currency portfolios, Working Paper, NOVA School of Business and Economics. Berger, D.W., A.P. Chaboud, S.V. Chernenko, E. Howorka, and J.H. Wright, 2008, Order ow and exchange rate dynamics in electronic brokerage system data, Journal of International Economics 75, 93109. Berger, D.W., A.P. Chaboud, and E. Hjalmarsson, 2009, What drives volatility persistence in the foreign exchange market?, Journal of Financial Economics 94, 192213. BIS, 2010, Triennial central bank survey - report on global foreign exchange market activity in 2010, Bank for International Settlements, Basel. Brandt, Michael W., and Kenneth A. Kavajecz, 2004, Price discovery in the U.S. treasury market: The impact of orderow and liquidity on the yield curve, Journal of Finance 59, 26232654. Brunnermeier, Markus, Stefan Nagel, and Lasse Pedersen, 2009, Carry trades and currency crashes, NBER Macroeconomics Annual 2008 23, 313347. Burnside, Craig, Martin Eichenbaum, Isaac Kleshchelski, and Sergio Rebelo, 2011, Do Peso problems explain the returns to the carry trade?, Review of Financial Studies 24, 853891. Carhart, Mark M., 1997, On persistence in mutual fund performance, Journal of Finance 52, 5782. Chinn, Menzie D., and Michael Moore, 2011, Order ow and the monetary model of exchange rates: Evidence from a novel data set, Journal of Money, Credit and Banking 43, 1599 1624. Due, Darrell, 2012, Dark Markets: Asset Pricing and Information Transmission in Overthe-Counter Markets (Princeton University Press).

32

Evans, Martin D., 2010, Order ows and the exchange rate disconnect puzzle, Journal of International Economics 80, 5871. , and Richard K. Lyons, 2002a, Order ow and exchange rate dynamics, Journal of Political Economy 110, 170180. , 2002b, Time-varying liquidity in foreign exchange, Journal of Monetary Economics 49, 10251051. , 2005, Meese-Rogo redux: Micro-based exchange rate forecasting, American Economic Review, Papers and Proceedings 95, 405414. , 2007, Exchange rate fundamentals and order ow, Working Paper, Georgetown University. , 2008, How is macro news transmitted to exchange rates?, Journal of Financial Economics 88, 2650. Fama, Eugene F., 1984, Forward and spot exchange rates, Journal of Monetary Economics 14, 319338. Fan, M., and Richard K. Lyons, 2003, Customer trades and extreme events in foreign exchange, in P. Mizen, ed.: Monetary History, Exchange Rates and Financial Markets: Essays in Honour of Charles Goodhart . pp. 160179 (Edward Elgar: Northampton). Farhi, Emmanuel, Samuel P. Fraiberger, Xavier Gabaix, Romain Ranciere, and Adrien Verdelhan, 2009, Crash risk in currency markets, Working Paper, MIT Sloan. Ferson, Wayne E., and R. Schadt, 1996, Measuring fund strategy and performance in changing economic conditions, Journal of Finance 51, 425462. Froot, Kenneth A., and Tarum Ramadorai, 2005, Currency returns, intrinsic value, and institutional-investor ows, Journal of Finance 60, 15351566. Fung, William, and David Hsieh, 2001, The risk in hedge fund strategies: Theory and evidence from trend followers, Review of Financial Studies 14, 313341. Fung, William, and David A. Hsieh, 2002, Asset-based style factors for hedge funds, Financial Analysts Journal 58, 1627. , 2004, Hedge fund benchmarks: A risk based approach, Financial Analysts Journal 60, 6580. Gibbons, Michael R., Stephen A. Ross, and Jay Shanken, 1989, A test of the eciency of a given portfolio, Econometrica 57, 11211152.

33

Gilmore, Stephen, and Fumio Hayashi, 2011, Emerging market currency excess returns, American Economic Journal: Macroeconomics 3, 85111. Goyal, Amit, and Alessio Saretto, 2009, Cross-section of option returns and volatility, Journal of Financial Economics 94, 310326. Hasbrouck, Joel, 1991a, Measuring the information content of stock trades, Journal of Finance 46, 179207. , 1991b, The summary informativeness of stock trades: An econometric analysis, Review of Financial Studies 4, 571595. Hau, Harald, and Helene Rey, 2004, Can portfolio rebalancing explain the dynamics of equity returns, equity ows, and exchange rates?, American Economic Review, Papers and Proceedings 94, 126133. Ito, Takatoshi, Richard K. Lyons, and Michael T. Melvin, 1998, Is there private information in the FX market?, Journal of Finance 53, 11111130. Kaniel, Ron, Gideon Saar, and Sheridan Titman, 2008, Individual investor trading and stock returns, Journal of Finance 63, 273310. King, Michael R., Carol Osler, and Dagnn Rime, 2012, Foreign exchange market structure, players and evolution, in Jessica James, Ian Marsh, and Lucio Sarno, ed.: Handbook of Exchange Rates . pp. 344 (Wiley). Love, Ryan, and Richard Payne, 2008, Macroeconomic news, order ows, and exchange rates, Journal of Financial and Quantitative Analysis 43, 467488. Lustig, Hanno, Nikolai Roussanov, and Adrien Verdelhan, 2011, Common risk factors in currency markets, Review of Financial Studies 24, 37313777. Lustig, Hanno, and Adrien Verdelhan, 2007, The cross section of foreign currency risk premia and consumption growth risk, American Economic Review 97, 89117. Lyons, Richard K., 1997, A simultaneous trade model of the foreign exchange hot potato, Journal of International Economics 42, 275298. , 2001, The Microstructure Approach to Exchange Rates (MIT Press: Cambridge). Marsh, Ian W., and C. ORourke, 2005, Customer order ow and exchange rate movements: Is there really information content?, Working Paper, Cass Business School. Menkho, Lukas, Lucio Sarno, Maik Schmeling, and Andreas Schrimpf, 2012a, Carry trades and global foreign exchange volatility, Journal of Finance 67, 681718.

34

, 2012b, Currency momentum strategies, Journal of Financial Economics 106, 660 684. Menkho, Lukas, and Maik Schmeling, 2010, Whose trades convey information? Evidence from a cross-section of traders, Journal of Financial Markets 13, 101128. Moore, Michael, and Richard Payne, 2011, On the existence of private information in FX markets, Journal of Banking and Finance 35, 12501262. Neely, Christopher, Paul Weller, and Joshua M. Ulrich, 2009, The adaptive markets hypothesis: Evidence from the foreign exchange market, Journal of Financial and Quantitative Analysis 44, 467488. Newey, Whitney K., and Kenneth D. West, 1987, A simple, positive semi-denite, heteroskedasticity and autocorrelation consistent covariance matrix, Econometrica 55, 703 708. Osler, Carol L., Alexander Mende, and Lukas Menkho, 2011, Price discovery in foreign exchange markets, Journal of International Money and Finance 30, 16961718. Patton, Andrew J., and Tarun Ramadorai, 2013, On the high-frequency dynamics of hedge fund risk exposures, Journal of Finance forthcoming. Patton, Andrew J., and Allan Timmermann, 2010, Monotonicity in asset returns: New tests with applications to the term structure, the CAPM, and portfolio sorts, Journal of Financial Economics 98, 605625. Payne, Richard, 2003, Informed trade in spot foreign exchange markets: An empirical investigation, Journal of International Economics 61, 307329. Phylaktis, Kate, and L. Chen, 2010, Asymmetric information, price discovery and macroeconomic announcements in FX market: Do top trading banks know more?, International Journal of Finance and Economics 15, 228246. Rime, Dagnn, Lucio Sarno, and Elvira Sojli, 2010, Exchange rates, order ow and macroeconomic information, Journal of International Economics 80, 7288. Sager, Michael, and Mark P. Taylor, 2008, Commercially available order ow data and exchange rate movements: Caveat emptor, Journal of Money, Credit and Banking 40, 583 625. Whaley, R.E., 2000, The investor fear gauge, Journal of Portfolio Management 26, 1217.

35

Table I. Descriptive Statistics for FX Customer Order Flows This table shows descriptive statistics for total order ows for the 15 currencies in our sample. Flows are measured in billions USD and all currencies are against the USD. A positive (negative) ow means that there is net buying (selling) pressure for the respective currency. The frequency is daily and the sample is from January 2001 to May 2011. Currencies included are the Australian Dollar (AUD), Canadian Dollar (CAD), Swiss Franc (CHF), Euro (EUR), Great Britain Pound (GBP), Japanese Yen (JPY), Norwegian Krone (NOK), New Zealand Dollar (NZD), Swedish Krona (SEK), Brazilian Real (BRL), Hong Kong Dollar (HKD), (South) Korean Won (KRW), Mexican Peso (MXN), Singapore Dollar (SGD), and South African Rand (ZAR). Mean Median Std Skew Kurt AC(1) p-val.

