Detecting Crowded Trades in Currency Funds

Momtchil Pojarliev Richard M. Levich

1st Draft: 12 May 2009

Revised: 26 August 2010

Final Version

The financial crisis of 2008 highlights the importance of detecting crowded trades due to
the risks they pose to the stability of the financial system and to the global economy.
However, there is a perception that crowded trades are difficult to identify. To date, no
single measure to capture the crowdedness of a trade or a trading style has developed.
We propose a methodology to measure crowded trades and apply it to professional
currency managers. Our results suggest that carry became a crowded trading strategy
towards the end of Q1 2008, shortly before a massive liquidation of carry trades. The
timing suggests a possible adverse relationship between our measure of style
crowdedness and the future performance of the trading style. Crowdedness in the trend
following and value strategies support this hypothesis.
Our sample period covers 63 months, of which 27 months are effectively an out-of-
sample period. The out-of-sample results confirm the usefulness of our measure of
crowdedness. After a period when carry returns were very favorable, carry became
crowded again in fall 2009, and then experienced a sharp reversal during the European
sovereign debt crises and after the “flash crash” in May 2010.
We apply our approach to currencies but the methodology is general and could be used to
measure the popularity or crowdedness of any trade with an identifiable time series
return. Our methodology may offer useful insights regarding the popularity of certain
trades – in currencies, gold, or other assets – among hedge funds. Further research in this
area may be relevant for investors, managers and regulators.

Key words: Foreign Exchange, Hedge Funds, Style Investing, Crowded Trades
JEL Classification: F31

Contact details:

Momtchil Pojarliev, Hathersage Capital Management LLC, 77 Bleecker St., Ste. 531,
New York, NY 10012, USA; Tel: 212-260-4606; Email: momtchil@hathersage.com .

Richard M. Levich, New York University Stern School of Business, Finance Department,
44 West 4th Street, New York, NY 10012-1126, USA; Tel: 212-998-0422, Fax: 212-995-
4256; Email: rlevich@stern.nyu.edu .
 Detecting Crowded Trades in Currency Funds

1. Introduction

Over the last twenty years, institutional investors have greatly increased their allocations

to alternative investments like hedge funds, and away from traditional assets like equities

and bonds. For example, a survey conducted by the National Association of College and

University Business Officers (2008) found that US university endowments larger than $1

billion allocated more than 20% of their assets to hedge funds. This strategy was partly

the result of a conventional belief that hedge funds could pursue more diverse strategies,

that diversification is the key for successful investing and that the returns on alternative

assets will have little or no correlation with returns on traditional investments.

However, the transfer of substantial assets under management to hedge funds harbored

considerable risks for investors and the broader financial system. Addressing the

Economic Club of New York in 2004, Timothy Geithner, then President of the Federal

Reserve Bank of New York, put the matter quite bluntly. “While there may well be more

diversity in the types of strategies hedge funds follow, there is also considerable

clustering, which raises the prospect of larger moves in some markets if conditions lead

to a general withdrawal from these ‘crowded’ trades.”1 In many ways, Geithner’s

conjecture about returns in crowed trades was realized during the global financial crisis.

1
Remarks by Timothy F. Geithner before the Economic Club of New York, May 27, 2004.
1
In turbulent periods, positioning and being aware of crowded trades become crucial as

traders may try to exit trades at the same time and in the same direction. The

phenomenon of large numbers of traders exiting similar trades at the same time creates

liquidity problems as everyone is rushing to exit a “burning house.” However, in order to

leave a “burning house,” it is not enough to reach the exit, but rather to persuade

someone from the outside to take your place, i.e. to take the other side of the trade.

Therefore, it is hardly a surprise that “positioning” and the concentration, or popularity,

or the crowded nature of certain trades and trading styles are highly discussed topics

among investment managers.2

To illustrate the importance of positioning, as a part of their periodic foreign exchange

research commentaries some banks have introduced Commodities Futures Trading

Commission data on market positioning in currency futures.3 These analyses focus on

the positioning of the non-commercial or speculative accounts. Custodians also are able

to take advantage of their proprietary equity flow data in order to gauge positioning. For

example, State Street Bank and Trust (SSBT) is said to make use of their proprietary

flow data to gauge positioning across different currencies.4 To date, however, no single

measure to capture the crowdedness of a trade or trading style has developed. Identifying

crowded trades is challenging due to the large number of asset classes hedge funds could

2
At the Quant Invest 2009 conference in Paris, Robert Litterman of Goldman Sachs is reported to have
said that computer-driven hedge funds needed to identify new areas to exploit, as some areas had become
so overcrowded that they are no longer profitable. “Quant Hedgies Must Fish In Fresh Waters –
Goldman,” Reuters, December 1, 2009.
3
See Deutsche Bank, Global Markets Research, DB FX Positioning Indices,” May 18, 2009.
4
See Froot and Ramadorai (2005, 2008) for analyses that rely on the SSBT equity flow data.
2
be invested in. Furthermore, given that most databases collect monthly returns, these

data would allow gauging crowdedness only over long term periods.

