BIS on Forex

7/29/2019 BIS on Forex

1/74

BIS Working PapersNo 405

Information Flows in DarkMarkets: Dissecting CustomerCurrency 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.


2/74

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)


3/74

Information Flows in Dark Markets:

Dissecting Customer Currency Trades

Lukas Menkhoff Lucio Sarno Maik Schmeling Andreas Schrimpf

This version: March 5, 2013

Abstract

We study the information in order flows of different 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 distinguishingtrades by key types of foreign exchange end-users. We find that order flows are highly

informative about future exchange rates and provide significant economic value for the

few large dealers who have access to these flows. Moreover, customer groups system-

atically engage in risk sharing with each other and differ markedly in their predictive

ability, trading styles, and risk exposure.

JEL Classification: 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, TerrenceHendershott, 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, ChristianUpper, Adrien Verdelhan, Michel van der Wel, as well as participants at several conferences, workshops andseminars for helpful comments and suggestions. We are very grateful to Gareth Berry, Geoffrey Kendrickand UBS for providing us with the proprietary data used in this study, and for numerous conversationson the institutional details of foreign exchange trading at UBS. Sarno acknowledges financial support fromthe Economic and Social Research Council (No. RES-062-23-2340) and Menkhoff and Schmeling gratefullyacknowledge financial support by the German Research Foundation (DFG). The views expressed in this paperare those of the authors and do not necessarily reflect those of the Bank for International Settlements.

Kiel Institute for the World Economy and Department of Economics, Leibniz Universitat Hannover,Konigsworther Platz 1, 30167 Hannover, Germany, Tel: +49 511 7624552, Email: [email protected].

Cass Business School and Centre for Economic Policy Research (CEPR). Corresponding author: Facultyof 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: [email protected].

Faculty of Finance, Cass Business School, City University London, 106 Bunhill Row, London EC1Y 8TZ,UK, Email: [email protected].

Bank for International Settlements and CREATES, Centralbahnplatz 2, 4002 Basel, Switzerland. Tel:+41 61 280 8942. Email: [email protected].


4/74


5/74

The foreign exchange (FX) market is the largest financial 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 othermarket participants and available liquidity in different 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 cus-

tomer 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

different customer groups FX trading, e.g., do they speculate on trends or are they con-

trarian 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 different types of end-users and enhances the knowledge about the ecology

of the worlds largest financial market.

We empirically tackle these questions by means of a unique data set covering more than

ten years of daily end-user order flow for up to fifteen currencies from one of the top FX

1Duffie (2012) provides a general overview of how opaqueness and market structure impact price discoveryand trading in dark markets, that is, OTC markets.

1


6/74

dealers, UBS. The data are disaggregated into four different groups of financial (asset man-

agers and hedge funds) and non-financial (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 find that: (i) Order flow by end-users is highly informative for

future exchange rate changes and carries substantial economic value for the dealer observing

these flows; (ii) there is clear evidence that different end-user segments actively share risks

with each other; and (iii) end-user groups follow very heterogeneous trading styles and strate-

gies and differ in their exposures to risk and hedge factors. This heterogeneity across players

is crucial for risk sharing and helps explain the vast differences in the predictive content of

flows across end-user segments that we document in this paper.

To gauge the impact of order flow 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 flow and is a pure out-of-sample approach

in that it only conditions on past information. Specifically, 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) flows.2 The infor-

mation contained in customer trades is highly valuable from an economic perspective: We

find that currencies with the highest lagged total order flows (that is, the strongest net buy-

ing pressure across all customer groups against the U.S. dollar) outperform currencies with

the lowest lagged flows (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 flow, 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 flows

leads to a spread of about 10% p.a. Flows by corporate customers basically generate no spread

in returns, whereas private customers flows even lead to a highly negative spread (about -14%

p.a.). In sum, we find that order flow contains significant economic value for a dealer with

2Lustig and Verdelhan (2007) were the first to build cross-sections of currency portfolios.

2


7/74

access to such information. Hence, the trend towards more market concentration observed

in FX markets over recent years clearly benefits large financial 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 inOTC markets, as trades by different categories of customers convey fundamentally different

information for price movements.

What drives the predictive content in flows? We investigate three main channels. First,

order flow could be related to the processing of information by market participants via the

process of price discovery. According to this view, order flow acts as the key vehicle that

impounds views about (economic) fundamentals into exchange rates.3 If order flow contains

private information, its effect on exchange rates is likely to be persistent. Second, there

could be a price pressure (liquidity) effect 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 flows 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 flow is linked to returns due to the different risk sharing motives and risk exposures of

market participants. For example, order flow could reflect 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-financial customers

are primarily concerned about laying off 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 flow,

we find clear differences across the four segments of end-users. Asset managers flows are

associated with permanent shifts in future exchange rates, suggesting that their order flow is

3See, 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 flow in a structural way to volatility (Berger,Chaboud, and Hjalmarsson, 2009) or directly to exchange rate fundamentals (Chinn and Moore, 2011).