Panel A. Developed Markets AUD CAD CHF EUR GBP JPY NOK NZD SEK -0.003 0.007 0.020 -0.063 -0.001 0.027 0.003 -0.002 0.001 -0.001 0.003 0.012 -0.041 -0.002 0.019 0.000 0.000 0.000 0.197 0.169 0.324 0.656 0.484 0.412 0.067 0.070 0.070 -2.69 55.89 1.98 43.02 0.11 74.88 -3.31 79.53 -5.82 270.70 1.88 55.87 0.60 49.20 -1.78 51.30 1.60 39.84 -0.01 0.00 0.02 0.03 0.01 0.03 0.08 0.14 0.01 (0.15) (1.00) (0.00) (0.00) (0.03) (0.00) (0.00) (0.00) (0.04)

Panel B. Emerging Markets BRL HKD KRW MXN SGD ZAR -0.004 0.006 -0.005 -0.002 0.000 0.002 0.000 0.000 0.000 0.000 0.000 0.000 0.068 0.079 0.070 0.049 0.068 0.068 -1.15 30.50 2.32 35.39 -0.15 59.45 -0.67 27.78 -4.39 110.15 -0.91 36.38 0.03 0.01 0.05 0.06 0.06 0.16 (0.00) (0.02) (0.00) (0.00) (0.00) (0.00)

36

Table II. Order Flow Portfolios: Excess Returns This table shows average annualized portfolio excess returns for currency portfolios sorted on lagged order ow. We standardize order ow over a rolling window of 60 trading days prior to the order ow signal as outlined in the text. Column Av shows average excess returns across all currencies, column BMS (bought minus sold) reports average excess returns for longshort portfolios in currencies with the highest versus lowest order ow. Numbers in brackets are t-statistics based on Newey-West standard errors whereas numbers in parentheses show (annualized) Sharpe Ratios. Columns MR, Up, and Down report p-values for tests of return monotonicity. The frequency is daily and the sample is from January 2001 to May 2011. Panel A shows results for total order ows and all 15 markets (T15) as well as for total order ows and the subsample of nine developed markets (T9). Panel B reports results for order ows disaggregated by customer type (asset managers AM, hedge funds HF, corporate clients CC, private clients PC). Panel A. Total Order Flows P1 T15 P2 P3 P4 P5 Av. BMS MR Up Down

T9

0.82 1.05 6.15 [0.29] [0.37] [2.23] (0.09) (0.11) (0.71) 0.34 2.24 8.21 [0.10] [0.74] [2.60] (0.03) (0.23) (0.80)

6.77 11.13 5.18 [2.40] [4.04] [2.20] (0.77) (1.21) (0.69) 12.76 5.89 [4.17] [2.15] (1.23) (0.66)

10.31 0.00 0.00 [4.05] (1.26) 12.43 0.00 0.00 [4.68] (1.45)

Panel B. Disaggregated Order Flows AM -1.13 [-0.35] (-0.11) -0.32 [-0.10] (-0.03) 6.90 [2.15] (0.67) 12.71 [4.06] (1.23) 3.75 [1.24] (0.38) 6.05 [2.04] (0.61) 5.27 [1.73] (0.53) 6.69 [2.18] (0.67) 6.30 14.31 [2.04] [4.63] (0.62) (1.38) 6.26 9.78 [1.94] [3.02] (0.59) (0.94) 7.02 2.61 [2.16] [0.84] (0.66) (0.26) 2.90 -1.30 [0.93] [-0.41] (0.28) (-0.13) 15.43 [5.72] (1.79) 10.09 [3.94] (1.20) -4.29 [-1.66] (-0.51) -14.01 [-5.20] (-1.55) 0.00 0.00

HF

0.04

0.00

CO

0.35

0.09

PC

0.00

0.00

37

Table III. Order Flow Portfolios: Marginal Forecast Performance for Longer Horizons This table shows average excess returns (p.a.) for BMS portfolios sorted on lagged order ow as in Table II. t-statistics based on Newey-West standard errors are reported in brackets. However, we do not only sort on order ow of the previous day but also allow for longer lags of up to nine days between order ow signals and portfolio formation. Portfolios are rebalanced daily. T15 denotes portfolios sorts on total order ows and the sample of all 15 currencies, and T9 denotes portfolios sorts on total order ows and the sample of 9 developed currencies; AM, HF, CC, and PC denote portfolios sorts on asset managers, hedge funds, corporate clients, and private clients order ows, respectively. Lags between order ow signal and portfolio formation (days) 1 T15 T9 AM HF CC PC 10.31 [4.05] 12.43 [4.68] 2 24.63 [8.94] 24.27 [8.73] 3 4 5 6 7 0.31 [0.13] 2.28 [0.90] 8 9 10

10.22 -1.11 [4.38] [-0.44] 7.44 -4.17 [2.99] [-1.61]

3.02 0.20 [1.28] [0.09] 5.39 -1.55 [2.00] [-0.61]

1.93 -2.32 -0.43 [0.84] [-0.95] [-0.19] 1.33 -1.08 -1.75 [0.51] [-0.42] [-0.71]

15.43 24.86 8.27 -1.29 2.17 0.62 -0.20 3.37 2.26 -2.79 [5.72] [8.80] [3.03] [-0.47] [0.87] [0.23] [-0.07] [1.22] [0.82] [-0.97] 10.09 28.22 2.05 -2.94 0.14 -6.19 2.84 -0.29 -4.66 -1.05 [3.94] [9.26] [0.79] [-1.15] [0.05] [-2.39] [1.12] [-0.10] [-1.77] [-0.40] -4.29 -8.13 -1.47 2.25 -4.98 1.91 -0.01 1.40 -0.33 2.80 [-1.66] [-2.86] [-0.49] [0.88] [-1.93] [0.74] [0.00] [0.56] [-0.12] [1.08] -14.01 -33.77 3.21 1.82 -3.29 -0.77 2.27 -1.35 0.65 2.10 [-5.20] [-10.80] [1.24] [0.67] [-1.15] [-0.27] [0.86] [-0.52] [0.24] [0.78]

38

Table IV. Panel Regressions of Currency Returns on Lagged Order Flow This table reports results for panel regressions of currency excess returns (rxt+1 ) on lagged customer order ow (OFt ) and control variables (the interest rate dierential ij,t it , lagged excess returns over the previous day rxt , lagged excess returns over the prior 60 days rxt1;t60 ). Order ow is measured in billion USD. T15 and T9 refer to total order ow for all 15 currencies and the sample of developed market currencies, respectively. The regressions in (v) and (vi) also include disaggregated order ow for asset managers, AM, hedge funds, HF, corporate clients, CC, and private clients, CC). In each specication, we show results both for pooled regressions (pooling over all currency pairs) and for specications with currency pair- and time-xed eects. t-statistics based on clustered standard errors (by currency pair) are reported in brackets. (i) const. OFtT 15 OFtT 9 OFtAM OFtHF OFtCC OFtP C ij,t it rxt rxt1;t60 Country dummies Time dummies R2 obs 1.036 [7.74] 0.002 [0.27] 0.000 [-0.79] NO NO 1.720 [2.45] -0.010 [-1.21] -0.010 [-3.56] YES YES 0.897 [2.59] 0.001 [0.12] 0.000 [-1.37] NO NO 0.108 [0.11] -0.013 [-1.37] -0.015 [-2.77] YES YES 0.015 [4.92] 0.025 [3.72] (ii) -0.010 [-3.56] 0.023 [3.49] (iii) 0.020 [6.81] (iv) -0.015 [-2.77] (v) 0.020 [6.41] (vi) -0.015 [-2.72]

0.023 [3.42]

0.021 [3.10] 0.043 [4.94] 0.011 [2.12] -0.017 [-2.63] -0.028 [-3.43] 0.936 [2.75] -0.006 [0.57] 0.000 [-0.96] NO NO 0.038 [4.29] 0.010 [1.95] -0.014 [-2.23] -0.024 [-3.00] 0.348 [0.36] -0.018 [-1.86] -0.015 [-2.72] YES YES

0.002 0.024 0.001 0.030 0.006 0.045 37,936 37,936 23,436 23,436 23,436 23,436

39

Table V. Drivers of Customer FX Order Flow: Panel Regressions This table reports results for panel regressions of customer order ows (OF) on lagged customer order ow (OFt for asset managers, AM, hedge funds, HF, corporate clients, CC, and private clients, CC). The regressions also consider lagged returns on various asset classes as additional regressors (the interest rate dierential ij,t it , lagged exchange rate changes over the previous day st and over the prior 20 trading days st1,t20 , lagged country-level eq eq equity returns over the previous trading day rt and over the prior 20 trading days rt 1;t20 ), b and lagged country-level government bond returns rt (10-year maturity benchmark bonds). t-statistics based on clustered standard errors (by currency pair) are reported in brackets and we account for currency pair- and time-xed eects. Dependent variable: Customer order ows C OFtAM OFtHF OFtCC OFtP +1 +1 +1 +1 OFtAM OFtHF OFtCC OFtP C ij,t it st st1,t20
eq rt eq rt 1;t20 b rt b rt 1;t20