In this paper, we make use of a new data base of daily data on currency funds to develop

a new approach to detect positioning and identify crowded trades. Because currency

funds have a clearly defined investment universe, they offer a good laboratory for

developing an approach for detecting crowded trades. Furthermore, the high-frequency

data in our sample allow us to develop measures of crowdedness over economically

relevant horizons.5 We apply our approach to currencies but the methodology is general

and could be used to measure the popularity or crowdedness of any trade with an

identifiable time series return.

Currencies may become more highly correlated when investors pursue similar trading

strategies. For example, there is little fundamental reason to expect high correlation

between currency returns on the GBP/CHF cross and the NZD/JPY cross. However, a

carry trader, who establishes a long position in high yielding currencies by taking a short

position in low yielding currencies, could very likely have been long the two crosses

(long GBP vs. CHF and long NZD vs. JPY) over much of the last twenty years, as

interest rates in New Zealand and the United Kingdom were generally higher than

interest rates in Japan and Switzerland, respectively.6 Increasing popularity of the carry

trade could help account for the greater correlation of these two crosses in recent years

5
Gross (2005) suggests that three to four years is the “average life” of investment firms, i.e. the time
frame before an average client will leave if performance disappoints.
6
See Froot and Thaler (1990) for a survey of the carry trade.
3
(see Exhibit 1). Indeed, the rolling annual correlation of the GBP/CHF and NZD/JPY

cross rates rose above 0.50 in the late 1990s, and then fell sharply after the liquidation of

the Yen carry trade between June 1998 and December 1999.7 The correlation again rose

above 0.50 in 2007 and peaked in the fourth quarter of 2008, before falling sharply in the

first half of 2009, a period which once again experienced massive liquidation of carry

trades.8

Analyzed in this way, currency traders appear more focused on exposure to a particular

risk factor or trading strategy and less so on exposure to particular currencies. A large

short JPY exposure might be offset by long CHF exposure as both currencies rallied

during the recent carry trade liquidation. And a seemingly small short JPY exposure

becomes more risky when combined with exposures to other carry trade proxies. For

these reasons, measuring crowdedness by investment style rather than by individual

currency pairs seems preferable, which is a property our technique will exploit.

Previous research (see Pojarliev and Levich (PL), 2008a) has shown that four factors (or

styles), which represents the return on several well-known currency trading strategies

and the foreign exchange volatility explain a significant part of the variability of the

returns of professional currency managers. Thus, exposures to these factors might be a

7
The USD/JPY spot rate fell from levels above 145 to almost 100 during this time period.
8
This rolling correlation could be interpreted as a simple measure of carry crowdedness until the end of
2008. Indeed, the correlation between this measure and our measure of crowdedness from September 2005
until the end of 2008 is 49%. However, it drops to -72% from January 2009 until June 2010, a period in
which the Swiss National Bank (SNB) intervened massively to stem the appreciation of the Swiss franc.
The SNB sold CHF valued at roughly $200bn between 2009 and mid-2010, including $73bn in May 2010
alone. These calculations suggest that the Swiss franc was no longer considered as a funding currency
after 2008, as investors sought safety and preferred to own CHF.
4
useful way to gauge the popularity or crowdedness of a trading strategy.9 We define

style crowdedness as the percentage of the funds with significant positive exposure to a

given style less the percentage of the funds with significant negative exposure to the

same style (contrarians).10 To estimate crowdedness, we rely on data for 107 currency

managers covering a little over five years between April 2005 and June 2010.11 We

estimate style betas using the four factor model proposed in PL (2008a). We use higher

frequency, weekly return data to obtain efficient parameter estimates for rolling 26-week

periods.