4Several studies explore the underlying mechanism for the impact of order flow and discuss the evidencein terms of information versus liquidity effects (e.g. Berger, Chaboud, Chernenko, Howorka, and Wright,2008; Fan and Lyons, 2003; Marsh and ORourke, 2005; Osler, Mende, and Menkhoff, 2011; Menkhoff andSchmeling, 2010; Phylaktis and Chen, 2010; Moore and Payne, 2011; Ito, Lyons, and Melvin, 1998).

3


8/74

related to superior processing of fundamental information.5 In contrast, hedge funds flows

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 short-

term liquidity effects but not with fundamental information processing. Corporate customersand private clients flows, however, seem to reflect largely uninformed trading.

Our results also point to a substantial heterogeneity across customers in their trading

styles and risk exposures, giving rise to different motives for risk sharing. First, we find

that the trades of various end-user groups react quite differently to past returns. Asset

managers tend to betrend-followers(positive feedback traders) with regard to past currency

returns. By contrast, private clients tend to be contrarians (negative feedback traders).

The latter finding squares well with recent findings for equity markets by Kaniel, Saar, and

Titman (2008) who show that individual equity investors behave as contrarians, effectively

providing liquidity for institutional investors. Different from their results, however, private

clients do not directly benefit from serving as (implicit) counterparties of financial customers

in FX markets. Second, the flows of most customer groups are negatively correlated over

short to intermediate horizons, suggesting that different 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 significant

proportion of risk is shared among end-users in the customer-dealer segment. Third, we

find 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 flows is not due to risk taking but likely reflects

superior information. Hedge funds, by contrast, are significantly 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 effectively providing liquidity and selling insurance

to other market participants. For non-financial customers there is some evidence of hedging

but it is not strong enough to fully explain their negative forecast performance arising from

5This information processing can come in different ways, e.g., a more accurate and/or faster interpretationof macroeconomic news releases, and better forecasting of market fundamentals such as liquidity and hedgingdemands of other market participants.

4


9/74

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 flow is crucial for understanding

how information is incorporated into exchange rates. It is well known from the literature

that order flow 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 flow predicts exchange rates. A few papers have shown that FX order flow

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 find robust predictive power of exchange rates

by order flow in the first place (see, e.g., Sager and Taylor, 2008). Our work is also related

to a different 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 first to show

that order flow forecasts currency returns in an out-of-sample forecasting setting by forming

order flow portfolios. This multi-currency investment approach provides an intuitive mea-

sure of the economic value of order flow for the few large dealers observing these flows. 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 affected by trends in exchange rates, most

notably the U.S. dollar. Our portfolio procedure, by contrast, studies exchange rate pre-

6There is also evidence that marketwide private information extracted from equity order flow is useful forforecasting currency returns (Albuquerque, de Francisco, and Marques, 2008).

7Lustig 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 Menkhoff, Sarno, Schmeling, and Schrimpf (2012a,b) all build currencyportfolios to study return predictability and/or currency risk exposure.

5


10/74

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 first

to test whether risk exposure drives the information in customer order flows. We show how

different key FX market players trade, e.g., to which extent they rely on trend-following orbehave as contrarians, and in which ways they are exposed to systematic risk. We find strong

evidence of heterogeneity in the exposures and trading behavior across different groups of

market participants. These findings indicate that there is significant risk sharing between fi-

nancial and non-financial customers as well as between different groups of financial customers

(leveraged versus real money managers).

Taken together, these results have implications for our general understanding of informa-

tion flows in dark markets and how large dealers in OTC markets benefit from observing a

large proportion of the order flow. These results also add to our general understanding of

how risk is shared in financial markets due to different 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 flow, Section III empirically

investigates alternative underlying reasons for why order flow 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 flows 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 flow from the end-

user segment of the FX market and not from the inter-dealer market. This is important

since microstructure models suggest that the information in flows stems from trading with

customers and not from inter-dealer trading (e.g. Evans and Lyons, 2002a). Order flows 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


11/74

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 flow therefore does not measure trading volume but net buying (or selling) pressure, as

mentioned above. Our order flow data are available both in aggregated form and at a higherlevel of granularity allowing for a differentiation across end-user groups.

Aggregate order flow. Aggregate order flows, that is, aggregated across customers (re-

gardless 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 flows as total flows since they are aggregated across all customers of UBS.