0.035 [4.46] 0.034 [2.75] -0.017 [-2.58] -0.026 [-2.10]

const. Country dummies Time dummies R2 obs

0.008 [0.71] YES YES

0.033 [4.22] 0.031 [2.66] -0.016 [-2.53] -0.025 [-2.05] -0.150 [-2.05] 3.541 [4.97] 1.012 [1.97] 1.251 [2.56] 0.347 [1.44] -3.730 [-1.56] -0.019 [-0.03] -0.002 [0.03] YES YES

0.013 [1.79] 0.008 [0.57] 0.000 [0.02] -0.005 [-0.67]

-0.078 [-4.42] YES YES

0.012 [1.73] 0.007 [0.50] 0.000 [0.05] -0.004 [-0.61] 0.102 [0.80] 1.769 [1.47] 0.612 [0.50] 0.399 [0.41] -0.113 [-0.52] -5.170 [-1.26] 0.278 [-0.55] -0.089 [-3.47] YES YES

-0.010 [-1.13] -0.009 [-1.70] 0.035 [2.93] 0.025 [2.47]

-0.320 [-7.27] YES YES

-0.009 [-1.08] -0.008 [-1.55] 0.034 [2.88] 0.024 [2.46] 0.413 [2.51] -1.312 [-1.15] -0.741 [-0.45] -1.164 [-2.37] 0.205 [1.17] -1.135 [-0.57] 0.626 [1.04] -0.295 [-6.01] YES YES

-0.005 [-0.61] -0.037 [-2.59] -0.012 [-1.47] 0.027 [2.21]

-0.003 [-0.38] -0.350 [-2.56] -0.013 [-1.62] 0.025 [2.02] 0.185 [1.02] -4.187 [-2.52] -2.187 [-1.52] -0.226 [-0.34] -0.225 [-0.34] 10.145 [2.68] 1.151 [2.03] 0.039 0.076 [4.77] [4.41] YES YES YES YES

0.013 0.015 0.011 0.011 0.029 0.030 0.015 0.018 23,796 23,796 23,796 23,796 23,796 23,796 23,796 23,796 40

Table VI. Risk Exposures of Financial End-Users

This table reports regression results for the risk exposures of the BMS portfolios of nancial FX market end-users, that is, asset managers and hedge funds. The methodological framework in Panel A is a modied linear Fung-Hsieh model with eight factors as outlined in the main text. Panel B also accounts for conditional equity market exposures by including additional interaction terms. The three conditioning variables are rst dierences of the TED spread, the VIX and the 3-month T-Bill rate. The Table shows results for four parsimonious model specications where the factors are selected according to the Schwarz criterion as outlined in the main text. Results for the other factors are not reported. We further report the estimated intercept , the adjusted R2 and the BIC computed for the two-equation system (Sys-BIC). Below the regression coecients, t-statistics based on HAC standard errors are reported (in brackets).
A. (i) P T F SF X P T F SIR T S DF VOLF X 2 R Sys-BIC B. rm VIX(t-1) Asset Managers (ii) (iii) (iv) 2.35 [2.65] 3.07 2.18 [4.03] [2.86] -2.03 [-2.06] 3.15 3.65 [2.83] [2.87] 0.07 0.06 [2.44] [2.13] 1.46 1.40 1.26 1.23 [5.32] [5.45] [5.68] [5.25] 0.10 0.12 0.21 0.15 3.53 3.53 3.54 3.54 Asset Managers (i) (ii) (iii) (iv) -0.26 [-2.69] -0.19 [-2.82] -0.26 [-2.47] -0.20 -0.20 [-3.06] [-2.48] 2.52 [2.31] Hedge Funds (ii) (iii) -2.68 [-2.51] -1.33 [-1.85] 0.38 [0.59] -3.58 [-2.61] -1.16 [-1.67]

(i)

(iv)

-3.67 [-2.69]

-0.07 -0.05 [-2.50] [-2.09] 0.71 0.78 0.89 0.90 [3.10] [3.49] [4.01] [3.97] 0.11 0.14 0.10 0.10 3.53 3.53 3.54 3.54 (i) Hedge Funds (ii) (iii) (iv)

-0.25 [-2.95] rm TED(t-1) -0.18 [-2.44] P T F SF X P T F SIR VOLF X 2 R Sys-BIC

2.11 [2.79] 0.05 0.06 [2.08] [2.15] 1.35 1.18 1.20 1.44 [5.57] [5.55] [5.38] [5.31] 0.17 0.18 0.15 0.12 3.45 3.46 3.47 3.47

0.15 0.17 0.16 [2.30] [3.08] [3.32] 0.32 0.37 0.33 [4.14] [5.74] [4.58] -2.54 [-2.48] -0.87 [-1.35] -0.05 [-2.10] 0.72 0.85 0.86 [3.50] [3.89] [4.05] 0.21 0.18 0.21 3.45 3.46 3.47

0.34 [4.63]

-0.05 [-2.29] 0.67 [3.21] 0.19 3.47

41

Table VII. Risk Exposures of Private and Corporate Clients

This table reports regression results for the risk exposures of the BMS portfolios computed from the ows of corporate clients (CC) or private clients (PC). The methodological framework in Panel A is a modied linear Fung-Hsieh model with eight factors as outlined in the main text. Panel B also accounts for conditional equity market exposures by including additional interaction terms. The three conditioning variables are rst dierences of the TED spread, the VIX and the 3-month T-Bill rate. The Table shows results for four parsimonious model specications where the factors are selected according to the Schwarz criterion as outlined in the main text. Results for the other factors are not reported. We further report the estimated intercept , the adjusted R2 and the BIC computed for the two-equation system (Sys-BIC). Below the regression coecients, t-statistics based on HAC standard errors are reported (in brackets).
A. (i) rm PTFSIR DF 2 R Sys-BIC B. rm Corporate Clients (ii) (iii) (iv) -0.14 [-1.94] -1.83 [-1.04] Private Clients (ii) (iii) (iv) -0.10 [-2.43] -0.70 [-0.52]

(i)

-0.14 -0.11 [-2.32] [-1.63]

-0.07 -0.09 [-1.56] [-2.27]

-3.43 -2.20 [-1.38] [-0.99] -0.30 -0.31 -0.37 -0.25 [-1.49] [-1.5] [-1.73] [-1.42] 0.07 0.04 0.01 0.07 3.93 3.94 3.94 3.96 Corporate Clients (i) (ii) (iii) (iv)

2.57 3.18 [2.81] [3.51] -1.16 -1.15 -1.19 -1.12 [-4.27] [-4.13] [-4.34] [-4.04] 0.06 0.03 0.04 0.03 3.93 3.94 3.94 3.96 (i) Private Clients (ii) (iii) (iv)

-0.08 -0.10 [-1.30] [-1.50] rm TED(t-1) 0.47 0.44 0.40 0.49 [2.95] [2.29] [2.1] [2.56] PTFSIR -0.94 [-0.8] DF 0.39 [0.25] -0.46 -0.41 -0.37 -0.46 [-2.04] [-1.91] [-1.85] [-2.02] 2 R 0.13 0.14 0.15 0.12 Sys-BIC 3.82 3.83 3.87 3.88

-0.13 [-3.05] -0.31 -0.12 [-2.68] [-0.99] -1.38 [-1.45] 2.51 [1.71] -1.18 -1.12 -1.06 -1.19 [-4.25] [-4.27] [-4.20] [-4.46] 0.02 0.07 0.09 0.04 3.82 3.83 3.87 3.88

-0.11 [-2.48] -0.20 -0.24 [-2.53] [-2.27]

42

Figure 1. Cumulative Excess Returns on BMS Portfolios

All countries 140 120 Cumulative log return (in %) 100 80 60 40 20 0 -20 01 02 03 04 05 06 07 08 09 10 11 Cumulative log return (in %) 140 120 100 80 60 40 20 0 -20 01 02 03 04 05 06 07 08 09 10 11 Developed countries

Asset Managers 180 160 Cumulative log return (in %) 140 120 100 80 60 40 20 0 01 02 03 04 05 06 07 08 09 10 11 Cumulative log return (in %) 120 100 80 60 40 20 0

Hedge Funds

-20 01 02 03 04 05 06 07 08 09 10 11

Corporate Clients 10 Cumulative log return (in %) Cumulative log return (in %) 0 -10 -20 -30 -40 -50 -60 01 02 03 04 05 06 07 08 09 10 11 50

Private Clients

-50

-100

-150 01 02 03 04 05 06 07 08 09 10 11

This gure shows cumulative log excess returns for a long-short portfolio based on total order ows and all countries (T15), total ows and developed markets (T9), asset manager ows (AM), hedge fund ows (HF), corporate clients ows (CC), and private clients ows (PC). The sample period is daily from January 2001 May 2011.

43

Figure 2. Cumulative Post-Formation Exchange Rate Changes

This gure shows average cumulative spot exchange rate changes for BMS portfolios based on total ows and disaggregated ows over the rst 30 days after portfolio formation. We use daily data so that post-formation periods overlap. Shaded areas correspond to a 95% condence interval obtained from a moving-block bootstrap with 1,000 repetitions. 44

Figure 3. BMS Excess Returns in Event Time

This gure shows BMS portfolio excess returns (solid lines) in event time, from ve days prior to portfolio formation (t = 5, the day of portfolio formation (t = 0), up to ten days after portfolio formation (t = 10). BMS excess returns are annualized and in percent. The shaded areas correspond to 95% condence intervals based on Newey/West standard errors. The frequency is daily and the sample is from January 2001 May 2011.