Our overall results are quite promising and consistent with a view that crowded trades

harbour potential risk once a change in fundamentals or sentiment induces liquidation of

positions. Our analysis shows that Carry became a crowded strategy in spring 2008, just

a few months before the crash in October/November 2008. And Carry grew increasingly

crowded through 2009, peaking at the beginning of 2010, several months before the

European sovereign debt crises and the “flash crash” in spring 2010 when managers in

our sample became neutral on carry. Value trading exhibits a similar pattern. In spring

2008, Value became a popular contrarian strategy (a high proportion of the funds

exhibited significant negative value exposure) just a few months before performing

9
These factors are specific to the currency funds. Clearly, when measuring crowdedness for different
types of hedge funds, i.e. global macro, a researcher should make use of different factors, for example
those identified in Fung and Hsieh (2002).
10
Alternatively, each manager could be weighted by their assets under management (AUM). In a related
paper, Jylhä and Suominen (2009) find that AUM at hedge funds are significantly related to
contemporaneous and expedited future returns from a risk-adjusted carry trade. Unfortunately, we do not
have data on AUM for the managers in our sample to experiment with this alternative measure.
11
An earlier version of the paper was based only on a 3-year sample, from April 6, 2005 until March 26,
2008. We updated the paper when more data became available. The results post March 26, 2008 can be
interpreted as out-of-sample and highlight the usefulness of our framework to measure crowded trades.
5
exceptionally well. Value again became a popular contrarian strategy by the beginning

of 2010, to be followed by a surge upward in the performance of Value several months

later. Trend trading shows a somewhat different pattern. By the spring of 2008, investors

seemingly “gave up” on trend as a relatively highly level of crowdedness disappeared

only months before the trend trading style delivered big gains in the crisis period. Our

results suggest that style crowdedness varies considerably over time and may impact the

future performance of its respective style.

In the next section of the paper, we lay out our methodology for estimating crowdedness.

In section 3, we report our estimates of crowdedness and present some empirical

evidence on its determinants. Conclusions and implications of our findings are in the

final section.

2. Data Description and Definition of Crowdedness

To measure exposure to styles, we follow the approach used in PL (2008a) and use a

standard factor model of the form:

Rt = α + ∑i β i Fi ,t + ε t (1)

where

R is the excess return generated by the currency manager, defined as the total
return ( Rt* ) less the periodic risk-free rate ( RF , t )

α is a measure of active manager skill,

F is a beta factor, that requires a systematic risk premium in the market,

6
 β is a coefficient or factor loading that measures the sensitivity of the manager’s
returns to the factor, and

ε is a random error term.

To implement this approach, we require data on currency manager returns and factors

that proxy for types of trading strategies and exposures that currency managers would be

likely to utilize.

We make use of the same data base as used in Pojarliev and Levich (2008b), i.e. daily

return data for currency managers listed on the Deutsche Bank FXSelect trading

platform.12 While FXSelect is a new venture, the platform is designed to offer an

attractive means for professional currency managers to enhance their visibility and grow

their client base. As such, we believe that the FXSelect data offer a fair means of

assessing performance in the currency management industry.13 Because investors who

use FXSelect may buy and sell positions continuously, daily prices of funds are available

and allow us to measure crowdedness at shorter intervals. Our sample includes daily data

12
Launched in March 2005, FXSelect is an open platform, which allows clients of Deutsche Bank to
allocate their funds to different currency managers. Any currency manager can apply for registration in the
platform and be accepted if he satisfies the following criteria: a) Managers must be able to provide a daily
track record for at least the last 18 months verified by a third party, b) They cannot have had more than a
20% performance drawdown over the last 12 months, c) Assets under management must be at least 15
million USD, and d) Satisfactory criminal and regulatory searches on key individuals. We are grateful to
Neville Bulgin and Rashid Hoosenally from Deutsche Bank for supplying the data. More information
about FXSelect can be find in the brochure “FXSelect: An Asset Allocation Solution,” Deutsche Bank,
Global Markets Foreign Exchange, 2006.
13
Many (about 25%) of the managers in the FXSelect database are also included in other well known
hedge fund databases (CISDM and TASS). In our initial 3-year sample, the correlation between the
monthly returns on a “fund-of-funds” (FoF) portfolio, compromised of equally weighted positions in each
of the funds available on the platform and the monthly returns on two other well-known currency hedge
fund indices, the Parker FX index and the Barclay Trades Currency Index, is 67% and 65%, respectively.
As another example for the visibility of the platform, in February 2007 Deutsche Bank launched the
Mercer Currency Manager Index – a multimanager product based on managers from the FXSelect
platform chosen by Mercer Investment Consulting. According to information posted on its website,

7
on returns for 107 funds between April 2005 and June 2010.14 Only 10 of these funds

have a complete 63-month track record. However, there are 18 funds with more than 5

years of data, and 48 funds with 3 years or more data. To correct for accounting errors

and eliminate data outliers, we transform the daily returns into 156 weekly returns by

using Wednesday observations.15 The data base is especially useful as it provides us with

high frequency returns and allows for the correction of backfill and survivorship bias.16

Data for Risk Factors

As risk factors we make use of the same proxies as in PL (2008b).