A natural question is whether flows 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 flows of all other dealers, there are good reasons to believe

that the flows employed in our paper are highly correlated with a large portion of end-user

order flows. 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 significant 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 flows are highly correlated with flows observed at, e.g., Deutsche Bank, Barclays,

Citi, or JP Morgan, which in turn implies that our order flows are representative of the top

end of customer trading in the FX market.

8Note that most UBS FX customers are in fact big players and include other banks, many large assetmanagement firms and hedge funds, and a large fraction of wealthy private clients. According to the Eu-romoney survey, UBS has a particularly high market share in FX business with financial customers (banks,real money and leveraged funds).

7


12/74

Table I about here

Table I shows descriptive statistics for total flows (in billion USD). Daily order flows are

largest on average for EUR, JPY, and CHF. The pair with the largest average imbalance (inabsolute 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 flows are fairly small relative to gross daily trading

volume in FX markets.9 Flows are fairly volatile, however, which means that order flow

imbalances can frequently be very large. Daily flows tend to be positively autocorrelated,

but the degree of autocorrelation is very small albeit sometimes statistically significant. There

is also a clear pattern in standard deviations. Major currencies, such as the EUR, CHF, JPY,

GBP, have much larger variation in order flows and, hence, a larger absolute size of order

flows 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 flows 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 flows display a high kurtosis (especially the British pound), which is

largely driven by some days with extremely high (in absolute value) order flows. Eliminating

these few outliers does not change our results reported below.

Disaggregated order flow. We also obtain order flows 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 flows 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 classified9To 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) figures for both thecustomer-dealer segment and the inter-dealer market are based on data collected from about 4,000 reportingdealers worldwide.

10In addition, the volatility of flows also varies over time and flows tend to become increasingly volatiletowards the end of the sample. Also for this reason, some form of standardization is necessary.

11The nine currencies are: AUD, CAD, EUR, JPY, SEK, NZD, NOK, CHF, and GBP.

8


13/74

as hedge funds. Hedge funds are unregulated entities, whereas asset managers are regulated.

The corporate segment includes non-financial 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-financial corporations, with the exception of those pursuing ahighly leveraged investment strategy, which are classified 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 financial 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 differ considerably in the degree of

informedness and sophistication. One of the key features of OTC markets, that is, the non-

anonymous nature of transactions can thus further enhance the informational advantages ofdealers in dark markets.

The order flow data are assembled as follows. Each transaction booked in the UBS

execution system at any of its world-wide offices is tagged with a client type. At the end of

each business day, global transactions are aggregated for each customer group. Order flow is

measured as the difference 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 flow data

are reported in Table A.1 of the Internet Appendix.

Exchange rate returns and excess returns. For our empirical analysis below, we com-

plement these order flow data with daily spot exchange and forward rates from Reuters

(available from Datastream). We denote log changes of spot exchange rates as exchange rate

12Our data are raw order flow data with filtering limited to the most obvious cases. For instance, data

are adjusted for large merger and acquisition deals which are announced well in advance. Cross-bordermergers and acquisitions involve large purchases of foreign currency by the acquiring company to pay thecash component of the deal. These transactions are generally well-publicized and thus are anticipated bymarket participants. Finally, FX reserve managers, UBS proprietary traders and small banks not participatingin the inter-dealer market are excluded from the data. Flows from FX reserve managers are stripped outdue to confidentiality issues, flows from proprietary traders because they trade with UBS own money, whilesmall banks represent small customers less concerned about the FX market.

9


14/74

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 flows 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 differential

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 differentials. Interest rate differentials 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 condi-

tioning on lagged order flow. This provides a straightforward and intuitive assessment of the

13This approximation is exact if covered interest rate parity (CIP) holds, which tends to be the case atdaily or even shorter horizons in normal times (Akram, Rime, and Sarno, 2008). There have been violationsof this no-arbitrage relation over the recent financial crisis. As we show below, the results in this paper areentirely driven by changes in spot rates, whereas interest rate differentials only play a negligible role. Thus,the results do not depend on whether CIP holds or not.

10


15/74

economic value of order flow in predicting currency excess returns.