45

Figure 4. Correlation of Customer Order Flows Over Longer Horizons

This gure shows average correlation coecients between customer order ows (left panel) for horizons of 1, 2, ..., 60 trading days. Average correlations between ows are based on the average correlation across all nine currency pairs. A horizon of one day corresponds to (non-overlapping) daily observations, whereas correlations for longer horizons are based on (overlapping) sums of daily observations. Shaded areas correspond to bootstrapped 95% condence intervals based on a moving-block bootstrap with 1,000 repetitions. The sample period is January 2001 May 2011.

46

Figure 5. Rebalancing Frequency and Net Excess Returns

This gure shows average annualized excess returns for BMS portfolios based on total and disaggregated order ows for dierent rebalancing frequencies ranging from one to ten days. The dotted lines show excess returns and a 95% condence interval based on Newey-West standard errors before transaction costs whereas the solid line and shaded area show net excess returns and a 95% condence interval based on Newey-West standard errors after transaction costs.

47

Internet Appendix to accompany Information Flows in Dark Markets: Dissecting Customer Currency Trades
(Not for Publication)

48

Table A.1. Descriptive Statistics for Disaggregated Customer Order Flows This table shows descriptive statistics for customer order ows which are available for the nine major markets in our sample, that is, the Australian Dollar (AUD), Canadian Dollar (CAD), Swiss Franc (CHF), Euro (EUR), Great Britain Pound (GBP), Japanese Yen (JPY), Norwegian Krone (NOK), New Zealand Dollar (NZD), Swedish Krona (SEK). Flows are measured in billions (in USD) and all currencies are against the USD. A positive (negative) ow means that there is net buying (selling) pressure in the foreign currency against the USD. We report means, medians, standard deviations, skewness, kurtosis, and rst-order autocorrelation coecients (AC(1)) for all four customer groups ows and, for comparison, for total order ow in the nine currencies (T9). The rst number in each cell corresponds to the cross-sectional mean across currencies (e.g., the mean across time-series standard deviations of all nine currencies), whereas the second (parentheses) and third (brackets) number correspond to the 5% and 95% percentile of the cross-sectional distribution (across currencies), respectively. The frequency is daily and the sample is from January 2001 to May 2011. Mean Median Std Skew Kurt AC(1)

Panel A. Asset Managers -0.001 -0.001 0.272 (-0.063) (-0.041) (0.067) [0.027] [0.019] [0.656] -0.827 (-5.820) [1.978] 80.0 0.034 (39.8) (-0.009) [270.7] [0.140]

Panel B. Hedge Funds 0.002 0.001 0.205 (-0.004) (-0.002) (0.054) [0.009] [0.005] [0.494] -0.738 (-7.977) [4.810] 125.1 0.032 (17.5) (-0.117) [271.0] [0.128]

Panel C. Corporate Clients -0.003 -0.001 0.171 -1.091 (-0.028) (-0.022) (0.036) (-23.273) [0.012] [0.010] [0.387] [12.143] Panel D. Private Clients -0.003 -0.003 0.068 -0.137 (-0.049) (-0.038) (0.009) (-17.616) [0.007] [0.006] [0.165] [10.063] 208.8 0.072 (22.2) (-0.025) [638.8] [0.192] 176.8 0.004 (11.0) (-0.107) [898.1] [0.091]

Panel E. Total Flows (T9) 0.003 (-0.001) [0.014] 0.002 0.091 -2.857 225.0 0.024 (0.000) (0.009) (-30.643) (16.2) (-0.106) [0.012] [0.265] [5.212] [1,385.8] [0.075] 49

Table A.2. Correlation Between Customer Groups Order Flows This table reports correlation coecients between ows of customer groups for nine major currencies and for a pooled sample over all currencies. Correlation coecients AM/HF AM/CC AM/PC HF/CC HF/PC EUR JPY GBP CHF AUD NZD CAD SEK NOK Pooled -0.04 0.05 -0.03 0.01 0.00 0.00 -0.04 -0.03 -0.02 -0.01 -0.05 -0.05 0.02 -0.09 0.03 -0.05 -0.08 -0.01 -0.04 -0.04 -0.05 -0.12 -0.11 -0.08 -0.02 -0.06 -0.05 -0.01 -0.01 -0.05 -0.10 -0.02 -0.02 -0.07 -0.06 0.01 -0.01 -0.02 -0.03 -0.03 -0.20 -0.20 -0.17 -0.20 -0.07 -0.03 -0.15 0.04 0.03 -0.10

CC/PC -0.05 -0.05 0.02 -0.09 0.03 -0.05 -0.08 -0.01 -0.04 -0.04

50

Table A.3. Order Flow Portfolios: Dierent Standardization Schemes and Sub-Samples The setup of this table is identical to Table II, Panel A, in the main text but shows results for rolling (Panel A), recursive (Panel B) ,and in-sample standardization (Panel C) of customer order ow and for three dierent sample periods as opposed to the rolling standardization scheme employed in Table II. Panel A. Rolling Window P1 2001/01 2011/05 P2 P3 P4 P5 Av. BMS MR Up

0.82 1.05 6.15 6.77 11.13 5.18 10.31 0.01 0.01 [0.29] [0.37] [2.23] [2.40] [4.04] [2.20] [4.05] 2001/01 2007/06 2.14 4.21 5.06 6.02 11.84 5.85 9.69 0.00 0.04 [0.71] [1.41] [1.79] [2.23] [4.14] [2.55] [3.45] 2007/07 2011/06 -1.18 -3.70 7.79 7.90 10.07 4.18 11.25 0.18 0.05 [-0.21] [-0.67] [1.44] [1.37] [1.87] [0.87] [2.36]

Panel B. Recursive Window P1 2001/01 2011/05 P2 P3 P4 P5 BMS MR Up

-0.42 2.35 5.68 6.73 11.74 [-0.14] [0.83] [2.13] [2.40] [4.19] 2001/01 2007/06 0.56 5.82 3.86 7.62 11.68 [0.18] [2.07] [1.39] [2.83] [3.91] 2007/07 2011/06 -1.89 -2.87 8.41 5.4 11.83 [-0.34] [-0.51] [1.62] [0.94] [2.20] Panel C. In-Sample P1 2001/01 2011/05 P2 P3 P4 P5

12.16 0.00 0.00 [4.97] 11.12 0.19 0.00 [4.00] 13.72 0.02 0.01 [3.07]

BMS

MR

Up

0 1.91 7.16 6.09 10.98 [0.00] [0.68] [2.58] [2.14] [4.00] 2001/01 2007/06 1.86 4.47 6.54 6.4 10.36 [0.63] [1.52] [2.18] [2.31] [3.65] 2007/07 2011/06 -2.79 -1.92 8.09 5.61 11.91 [-0.51] [-0.35] [1.53] [0.97] [2.21]

10.98 0.11 0.00 [4.65] 8.5 0.01 0.07 [3.26] 14.7 0.15 0.01 [3.34]

51

Table A.4. Order Flow Portfolios: Exchange Rate Changes This table shows average portfolio exchange rate changes for ve portfolios (P1 , ..., P5 ) sorted on lagged order ow. Sorting is done based on standardized total ows of all customers. Column Av shows average excess returns across all currencies, column BMS (bought minus sold) reports average excess returns to investing in P5 and shorting P1 . Panel B reports the same information for spot exchange rate changes instead of excess returns. Flows are standardized by their standard deviation (i) using a rolling window over the previous 60 trading days (Panel A), (ii) using a recursive scheme with 60 days initialization horizon (Panel B), and (iii) their in-sample standard deviation. Average spot rate changes are annualized (assuming 252 trading days per year). Numbers in brackets are t-statistics based on NeweyWest standard errors. The frequency is daily and the sample is from January 2001 May 2011. Panel A. Rolling Window P1 Jan 2001 May 2011 Jan 2001 Jun 2007 Jul 2007 May 2011 P2 P3 P4 P5 Av. BMS

-1.28 -0.64 4.01 4.13 10.20 3.28 11.48 [-0.45] [-0.22] [1.47] [1.41] [3.72] [1.40] [4.57] -0.24 2.56 2.70 2.73 11.35 3.82 11.59 [-0.08] [0.86] [0.98] [0.91] [4.02] [1.68] [4.25] -2.85 -5.45 5.99 6.23 8.46 2.48 11.31 [-0.52] [-0.98] [1.11] [1.08] [1.57] [0.52] [2.37] Panel B. Recursive Window P1 P2 P3 P4 P5 BMS 12.73 [5.17] 12.03 [4.28] 13.77 [3.08]