As a benchmark test, we first sort currencies into portfolios based on (lagged) total order

flows for each currency. Specifically, we sort currencies into five portfolios (P1, P2, ..., P5)

depending on their total order flow 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 difference 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 flows. Before sorting currencies into portfolios, we need to make

sure that order flows are comparable across currencies. As the absolute size of order flows

differs across currencies (as shown above in Table I) it is not sensible to sort currencies based

on raw order flows. To allow for meaningful cross-currency comparisons, it is necessary to

standardize flows. We do this by dividing flows by their standard deviation to remove the

difference in absolute order flow sizes across currencies

xRj,t = xj,t(xj,t59;t) , (3)

where xRj,t denotes order flow standardized over a rolling window and xj,t denotes the raworder flow. In our baseline results, we compute the standard deviation of flows via a rolling

scheme over a 60-day rolling window. Robustness tests based on alternative approaches to

standardize flows are reported in a separate Internet Appendix.14

Portfolio excess returns. Table II shows average annualized excess returns for order flow

14In these robustness exercises, we also report results with longer rolling windows of up to three years aswell as for an expanding window. Furthermore, we provide tests where we standardize both with respect tovolatility as well as the mean. Finally, we also consider a standardization scheme based on gross FX turnoverdata for different currencies drawing on data from the BIS FX triennial survey. These tests, reported in theseparate Internet Appendix to conserve space, show that our results are not sensitive with regard to the wayflows are standardized.

11


16/74

portfolios (P1, P2, ..., P5), where P1 contains the three currencies with the lowest lagged stan-

dardized order flow and P5 contains the three currencies with the highest lagged standardized

order flow. 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 five 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 significant.

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 significance. Thus, order flows carry significant 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 significant share of end-user order flow 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 find 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

15Sub-sample tests for a pre-crisis subperiod from January 2001 to June 2007, and a crisis/post-crisissubperiod from July 2007 to May 2011 are reported in the Internet Appendix.

12


17/74

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 flow is informative about

future spot rates and not about interest rate differentials.

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

significantly 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 flow

is truly informative about future excess returns. This prediction is significantly 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 significant relationship between order flow 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

16The MR statistic tests for a monotonically increasing return pattern, whereas the Up (Down) test issomewhat less restrictive and simply tests for a generally increasing (decreasing) pattern without requiringmonotonicity in average portfolio returns. Specifically, the MR test requires that the return pattern ismonotonically increasing P1 < P2 < ... < P5 and formulates the null hypothesis as H0 : 0 and thealternative hypothesis as Ha : mini=1,...,4i > 0, where is a vector of differences in adjacent average

portfolio excess returns (P2 P1, P3 P2, P4 P3, P5 P4) and i is element i of this vector. TheUp test formulates the null hypothesis of a flat pattern H0 : = 0 and the alternative hypothesis asHa :

4

n=1 |i|1{i > 0} > 0, so that the test is less restrictive and also takes into account the size andmagnitude of deviations from a flat return pattern. The Down test follows in an analogous way.

13


18/74

B. Portfolios Conditioning on Disaggregated Order Flow

If superior information processing or genuine forecasting ability drive our results above, one

expects clear differences in the forecasting power of different customers flows, depending on

the groups characteristics (see, e.g., Fan and Lyons, 2003; Evans and Lyons, 2007, among

others). Specifically, one would expect to see superior information in flows of financial cus-

tomers, given that non-financial 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

flows. We closely follow the earlier approach with the exception that we only build four

portfolios (rather than five) here since we only have disaggregated flows 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 clear-

cut. Asset managers net buying or selling pressure of currencies is the most informative

about subsequent exchange rate behavior. Conditioning on asset managers flows 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 flows actually generate

a negative spread in portfolio excess returns of about 4% and 14%, respectively.17

Theresults point towards substantial differences 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 flow of

17Table A.5 shows results for spot rate changes instead of excess returns, which display no qualitativedifferences.

18Table A.6 in the Internet Appendix also shows that excess returns to the BMS portfolios based ondifferent customers flows are not highly correlated. Hence, the information contained in the different flowsappears to stem from different sources. In practice, this also means that BMS portfolios could be combinedto obtain even higher Sharpe Ratios. For example, a combined portfolio long in the asset managers BMSportfolio and short in the private clients BMS portfolio yields an annualized Sharpe Ratio of 2.19, which issubstantially higher than the individual Sharpe Ratios.

14


19/74

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 flows whichis highly significant. By contrast, we find a monotonically decreasing pattern in average

excess returns for portfolios based on private customers flows, and marginally significant

evidence for a decreasing pattern in portfolios based on corporate flows.

Hence, it is not the case that all order flow is equal in terms of its information content for

exchange rates. Instead, financial customers flows (asset managers and hedge funds) account

for the positive relation between lagged flows and future exchange rate returns uncovered in

the previous section. Flows by corporates are more or less uninformative, whereas private

clients flows even forecast returns in the wrong direction. Using total end-user order flow,

which is likely to be dominated by financial customers due to their higher trading volume19,

masks these differences and might even lead to wrong inference about the link between flows

and returns. In a nutshell, what matters for the relation between end-user order flows and

future returns is disaggregated data since the information content of flows 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 different across customer groups, even

when comparing, for example, asset managers and hedge funds. Both groups BMS portfo-

lios generate significant excess returns but returns for hedge funds are much more volatile

than those of asset managers. Hence, we will investigate possible sources of these different

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 flows and returns over

the subsequent trading day. An interesting question, however, is whether the information

19This is especially true for the order flow employed in this paper since UBS is one of the largest dealersin FX and has a high proportion of financial customers (relative to corporate clients).

15


20/74

contained in order flow quickly decays or whether it is useful for forecasting returns over more

than one trading day.

We examine the marginal predictive content of flows by forming portfolios as in the

analysis above, but we now allow for a longer lag between the order flow signal and portfolio

formation. Table III contains the results for different lags of 0, 1, 2, . . . , 9 days. To be more

specific, a lag of 0 days means that flows 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 flows of day t are used to forecast returns of trading day t + 3.

Table III about here

Results in Table III show that order flow appears to be most informative for the first

two to three days after portfolio formation and that the information in flows becomes in-

significant afterwards. Hence, the information contained in daily flows is fairly short-lived

and is impounded relatively quickly into exchange rates, especially considering that the order

flows 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 finding is in contrast to, e.g., Evans and Lyons (2005) who study a shorter and smallersample and find that times-series predictability of returns by order flow 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 flow 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 profits 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 flow 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 fixed-effects. For example, it could be the

16


21/74

case that asset managers and hedge funds order flow mimicking portfolios simply reproduce

a carry trade (Burnside, Eichenbaum, Kleshchelski, and Rebelo, 2011; Lustig, Roussanov,

and Verdelhan, 2011; Menkhoff, Sarno, Schmeling, and Schrimpf, 2012a) or that their order

flow just picks up momentum effects in currency returns (Menkhoff, Sarno, Schmeling, andSchrimpf, 2012b).

Specifically, we run panel regressions of the general form

rxj,t+1 = cOFct + 1(i

j,t it) + 2rxj,t + 3rxj,t60;t1 + j,t+1 (4)

where j (1,...,N) indexes currencies, rx denotes currency excess returns, OFc denotes order

flow of customer group c, (i

j,t it) denotes interest rate differentials (carry), and rxt andrxt60;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 fixed-effects (we also report results without fixed-effects

below). Standard errors are clustered by currency pair. Note that these panel regressions

employ non-standardized order flows and are based on individual currency returns and not

on portfolio returns.

Results are shown in Table IV and corroborate our findings based on our portfolio ap-

proach above, namely that order flows of financials positively predict future excess returns,

whereas flows by non-financial end-users negatively forecast returns. Based on specification

(vi) in Table IV, the coefficients on lagged order flow imply that a positive order flow of USD

1 billion forecasts a four basis point (b.p.) higher excess return on the following day for asset

managers flows, 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

effects 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

20Using other windows of less or more than 60 trading days does not yield qualitatively different results.

17


22/74

flow and future FX excess returns remains very strong when controlling for two common

predictors of returns in FX markets, interest rate differentials 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 differentials, however, do not drive out the information contained in

order flows and they become insignificant once we include cross-sectional fixed-effects in the

regression. In our panel regressions, lagged currency returns do not have consistent predictive

content beyond order flow and carry, which is in line with recent evidence in Menkhoff, Sarno,

Schmeling, and Schrimpf(2012b), who show that FX momentum strategies are not profitable

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 flow 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 flow has a permanent or transitory effect in prices is a

central theme in the earlier microstructure literature (see Hasbrouck, 1991a,b). A transitory

movement is interpreted as suggesting that order flow effects are merely due to short-term

liquidity or price pressure effects which eventually die out, whereas a permanent movement

in spot rates would indicate that order flow conveys information about fundamentals.21 More

specifically, a permanent price impact would most probably indicate that order flow 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 find substantial heterogeneity with

21One strand of literature argues that order flow is the conduit by which information about fundamentalsis impounded in prices and therefore has permanent effect on exchange rates (e.g., Evans and Lyons, 2002a;Brandt and Kavajecz, 2004; Evans and Lyons, 2007, 2008). Another strand of the literature suggests thatorder flow matters due to downward sloping demand curves or illiquidity and, hence, that order flow onlyhas a transitory impact on prices (e.g., Froot and Ramadorai, 2005).

18


23/74

regard to the forecasting power of different customer groups order flows. Therefore it is

particularly interesting to find out if all (or some) customers flows signal information relevant

for permanent changes in FX rates or whether some customer groups order flow simply exerts

price pressure and liquidity effects.