Jan 2001 May 2011 Jan 2001 Jun 2007 Jul 2007 May 2011

-2.40 0.71 3.40 4.32 10.33 [-0.83] [0.25] [1.27] [1.50] [3.70] -1.42 4.17 1.02 4.62 10.61 [-0.46] [1.46] [0.36] [1.60] [3.58] -3.87 -4.49 6.97 3.87 9.90 [-0.69] [-0.80] [1.34] [0.67] [1.84] Panel C. In-Sample P1 P2 P3 P4 P5

BMS 10.85 [4.57] 8.27 [3.14] 14.74 [3.34]

Jan 2001 May 2011 Jan 2001 Jun 2007 Jul 2007 May 2011

-1.17 -0.98 4.63 4.26 9.68 [-0.41] [-0.34] [1.68] [1.49] [3.51] 1.16 0.78 3.34 4.44 9.42 [0.40] [0.26] [1.13] [1.55] [3.30] -4.67 -3.61 6.58 4.01 10.07 [-0.84] [-0.65] [1.25] [0.70] [1.87]

52

Table A.5. Order Flow Portfolios: Customer Groups and Exchange Rate Changes This table is similar to Panel B of Table II but here we report results for spot exchange rate changes (and not excess returns). P1 AM HF CO PC T9 -1.65 [-0.51] -0.90 [-0.29] 6.30 [1.97] 12.08 [3.85] P2 2.97 [0.98] 5.32 [1.80] 4.47 [1.47] 5.99 [1.96] P3 P4 Av. BMS 15.52 [5.75] 10.15 [3.96] -4.04 [-1.56] -13.91 [-5.16] 12.55 [4.72]

5.62 13.86 [1.81] [4.49] 5.70 9.25 [1.77] [2.85] 6.37 2.26 [1.96] [0.73] 2.28 -1.84 [0.73] [-0.57] 12.24 5.27 [4.00] [1.93]

-0.31 1.54 7.58 [-0.09] [0.51] [2.40]

53

Table A.6. Correlation of Excess Returns This table reports correlation coecients between excess returns of dierent BMS portfolios based on (i) lagged total ows of all 15 currency pairs (T15), (ii) lagged total ows of nine developed countries (T9), (iii) lagged ows of asset managers (AM), (iv) lagged ows of hedge funds, (v) lagged ows of corporate clients (CC), and lagged ows of private clients (PC). All ows are standardized by their lagged volatility over a 60-day rolling window. The frequency is daily and the sample period is January 2001 to Ma 2011. T15 T15 T9 AM HF CC PC T9 AM HF CC

1.00 0.63 1.00 0.27 0.42 1.00 0.30 0.42 0.06 1.00 0.00 -0.06 -0.08 -0.13 1.00 -0.04 -0.02 -0.07 0.03 0.01

54

Table A.7. Order Flow Portfolios: Standardizing Flows (One Year) This table shows average annualized portfolio excess returns for ve (or four) portfolios (P1 , ..., P5 ) sorted on lagged order ow. Sorting is done based on standardized total ows and standardized customer ows. Column Av shows average excess returns across all currencies, column BMS (buying minus selling pressure) reports average excess returns to investing in P5 (or P4 ) and shorting P1 . Flows are standardized by their standard deviation using a rolling window over the previous 252 trading days (that is, roughly one year). We form 5 portfolios for total ows of all 15 currency pairs (T15) and 4 portfolios for total ows of the nine currencies for which we have disaggregated ows available (T9), for asset managers ows (AM), hedge funds ows (HF), corporate clients ows (CC), and private clients ows (PC). Numbers in brackets are t-statistics based on Newey-West standard errors. The frequency is daily and the sample is from January 2001 to May 2011.
Panel A. Excess Returns P1 T15 T9 AM HF CC PC 0.71 [0.23] 0.66 [0.19] -1.15 [-0.33] -0.63 [-0.19] 7.81 [2.36] 16.30 [4.96] P2 2.61 [0.87] 2.56 [0.82] 3.79 [1.19] 7.35 [2.32] 6.09 [1.86] 5.55 [1.65] P3 7.91 [2.77] 8.44 [2.54] 6.66 [2.02] 5.62 [1.65] 8.55 [2.48] 3.36 [0.99] P4 5.95 [2.00] 13.81 [4.24] 16.10 [4.98] 10.75 [3.13] 1.10 [0.34] -1.48 [-0.45] P5 12.60 [4.29] Av. 5.96 [2.38] 6.37 [2.19] BMS 11.89 [4.59] 13.15 [4.81] 17.25 [6.23] 11.38 [4.23] -6.71 [-2.39] -17.79 [-6.49]

Panel B. Exchange Rate Changes P1 T15 T9 AM HF CC PC -1.42 [-0.47] -0.05 [-0.01] -1.68 [-0.49] -1.27 [-0.39] 7.18 [2.16] 15.67 [4.77] P2 0.66 [0.22] 1.93 [0.62] 3.03 [0.96] 6.68 [2.11] 5.31 [1.62] 4.87 [1.45] P3 6.26 [2.20] 7.86 [2.36] 5.98 [1.82] 5.10 [1.49] 7.92 [2.30] 2.77 [0.82] P4 3.18 [1.04] 13.23 [4.06] 15.65 [4.84] 10.13 [2.95] 0.76 [0.23] -2.05 [-0.63] P5 11.20 [3.80] Av. 3.97 [1.59] 5.74 [1.98] BMS 12.62 [4.79] 13.27 [4.85] 17.33 [6.25] 11.41 [4.24] -6.42 [-2.28] -17.72 [-6.47]

55

Table A.8. Order Flow Portfolios: Standardizing Flows (Three Years) This table shows average annualized portfolio excess returns for ve (or four) portfolios (P1 , ..., P5 ) sorted on lagged order ow. Sorting is done based on standardized total ows and standardized customer ows. Column Av shows average excess returns across all currencies, column BMS (buying minus selling pressure) reports average excess returns to investing in P5 (or P4 ) and shorting P1 . Flows are standardized by their standard deviation using a rolling window over the previous 750 trading days (that is, roughly three years). We form 5 portfolios for total ows of all 15 currency pairs (T15) and 4 portfolios for total ows of the nine currencies for which we have disaggregated ows available (T9), for asset managers ows (AM), hedge funds ows (HF), corporate clients ows (CC), and private clients ows (PC). Numbers in brackets are t-statistics based on Newey-West standard errors. The frequency is daily and the sample is from January 2001 to May 2011.
Panel A. Excess Returns P1 T15 T9 AM HF CC PC -2.21 [-0.61] -3.02 [-0.75] -4.89 [-1.22] -2.44 [-0.64] 6.48 [1.65] 13.26 [3.47] P2 -0.55 [-0.16] -1.86 [-0.50] 1.15 [0.30] 2.31 [0.61] 2.16 [0.55] 4.32 [1.09] P3 5.97 [1.75] 7.29 [1.91] 3.20 [0.84] 3.97 [0.99] 7.25 [1.78] 0.98 [0.24] P4 6.28 [1.74] 13.28 [3.40] 15.39 [3.99] 9.92 [2.42] -2.37 [-0.63] -6.05 [-1.54] P5 10.25 [3.04] Av. 3.95 [1.31] 3.92 [1.14] BMS 12.46 [4.30] 16.30 [5.01] 20.28 [6.30] 12.36 [3.85] -8.85 [-2.70] -19.31 [-5.78]

Panel B. Exchange Rate Changes P1 T15 T9 AM HF CC PC -3.89 [-1.08] -3.51 [-0.86] -5.16 [-1.29] -2.93 [-0.77] 6.15 [1.57] 12.76 [3.34] P2 -2.04 [-0.58] -2.09 [-0.56] 0.75 [0.20] 2.06 [0.54] 1.71 [0.44] 4.06 [1.02] P3 5.17 [1.52] 7.05 [1.84] 2.88 [0.76] 3.82 [0.95] 6.93 [1.70] 0.79 [0.19] P4 4.95 [1.37] 12.92 [3.31] 15.14 [3.93] 9.50 [2.32] -2.47 [-0.66] -6.40 [-1.63] P5 8.60 [2.55] Av. 2.56 [0.85] 3.59 [1.05] BMS 12.49 [4.28] 16.43 [5.04] 20.30 [6.31] 12.43 [3.87] -8.62 [-2.62] -19.16 [-5.73]

56

Table A.9. Order Flow Portfolios: Order Flows Scaled By Currency Trading Volume This table is identical to Table II but here we do not standardize order ows by rolling windows of the previous 60 trading days volatility but by total currency trading volume (from the BIS FX triennial surveys for 2001, 2004, 2007, 2010). We linearly interpolate between the turnover gures to obtain a daily measure of total trading volume for each of the fteen currencies in our sample. Panel A. Total order ows P1 T15 P2 P3 P4 P5 Av. BMS MR Up Down

T9

-1.14 3.18 3.19 [-0.38] [1.19] [1.26] (-0.12) (0.37) (0.39) -0.73 2.35 6.09 [-0.22] [0.83] [1.98] (-0.07) (0.25) (0.61)

6.73 11.23 5.01 [2.37] [3.99] [2.00] (0.76) (1.19) (0.66) 12.46 5.45 [4.01] [2.11] (1.17) (0.63)

12.37 0.02 0.00 [5.05] (1.53) 13.19 0.00 0.00 [5.09] (1.55)

Panel B. Disaggregated order ows AM -0.97 [-0.29] (-0.09) 0.14 [0.04] (0.01) 5.94 [1.88] (0.58) 14.06 [4.52] (1.35) 2.40 [0.85] (0.25) 4.05 [1.40] (0.42) 5.16 [1.72] (0.52) 5.01 [1.67] (0.51) 5.98 13.23 [2.02] [4.19] (0.61) (1.24) 5.94 9.43 [1.95] [2.87] (0.59) (0.87) 4.25 3.07 [1.34] [1.01] (0.41) (0.31) 0.69 -1.14 [0.22] [-0.36] (0.07) (-0.11) 14.19 [5.33] (1.66) 9.29 [3.50] (1.10) -2.88 [-1.15] (-0.35) -15.20 [-5.68] (-1.71) 0.00 0.00

HF

0.00

0.04

CO

0.01

0.65

PC

0.00

0.00

57

Table A.10. Order Flow Portfolios: Demeaning Flows This table is similar to Table II but here we present results for total ows (T15 and T9) and customer groups ows and we standardize order ows by subtracting the rolling mean and dividing by the rolling standard deviation. The frequency is daily and the sample is from January 2001 May 2011. P1 T15 T9 AM HF CO PC P2 P3 P4 P5 Av BMS 11.84 [4.89] 13.49 [5.25] 14.77 [5.44] 9.99 [3.88] -4.17 [-1.50] -14.19 [-5.00]

-0.68 5.02 5.54 [-0.24] [1.86] [1.95] 0.11 3.21 6.12 [0.03] [1.08] [1.91] -0.28 [-0.08] -0.04 [-0.01] 7.91 [2.48] 12.11 [3.77] 4.64 [1.56] 6.29 [2.09] 5.41 [1.82] 6.37 [2.15]

5.67 11.16 5.34 [2.07] [3.88] [2.26] 13.60 [4.40]

3.45 14.50 [1.07] [4.77] 5.26 9.95 [1.69] [3.02] 4.85 3.75 [1.54] [1.17] 5.05 -2.09 [1.60] [-0.63]

58

Table A.11. Order Flow-Weighted BMS Portfolios This table shows returns for BMS portfolios based on total ows and customer ows but here we employ portfolio weights directly based on lagged order ows. For each trading day t, we cross-sectionally standardize order ows, rescale these standardized ows so that they sum to two in absolute value, and then use these rescaled and standardized ows as portfolio weights for day t to t + 1. Numbers in squared brackets are based on Newey/West standard errors. The frequency is daily and the sample is from January 2001 May 2011. T15 T9 AM HF CO PC

Mean 12.70 11.12 13.94 6.61 -1.94 -9.77 t-Stat. 5.07 4.70 5.86 2.91 -0.89 -4.27 St. Dev. 8.38 7.60 7.72 7.46 7.04 7.12 Sharpe Ratio 1.52 1.46 1.81 0.89 -0.28 -1.37 Skewness 0.25 1.42 1.76 0.09 -0.12 -0.20 Kurtosis 19.63 33.82 28.54 11.75 12.87 11.05 Maximum 6.57 7.79 7.42 3.95 2.96 3.21 Minimum -5.33 -3.68 -2.46 -3.69 -3.95 -4.12

59

Table A.12. BMS Portfolios: Longer Horizons This table shows average annualized BMS portfolio excess returns for longer forecast horizons of 1, 2, ..., 5, 10, 20, 40, and 60 trading days. We use an exponential moving average (EMA) with a decay parameter of 0.25 (Panel A) and 0.75 (Panel B) for lagged order ows to consider longer histories of order ows for forecasting. BMS portfolios are based on 5 portfolios for total ows of all 15 currency pairs (T15) and 4 portfolios for total ows of the nine developed markets for which disaggregated ows are available (T9), for asset managers ows (AM), hedge funds ows (HF), corporate clients ows (CC), and private clients ows (PC). Numbers in brackets are t-statistics based on Newey-West standard errors. The frequency is daily and the sample is from January 2001 to May 2011. Rebalancing Frequency (Trading Days) 3 4 5 10 20 40 EMA with Decay Parameter 0.25 T15 T9 AM HF CC PC 17.20 12.31 9.82 7.68 4.57 4.15 2.83 3.60 [8.22] [5.68] [4.76] [3.66] [2.46] [2.40] [1.92] [1.68] 20.74 14.25 11.27 7.98 5.19 4.84 2.14 -0.41 [7.99] [5.32] [4.46] [3.11] [2.99] [2.13] [1.76] [-0.15] 20.28 16.30 11.99 6.63 5.72 5.12 2.99 1.56 [7.39] [5.91] [4.75] [2.61] [2.70] [2.08] [1.42] [0.59] 15.78 7.60 3.78 2.31 -1.03 -0.72 1.50 3.14 [5.30] [2.67] [1.30] [0.81] [-0.36] [-0.26] [0.58] [1.09] -4.74 -3.53 -2.85 -2.61 -0.61 -1.03 0.13 -3.51 [-1.84] [-1.32] [-1.12] [-1.03] [-0.25] [-0.42] [0.05] [-1.53] -16.65 -8.58 -6.44 -2.42 -2.66 0.23 -1.78 -1.77 [-5.73] [-3.33] [-2.31] [-0.87] [-0.99] [0.08] [-0.58] [-0.71] EMA with Decay Parameter 0.75 T15 T9 AM HF CC PC 18.97 16.21 9.48 10.67 5.78 4.27 0.63 3.09 [7.92] [7.46] [4.40] [5.36] [2.94] [1.94] [0.32] [1.58] 23.05 18.09 12.03 10.14 7.83 4.90 1.14 3.87 [8.00] [6.76] [4.62] [4.12] [3.28] [2.22] [0.46] [1.54] 23.75 20.92 16.19 11.74 7.27 5.98 2.14 0.82 [8.30] [7.68] [5.89] [4.32] [3.61] [2.54] [0.87] [0.31] 20.66 11.34 6.83 8.15 6.84 1.46 1.37 3.71 [7.58] [3.73] [2.51] [2.91] [2.67] [0.55] [0.56] [1.41] -3.61 -3.47 -1.54 -4.52 -1.01 -2.20 -2.67 -1.76 [-1.31] [-1.29] [-0.62] [-1.70] [-0.42] [-0.96] [-1.10] [-0.76] -24.34 -14.51 -9.43 -7.12 -3.94 1.10 -0.78 0.72 [-7.96] [-5.25] [-3.51] [-2.61] [-1.62] [0.45] [-0.29] [0.29]

60

60

Table A.13. Order Flow Portfolios: Four Liquid Currencies This table is similar to Table II in the main text but here we only include EUR/USD, JPY/USD, GBP/USD, and CHF/USD in our sample of currencies and only form two portfolios. T4 denotes portfolios based on total order ows. Excess Returns P1 T4 P2 Av. BMS Exchange Rate Changes P1 1.95 [0.70] -0.79 [-0.28] 2.18 [0.79] 5.61 [2.00] 8.14 [2.95] P2 Av. BMS

1.55 5.48 3.51 3.93 [0.56] [2.00] [1.40] [1.70] AM -1.31 8.34 9.65 [-0.46] [3.11] [4.23] HF 1.68 5.35 3.67 [0.61] [1.93] [1.55] CC 5.16 1.90 -3.25 [1.83] [0.70] [-1.41] PC 7.59 -0.56 -8.15 [2.75] [-0.20] [-3.47]

6.20 4.07 4.25 [2.24] [1.62] [1.81] 8.94 9.73 [3.34] [4.27] 5.97 3.79 [2.15] [1.60] 2.57 -3.05 [0.95] [-1.32] 0.00 -8.14 [0.00] [-3.47]

61

Table A.14. Order Flow Portfolios: Sensitivity to Individual Currencies This table show average annualized excess returns to BMS portfolios based on total ows, asset managers (AM) ows, hedge fund (HF) ows, corporate clients (CC) ows, and private clients (PC) ows for a cross-validation setting in which we discard one of the available currencies in our sample. We do this for each available currency and the rst column indicates which currency is left out when computing returns to the BMS portfolio. Hence, BMS returns for total ows are based on 14 currencies and BMS returns for customer ows are based on 8 currencies instead of 15 and 9 currencies, respectively. t-statistics in brackets are based on Newey/West standard errors. Total ows rx t EUR JPY GBP CHF AUD NZD CAD SEK NOK MXN BRL ZAR KRW SGD HKD 10.79 8.26 10.18 10.29 10.39 9.35 11.98 9.33 10.04 10.10 11.22 6.71 11.75 10.84 11.68 [4.12] [3.34] [3.99] [4.09] [4.18] [3.67] [4.70] [3.70] [3.91] [4.06] [4.56] [2.79] [4.78] [4.27] [4.75] AM rx 15.40 9.62 15.29 13.58 10.39 14.75 14.70 15.51 16.52 HF rx CC t rx t PC rx

[5.35] 8.86 [3.10] -2.15 [-0.78] -14.62 [-5.23] [3.95] 11.30 [4.73] -3.13 [-1.33] -9.12 [-3.87] [5.65] 8.95 [3.35] -3.04 [-1.17] -13.65 [-4.93] [5.03] 11.49 [4.51] -2.08 [-0.79] -16.31 [-5.90] [4.03] 9.78 [3.83] -5.58 [-2.19] -10.77 [-4.26] [5.52] 8.16 [3.15] -2.22 [-0.87] -11.93 [-4.45] [5.34] 9.65 [3.81] -4.07 [-1.61] -10.98 [-4.12] [5.76] 6.32 [2.28] -4.46 [-1.76] -14.91 [-5.45] [6.00] 9.38 [3.72] -3.37 [-1.24] -16.20 [-5.84]

62

Table A.15. Order ow portfolio returns around customer trading

This table shows average annualized portfolio excess returns for currency portfolios sorted on lagged disaggregated customer order ow in event time around portfolio formation (t = 5, 4, , 0, 1, 2, , 10 days where t = 0 is the day of portfolio formation) for the high (P4 ) and low (P1 ) portfolios.