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 flow. The solid lines

show the cumulative excess returns (in basis points), whereas the shaded areas show 95%

confidence intervals based on a moving-block bootstrap with 1,000 repetitions. Total flows

for all 15 currencies (T15) forecast a permanent change in spot rates which is statistically

significantly different 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

effect of the order flow signal levels out. Importantly, these findings suggest that order flow

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 ob-

serve 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 confidence interval includes zero. Hence, there is much less evidence that

order flow conveys information about fundamentals when only looking at major developed

markets. This finding makes sense, however, since the major currency markets are most

probably more researched and more efficient than smaller currency markets so that the scope

for superior information processing is reduced.22

22This may be interpreted in the context of the adaptive markets hypothesis (see e.g. Neely, Weller, andUlrich, 2009, for an analysis in FX markets).

19


24/74

As a natural next step, we also examine the same question separately for disaggregated

order flows (lower panels of Figure 2). Results are clear-cut. The only end-user group with a

statistically significant 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 finding for total flows of the nine major currencies above, it is interesting to see

that asset managers flows are indeed associated with permanent spot rate changes. Hence,

order flow of asset managers seems to be related to the processing of fundamental information

whereas hedge funds order flow corresponds to short-lived information unrelated to funda-

mental information. Similarly, it seems reasonable that the negative relation between private

clients flows and future spot rates eventually dies out over time.

These findings are novel in the literature and suggest that order flows by different end-

user groups even by the two financial customer groups embed different information for

future exchange rates. These differences can arise either because they are based on different

mechanisms to process information or because of different trading motives and hedging needs.

To explore this further, we investigate the drivers of order flow in more detail and shed light

on the observed differences in end-user order flows.

B. Risk Sharing Among Foreign Exchange End-Users

The analysis above suggests that asset managers flows are related to the processing of fun-

damental information that is quickly, but permanently, impounded into prices whereas the

other customer groups flows 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 reflect a compensation for taking on such risks. While

these are just examples, a risk sharing story in general implies that we observe customers sys-

tematically trading in opposite directions and that their portfolios load on different sources

of systematic risk. We investigate these issues below.

20


25/74

Portfolio returns in event time. We first provide a more detailed look at the return be-

havior around portfolio formation dates to better understand differences in customer groups.

Figure 3 shows the average annualized BMS excess return for the five days prior to portfo-

lio formation (5,4, ...,1), the day of portfolio formation 0, and the first ten days afterportfolio formation (1, 2, ...,10). Shaded areas correspond to 95% confidence 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 con-

trarian fashion. The pattern for hedge funds and corporates is less clear. Second, formation

day returns (day 0) are significantly different from zero for all four customer groups. How-

ever, 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 expla-

nation seems more reasonable especially for private clients who do not trade large enough

volumes to move prices in FX markets.

Overall, these findings suggest that customer groups trading positions at least partly

offset each other, as asset managers and private clients clearly differ in terms of their trend-

following behavior. This finding is different from equity markets where Kaniel, Saar, and

Titman (2008) find 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 find 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 findings, we next look at the corre-

lation of customer groups flows directly. While there is little contemporaneous correlation in

flows, as noted above (see Table A.2 in the Internet Appendix), it is nevertheless interesting

21


26/74

to look at flows over longer horizons to find 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 flow correlations between customer groups at some horizons.

Figure 4 plots contemporaneous correlations between flows of all four customer groups for

horizons of one to 60 days (using overlapping observations) where the shaded areas correspond

to 95% bootstrap confidence intervals. For the two financial customer groups, there is a small

and short-term negative flow 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 financial and

non-financial customers are significantly negative at all horizons while there is no significant

correlation between flows of the non-financial customer groups. These results are generally

in line with a risk sharing story where financial players trade in the opposite direction of

non-financial market participants. This finding 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 off 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 efficiently.

Given the negative correlation of flows 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 flows. As a natural next step we seek to provide a better understanding of the

drivers of end-user order flows and shed light on the source of the negative flow correlations

discussed above. First, we examine whether the flows of some customer groups systematically

lead the flows of other groups. Second, we study whether customers flows differ in their

response to lagged asset returns in other key asset classes. In this context we are interested

in the possible effects of portfolio re-balancing on the end-user demand for currencies (Hau

and Rey, 2004). To investigate this, we run panel regressions of order flows on lagged flows

and further explanatory variables, such as interest rate differentials (it it), lagged exchange

22


27/74

rate changes over one and 20 days (st,st1;t20), lagged stock returns (reqt , r

eqt1;t20), and

lagged bond returns (rbt , rbt1;t20)