5 0 4 5 6

Event time (day) 1 2 3

10

AM

P4

P1

2.41 [0.74] 7.31 [2.35]

5.68 [1.78] 4.47 [1.47]

6.90 [2.02] 3.61 [1.12]

9.61 [3.10] 3.99 [1.28]

11.37 [3.62] 1.43 [0.48]

13.60 [4.27] -1.20 [-0.37]

14.31 [4.65] -1.13 [-0.34]

18.67 [5.90] -6.18 [-1.89]

7.75 [2.50] -0.52 [-0.16]

5.12 [1.51] 6.41 [2.12]

6.62 [2.15] 4.44 [1.47]

4.52 [1.43] 3.90 [1.20]

5.65 [1.74] 5.86 [1.84]

9.20 [2.95] 5.82 [1.81]

7.25 [2.23] 4.99 [1.56]

3.22 [1.00] 6.01 [1.88]

HF

P4

63
2.06 [0.68] 7.27 [2.36] 6.90 [2.19] 4.53 [1.47] 1.51 [0.49] 7.27 [2.24] 0.49 [0.16] 8.52 [2.49] 2.61 [0.84] 6.90 [2.15] 2.80 [0.91] 10.93 [3.41] 4.38 [1.35] 5.84 [1.91] 2.75 [0.87] 8.77 [2.70] 1.76 [0.57] 9.19 [2.86] -3.68 [-1.18] 13.85 [4.37] -9.93 [-3.00] 18.24 [5.61] -1.30 [-0.40] 12.71 [4.05] -10.66 [-2.99] 23.11 [7.31] 6.89 [2.14] 3.67 [1.20]

P1

0.74 [10.10] 5.30 [1.74]

1.78 [7.40] 3.03 [0.97]

2.02 [5.48] 6.27 [2.07]

3.10 [7.27] -0.74 [-0.24]

3.62 [8.60] 0.20 [0.06]

4.27 [17.90] -9.11 [-2.77]

4.65 [9.78] -0.32 [-0.10]

5.90 [19.67] -8.55 [-2.47]

2.50 [6.67] 4.61 [1.52]

1.51 [4.15] 7.08 [2.33]

2.15 [4.53] 4.39 [1.50]

1.43 [1.92] 8.11 [2.62]

1.74 [7.04] 4.20 [1.41]

2.95 [5.60] 5.89 [1.93]

2.23 [2.96] 7.61 [2.51]

1.00 [4.90] 5.95 [1.96]

CC

P4

P1

4.87 [1.62] 5.67 [1.83]

3.73 [1.24] 7.13 [2.27]

5.89 [1.94] 3.64 [1.15]

3.40 [1.03] 8.38 [2.77]

8.09 [2.68] 6.18 [2.01]

5.60 [1.83] 5.61 [1.78]

6.57 [2.12] 5.16 [1.68]

5.36 [1.69] 5.69 [1.81]

7.65 [2.52] 4.86 [1.58]

PC

P4

P1

2.93 [0.97] 7.31 [2.22]

3.16 [1.02] 8.11 [2.45]

5.82 [1.92] 4.00 [1.18]

4.34 [1.33] 7.63 [2.36]

6.77 [2.06] 7.54 [2.35]

5.49 [1.75] 3.22 [0.98]

5.22 [1.62] 6.57 [2.12]

4.94 [1.55] 4.29 [1.36]

6.31 [2.05] 4.21 [1.26]

Table A.16. Pricing Error Statistics For The Cross-Section

This table reports pricing error statistics based on estimating the models of Table VI for the broader cross-section of the order ow mimicking portfolios of nancial end-users. GRS is the test statistic by Gibbons, Ross, and Shanken (1989). We compute the joint test for zero alphas for the entire cross section of eight portfolios of Asset Managers and Hedge Funds. We also compute the GRSstatistic separately for the portfolios (P1-P4) of each group. Model specications (i) - to (iv) follow the setup of table VI. GRS p-val. GRS AM p-val. GRS HF p-val.

A. Linear Exposures (i) 6.87 0.00 10.37 (ii) 6.49 0.00 9.51 (iii) 7.61 0.00 10.77 (iv) 7.14 0.00 9.51 B. Conditional Exposures (i) 6.19 0.00 9.46 (ii) 6.51 0.00 9.04 (iii) 6.53 0.00 9.34 (iv) 6.58 0.00 10.24

0.00 0.00 0.00 0.00 0.00 0.00 0.00 0.00

3.79 4.00 6.69 6.75 4.14 6.31 5.81 3.79

0.01 0.00 0.00 0.00 0.00 0.00 0.00 0.01

64

Table A.17. Risk Exposures: Equity Factors

This table reports regression results for the risk exposures of the BMS portfolios of nancial FX market end-users, that is, asset managers and hedge funds. The risk factors include the excess return on the market portfolio (rm ), and the Fama-French size (SMB) and value (HML) factors. UMD denotes the return on Carharts momentum factor. The Table shows results for four parsimonious model specications (i-iv) where the factors are selected according to the Schwarz criterion (joint estimation of the equation for asset managers and hedge funds BMS returns). Specication (v) includes all factors jointly. We further report the estimated intercept , the adjusted R2 and the BIC computed for the two-equation system (Sys-BIC). Below the regression coecients, t-statistics based on HAC standard errors are reported (in brackets). Asset Managers (ii) (iii) (iv) -0.02 [-0.28] 0.05 [-0.66] -0.05 [-0.47] 0.01 [0.16] 1.28 [4.77] -0.01 3.75 Hedge Funds (ii) (iii) (iv) 0.03 [0.61] 0.06 [0.46] 0.03 [0.46] 0.00 [-0.07] 0.87 [3.77] -0.01 3.75

(i) rm SMB HML UMD

(v) -0.03 [-0.38] 0.07 [1.01] -0.04 [-0.47] 0.00 [-0.13] 1.28 [4.72] -0.03 3.97

(i)

(v) 0.03 [0.52] 0.05 [0.39] 0.02 [0.3] 0.02 [0.37] 0.83 [3.45] -0.03 3.97

1.26 1.29 1.30 [4.57] [4.85] [4.74] 2 R -0.01 -0.01 -0.01 Sys-BIC 3.74 3.74 3.74

0.84 [3.38] 0.00 3.74

0.86 [3.69] 0.00 3.74

0.86 [3.7] -0.01 3.74

65

Table A.18. Risk Exposures: FX Factors

This table reports regression results for the risk exposures of the BMS portfolios of nancial FX market end-users, that is, asset managers and hedge funds. The FX factors include the excess Dollar risk factor and the carry risk factor by Lustig, Roussanov, and Verdelhan (2011). VOLF X is the global FX volatility risk factor (factor mimicking portfolio). The Table shows results for four parsimonious model specications (i-iv) where the factors are selected according to the Schwarz criterion (joint estimation of the equation for asset managers and hedge funds BMS returns). Specication (v) includes all factors jointly. We further report the estimated intercept , the adjusted R2 and the BIC computed for the two-equation system (Sys-BIC). Below the regression coecients, t-statistics based on HAC standard errors are reported (in brackets). Asset Managers (ii) (iii) (iv) Hedge Funds (iii) (iv)

(i) DOL HMLF X 0.07 [2.44] 1.46 [5.32] 2 R 0.10 Sys-BIC 3.53 VOLF X

(v)

(i)

(ii)

(v)

-0.21 -0.04 [-1.75] [-0.33] 0.07 [2.34] 1.43 1.47 [4.75] [5.15] 0.06 0.09 3.55 3.57

0.04 0.04 [0.28] [0.28] -0.03 [-0.32] 0.08 0.07 [2.24] [2.55] 1.45 1.46 [5.43] [5.26] 0.09 0.09 3.60 3.64