OFcj,t+1

= + AM

OFAMj,t

+ HF

OFHFj,t

+ CC

OFCCj,t

+ PC

OFPCj,t

+1(ij,t it) + 2sj,t + 3sj,t1;t20 (5)

+4reqj,t + 5r

eqj,t1;t20 + 6r

bj,t + 7r

bj,t1;t20 + j,t+1,

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 fixed-effects. 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 specification which only includes lagged flows and one which additionally includes interest

rate differentials and lagged returns.23 Looking first at the specifications which only include

lagged flows, we find that the flows of asset managers are significantly related to the flows

by the other groups. These results (akin to simple Granger causality tests) indicate again

that asset managers trade in the opposite direction of non-financial customers. Flows by

hedge funds, on the other hand, do not load significantly on lagged flows of any group, which

shows that asset managers and hedge funds show a quite different behavior. Corporate

flows are positively driven by own lagged flows and those of private clients, whereas flows by

private clients are significantly negatively related to lagged hedge funds flows and significantly

positively autocorrelated. In sum, there is a wealth of interrelationships between customer

flows 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 find that asset managers trade

against the interest rate differential, whereas corporate customers trade with the interest

23Using more than one lag of flows in the regressions generally yields insignificant coefficient estimates sowe restrict the regressions to include one lag of flows.

23


28/74

rate differential. Surprisingly, flows by hedge funds (and private clients) are not affected

by the interest differential suggesting that, on average, carry trading is not a dominant

driver of their flows over our sample.24 Results for lagged exchange rates indicate that

asset managers are trend-followers (positive feedback traders), whereas private clients canbe described as contrarians (negative feedback traders). Asset managers flows also react

significantly positively to lagged equity returns, whereas private clients flows 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 flows 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 inHau and Rey, 2004). However, the results strongly support the notion that flows of different

groups are to some extent driven by the returns of other asset classes, although the factors

that influence flows clearly differ across end-user groups.

Finally, it seems worthwhile mentioning that the estimated overall constant in the panel

regression is significantly negative for hedge funds and corporates but significantly 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

deficit of the U.S. over the last decade, the negative coefficient for corporate clients is not

surprising. Also, the strong inflow of foreign savings into U.S. capital markets over the sample

period makes sense of the positive coefficient for private clients.

24While this result may be surprising, it is worth bearing in mind that it relates to the aggregate hedgefunds community and on average over the full sample. It is therefore entirely possible, or even likely, thatthere is variation across hedge funds and across time: For example, it may well be the case that some hedgefunds follow carry trade strategies and some follow uncovered interest parity (anti-carry trade) strategies, orthat for a particular hedge fund carry trades were implemented for the first part of the sample and deleveraged

during the second part of the sample which is characterized by the recent global crisis. Thus, our result isnot intended to detect the individual behavior of hedge funds.

24


29/74

C. Differences in Risk Exposures

Finally, we investigate if differences 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 pre-

mium 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 option-

like payoffs 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

rxp;t = +Kk=1

kFk;t +J

j=1

jrm;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 (TS), and changes in credit spreads (DF).

It further includes returns on portfolios of lookback straddles for FX futures and interest

rate futures, denoted by PTFS FX and PTFS IR respectively. We augment this sub-set of

factors from Fung and Hsieh (2004) by additional factors that are intended to capture FX-

related risk. We include the Dollar risk factor (DOL) and the carry factor (HMLFX) by

Lustig, Roussanov, and Verdelhan (2011) as well as a factor-mimicking portfolio of global FX

volatility (V OLFX) (Menkhoff, Sarno, Schmeling, and Schrimpf, 2012a). Following Patton

25


30/74

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 overfitting, 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

differences which might explain the variation in BMS excess returns. However, as financial

and non-financial customers are likely to be very different, we focus on asset managers versus

hedge funds in the first set of results and private clients versus corporates in the second

set of results. More specifically, 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 OLFX) as the single factor.

Other model specifications 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 find that the signs are always

opposite. While asset managers BMS returns load positively on FX volatility shocks, trend-

following 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 different FX trading behavior and exposure

26


31/74

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 fit. 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 different and that their FX trading positions are exposed differently to market

stress. It should be noted, though, that the alphas of asset managers and hedge funds aresignificantly different from zero and still quite large, that is, exposure to risk does not drive

the information in flows 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-financial customers. Still, we find 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 first examine the

effect of transaction costs and then turn to a brief discussion of various other robustness tests25Additional 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 TableA.20 reports results for the BMS portfolio based on total flows for completeness.