-0.07 [-2.5] 0.71 [3.1] 0.11 3.53

0.27 0.19 [2.63] [1.66] -0.03 [-1.30] 0.69 0.67 [2.95] [2.93] 0.13 0.13 3.55 3.57

-0.11 -0.11 [-0.99] [-0.98] 0.19 [1.62] -0.08 -0.05 [-2.58] [-1.46] 0.73 0.69 [3.11] [2.96] 0.11 0.13 3.60 3.64

66

Table A.19. Risk Exposures: Fung-Hsieh Factors

This table reports regression results for the risk exposures of the BMS portfolios of nancial FX market endusers, that is, asset managers and hedge funds. The options-based factors are intended to capture non-linear payo features that are typical of hedge fund returns (Fung and Hsieh, 2001). Panel A considers the ve market timing-factors for various asset classes (BN - Bonds, FX, CM - commodities, Eq - equities, and IR short term interest rates) as the starting point. Panel B uses the Fung-Hsieh 7-factor model as the starting point, as outlined in the main text. The Table shows results for four parsimonious model specications (i-iv) where the factors are selected according to the Schwarz criterion (joint estimation of the equation for asset managers and hedge funds BMS returns). Specication (v) includes all factors jointly. We further report the estimated intercept , the adjusted R2 and the BIC computed for the two-equation system (Sys-BIC). Below the regression coecients, t-statistics based on HAC standard errors are reported (in brackets).
A. (i) PTFSBD PTFSF X PTFSCM PTFSIR PTFSEQ 2 R Sys-BIC B. (i) rm SMB PTFSF X PTFSCM PTFSBD T S DF 2 R Sys-BIC Asset Managers (ii) (iii) (iv) Hedge Funds (ii) (iii) (iv)

(v)

(i)

(v)

1.19 1.20 [0.54] [0.54] 2.69 3.67 2.50 3.00 [2.67] [2.70] [2.78] [2.78] -1.88 [-1.24] 2.28 2.70 2.20 2.50 [3.45] [3.61] [3.47] [3.78] -0.45 [-0.23] 1.20 1.25 1.22 1.24 1.20 [5.51] [5.18] [5.4] [4.8] [4.11] 0.14 0.07 0.11 0.13 0.13 3.55 3.58 3.59 3.61 3.75 Asset Managers (ii) (iii) (iv)

-1.95 -1.96 [-1.24] [-1.30] -3.37 -3.87 -3.06 -3.49 [-2.97] [-2.62] [-2.70] [-3.13] 1.63 [1.07] -1.16 -1.69 -1.04 -1.29 [-1.32] [-1.49] [-1.12] [-1.78] 0.36 [0.17] 0.93 0.91 0.91 0.86 0.90 [4.25] [4.24] [4.02] [4.22] [4.15] 0.11 0.09 0.04 0.12 0.11 3.55 3.58 3.59 3.61 3.75 Hedge Funds (ii) (iii) (iv)

(v)

(i)

(v)

0.03 [0.53] 0.07 [1.13] 2.48 3.67 2.13 [2.39] [2.70] [1.77] 0.46 [0.34] 1.68 0.98 [0.82] [0.54] -0.90 [-1.1] 3.72 4.75 4.38 3.99 [2.05] [2.49] [2.48] [2.02] 1.26 1.29 1.25 1.34 1.27 [5.11] [4.97] [5.18] [4.58] [4.86] 0.12 0.09 0.07 0.09 0.11 3.55 3.56 3.58 3.61 3.89

-0.03 [-1.10] 0.06 [0.52] -2.89 -3.87 -3.05 [-2.85] [-2.62] [-2.64] 0.96 [0.61] -2.31 -1.95 [-1.60] [-1.38] -0.24 [-0.37] -3.06 -4.26 -3.75 -2.79 [-1.86] [-2.28] [-1.88] [-1.56] 0.90 0.87 0.91 0.79 0.83 [4.17] [3.91] [4.24] [3.78] [4.03] 0.13 0.09 0.09 0.10 0.11 3.55 3.56 3.58 3.61 3.89

67

Table A.20. Risk Exposures T15/T9

This table reports regression results for the risk exposures of the BMS portfolios computed from the total ows based on either 15 (T15) or 9 currencies. The methodological framework in Panel A is a modied linear Fung-Hsieh model with eight factors as outlined in the main text. Panel B also accounts for the conditional exposure to stock market returns by including additional interaction terms of market returns. The three conditioning variables are rst dierences of the 3-month T-Bill rate, the VIX and the TED spread. The Table shows results for four parsimonious model specications where the factors are selected according to the Schwarz criterion (joint estimation of the equation for T15 and T9 BMS returns). Results for the other factors are not reported. We further report the estimated intercept , the adjusted R2 and the BIC computed for the two-equation system (Sys-BIC). Below the regression coecients, t-statistics based on HAC standard errors are reported (in brackets). A. (i) rm DOL T S DF 2 R Sys-BIC B. (i) DOL rm TED(t-1) rm TB(t-1) DF 2 R Sys-BIC 0.26 [2.73] -0.36 [-2.46] -4.79 -3.70 -3.36 -5.16 [-2.24] [-2.03] [-2.05] [-2.19] 0.86 0.85 0.82 0.89 [4.63] [4.67] [4.53] [4.22] 0.11 0.13 0.14 0.10 3.19 3.20 3.21 3.21 (ii) -0.08 [-0.54] -0.96 [-1.45] -4.72 -5.20 -5.06 -4.74 [-2.17] [-2.17] [-2.11] [-2.41] 0.88 0.92 0.90 0.91 [4.87] [4.51] [4.97] [4.85] 0.10 0.10 0.10 0.11 3.19 3.21 3.23 3.25 T15 (iii) (iv) -0.07 [-0.42] 0.26 [2.03] 0.05 [0.19] 1.30 1.15 2.71 -0.20 [0.92] [0.76] [1.15] [-0.15] 1.03 1.03 0.99 1.15 [4.3] [4.3] [4.13] [4.73] 0.00 -0.01 0.04 0.04 3.19 3.20 3.21 3.21 (i) (ii) (ii) T15 (iii) -0.04 [-0.79] -0.28 [-2.74] -0.60 [-1.18] -0.22 0.59 1.35 [-0.17] [0.48] [0.94] 1.17 1.09 1.06 [4.94] [4.48] [4.41] 0.05 0.03 0.00 3.21 3.23 3.25 T9 (iii) (iv) -0.27 [-2.56] (iv) (i) (ii) T9 (iii) -0.09 [-2.11] (iv)

1.37 [0.97] 1.05 [4.46] 0.00 3.19

68

Table A.21. Marginal Forecast Performance: Four-Factor Adjusted Excess Returns This table shows excess returns for BMS portfolios sorted on lagged order ow as in Table III. We do not only sort on order ow of the previous day but also allow for longer lags of up to nine days between order ow signals and portfolio formation. Portfolios are rebalanced daily. T15 denotes portfolios sorts on total order ows and the sample of all 15 currencies, and T9 denotes portfolios sorts on total order ows and the sample of 9 developed currencies; AM, HF, CC, and PC denote portfolios sorts on asset managers, hedge funds, corporate clients, and private clients order ows, respectively. Compared to Table III which reports unadjusted excess returns, we report adjusted excess returns based on the Carhart (1997) four-factor model. Controlling for M KT RF , HM L, SM B , and U M D T15 T9 AM HF CC PC 10.31 [3.92] 12.52 [4.53] 24.87 [8.90] 24.59 [8.89] 10.55 -1.11 [4.50] [-0.44] 7.83 -4.10 [3.10] [-1.56] 3.31 0.33 [1.38] [0.14] 6.20 -1.76 [2.29] [-0.68] 0.27 [0.11] 2.26 [0.88] 2.21 -2.15 -0.72 [0.94] [-0.86] [-0.31] 1.45 -0.97 -1.93 [0.54] [-0.37] [-0.78]

16.06 25.30 8.85 -1.55 2.68 0.23 -0.06 3.67 2.31 -2.98 [5.74] [8.88] [3.15] [-0.55] [1.06] [0.09] [-0.02] [1.29] [0.82] [-1.03] 9.95 28.37 1.44 -2.90 0.05 -6.09 2.88 -0.20 -4.69 -0.72 [3.83] [9.42] [0.53] [-1.14] [0.02] [-2.34] [1.11] [-0.07] [-1.78] [-0.26] -4.24 -8.30 -1.82 2.48 -5.23 2.13 -0.52 1.48 -0.09 3.41 [-1.54] [-2.88] [-0.59] [0.96] [-2.01] [0.82] [-0.20] [0.57] [-0.03] [1.32] -14.35 -34.38 3.45 2.29 -3.37 -1.08 2.33 -1.65 0.80 1.83 [-5.09] [-10.97] [1.29] [0.84] [-1.13] [-0.36] [0.87] [-0.61] [0.30] [0.68]

69

Figure A.1. BMS Excess Returns in Event Time: Four Liquid Currency Pairs

This gure is similar to Figure 3 but here BMS returns are based on sorting four liquid currencies (EUR/USD, JPY/USD, GBP/USD, CHF/USD) into two portfolios.

70