26Table A.16 reports pricing errors for the cross-section of order flow portfolios. Specifically, we report theGibbons, Ross, and Shanken (1989) test for the null that the alphas are jointly equal to zero. Corroboratingthe time-series regressions in Tables VI and VII, the test always rejects the null of zero alphas.

27


32/74

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 flow infor-

mation for trading strategies or the effects of transaction costs. This is because our data on

customer order flow 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 flows. 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 flow 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 effects 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% confidence intervals for excess returns before

transaction costs to show the effect of different re-balancing periods. The solid line andshaded area show average net excess returns (p.a.) and 95% confidence intervals when taking

transaction costs into account.

Figure 5 about here

We find that exploiting the information in flows should in practice be feasible for a dealer.

This holds for both sets of currencies T15 and T9. Average excess returns are significantly

27Obviously the data should be used in respect of clients confidentiality and the specific compliance agree-ments governing customers transactions.

28The bid-ask spread data available is for quoted spreads and not effective spreads. It is well known thatquoted spreads are much higher than effective 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 thisnumber seems conservative, though. First, banks with access to this kind of customer order flow data arebig dealers and pay very low spreads since they are key market makers. Second, Gilmore and Hayashi (2011)find in a recent study that transaction costs due to bid-ask spreads are likely to be much lower than our 50%rule. This finding was corroborated by our own conversations with UBS dealers.

28


33/74

different 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 flow 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 flows. First,

we check whether standardizing flows 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 flows relative to total currency trading volume (obtained from the BIS FX triennial

surveys).29 Table A.9 shows the results, which also indicate significant predictability of

returns by order flows. Third, we standardize flows by additionally demeaning flows over

the rolling window (Table A.10). As a final step, we form portfolios based on flows for all

currencies via the following procedure: We cross-sectionally standardize order flows for day

t (we subtract the cross-sectional mean and divide by the standard deviation), rescale these

standardized flows 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 flows forecast returns at longer horizons. To this end, we use

an exponential moving average to sum order flows into the past and then use these lower

frequency flows to build BMS portfolios which we rebalance every 2, 3, 4, 5, 10, 20, 60 trading

days. We report results for two different decay parameters (0.25 and 0.75) in the exponential

moving average in Table A.12. We find that predictability dies out fairly quickly and that

only asset managers flows 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,

29We linearly interpolate between the data of the BIS survey to obtain a daily time-series of trading volumesin USD for the nine developed currencies and then use the ratio of customer flows to total trading volumesas our sorting variable.

29


34/74

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

find that our main results remain qualitatively unchanged.

We next explore whether a specific currency is driving the profitability of the order flow

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 flows) or 8 currency pairs (disaggregated order flows). Table A.14 summarizes the

results from this exercise. We always find the same general return pattern, that is, our main

findings 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 financial 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 flow? 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 find that observing customer order flows in a dark market is highly valuable from

the dealers perspective. Currency excess returns to portfolios mimicking aggregate customer

order flows in real-time are about 10% p.a. and highly significant. 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 find excess returns as high as

15% p.a., that is, non-anonymity further increases the informational advantage of dealers.

The flows by asset managers have the strongest predictive power for exchange rates, likely

reflecting the processing of fundamental information. Their flows have a permanent forecast

30


35/74

power, whereas flows 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 infrastruc-

ture, for example) and to set up trading in a way that reveals end-users identities. Thesefindings 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 differ markedly in their trading strate-

gies 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, flows of different end-user segments tend to be

negatively correlated over longer horizons. These findings 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 financial mar-

kets. Our findings 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 different predictive ability, trading motives, trading styles, and risk

exposures.

31


36/74

References

Akram, Q. Farooq, Dagfinn 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 informa-

tion 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 flow

and exchange rate dynamics in electronic brokerage system data, Journal of InternationalEconomics 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 orderflow 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 flow and the monetary model of exchangerates: Evidence from a novel data set, Journal of Money, Credit and Banking 43, 1599

1624.

Duffie, Darrell, 2012, Dark Markets: Asset Pricing and Information Transmission in Over-

the-Counter Markets (Princeton University Press).

32


37/74

Evans, Martin D., 2010, Order flows and the exchange rate disconnect puzzle, Journal of

International Economics 80, 5871.

, and Richard K. Lyons, 2002a, Order flow 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-Rogoff redux: Micro-based exchange rate forecasting, American Eco-

nomic Review, Papers and Proceedings 95, 405414.

, 2007, Exchange rate fundamentals and order flow, 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 ex-

change, in P. Mizen, ed.: Monetary History, Exchange Rates and Financial Markets: Es-

says 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 flows, 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 effic

BIS on Forex

Documents

Transcript of BIS on Forex