Carry Factor

We use the Deutsche Bank G10 Harvest Index as the proxy for the returns of a carry

strategy. This index reflects the return of being long the 3 high-yielding currencies

against being short the 3 low-yielding currencies within the G10 currency universe. The

index is rebalanced quarterly. Every quarter the currencies are re-ranked according to

their current 3-month Libor rate. The Bloomberg code for this factor is DBHVG10U

Index.

FXSelect has attracted $3.5 billion in AUM from pension funds, fund of funds, private banks, insurance
companies, and other investors.
14
We use the terms “fund” and “manager” interchangeably. A currency management firm could have
multiple funds or programs on the platform.
15
We use Wednesday as fewer bank holidays fall on Wednesday. Managers are based in different
locations (US, UK, Australia, Switzerland, Monaco, Spain, Sweden, Germany, Ireland and Canada).
16
For more information see Pojarliev and Levich (2008b).
8
Trend Factor

As a proxy for the trend-following factor, we use the AFX Currency Management

Index.17 The AFX Index is based on trading in seven currency pairs weighted by their

volume of turnover in the spot market, with returns for each pair based on an equally-

weighted portfolio of three moving average rules (32, 61 and 117 days).18

Value Factor

We use the Deutsche Bank FX PPP Index as the proxy for the returns of a value strategy.

To gauge relative value, Deutsche Bank prepares a ranking based on the average daily

spot rate over the last three months divided by the PPP exchange rate as published

annually by the OECD. The FX PPP index reflects the return of being long the 3

currencies with the highest rank (undervalued currencies) against being short the 3

currencies with the lowest rank (overvalued currencies) within G10 currency universe.

The Bloomberg code for this factor is DBPPPUSF Index.

Currency Volatility Factor

We use the Deutsche Bank Currency Volatility Index as the proxy for foreign exchange

volatility. This index is calculated as the weighted average of 3-month implied volatility

for nine major currency pairs (as provided by the British Bankers Association) with

17
Monthly data for this index are available at the AFX web site (http://www.ljmu.ac.uk/LBS/95327.htm).
We are grateful to Pierre Lequeux from Aviva Investors for providing daily data. We transformed the
daily returns into weekly returns by using the Wednesday observations.
18
The seven currency pairs are EUR-USD, USD-JPY, USD-CHF, GBP-USD, EUR-JPY, EUR-GBP, and
EUR-CHF.
9
weights based on trading volume in the BIS surveys.19 The Bloomberg code for this

factor is CVIX Index. We use the first difference for this factor in equation (1) as it is

not a trading strategy. In the case of the previous three factors, we use returns.

Definition of Crowdedness

We define the crowdedness of style F at time t (CF,t) as the percentage of the funds with

significant positive exposure to style F less the percentage of the funds with significant

negative exposure to the same style (contrarians).

C F , t = a F , t − bF , t (2)

where

aF,t is the percentage of funds with significant positive exposure to risk factor F
over the period t-25 through t, i.e. we use rolling windows of 26 weeks to
estimate the exposures to the risk factors with equation (1).

bF,t is the percentage of funds with significant negative exposure to risk factor F
over the period t-25 through t-25, i.e. we use rolling windows of 26 weeks to
estimate the exposures to the risk factors with equation (1).

For both positive and negative exposures, we use a standard 95% confidence
level and t-value with absolute value greater than or equal to 1.96 to indentify
significant exposure.

By restricting our measure to only those funds with significant style betas, we intend to

exclude funds where the point estimate of exposure while non-zero may not be

meaningful. As a robustness check, in the empirical section we present several other

measures of crowdedness based only on the magnitude of style betas regardless of their

significance.

19
The nine currency pairs are EUR-USD, USD-JPY, USD-CHF, USD-CAD, AUD-USD, GBP-USD,
EUR-JPY, EUR-GBP, and EUR-CHF.
10
3. Empirical Results

a. Time Variation in Crowdedness

To determine which funds have significant exposure to each trading strategy, we

estimate equation (1) using a rolling sample of 26 weekly observations over the 63-

month (274 weeks) sample period, April 6, 2005 – June 30, 2010. Thus, we are able to

estimate crowdedness on 249 dates commencing September 28, 2005 running through

June 30, 2010. Funds on the platform for less than 26 weeks are excluded from our

analysis.20

The number of funds used to estimate crowdedness varies from week to week, as new

funds join the platform and some funds exit the platform. Exhibit 2 plots the number of

funds used to estimate crowdedness. The number of funds is the lowest (22) at the

beginning of the sample. It then rises steadily toward 60 in late 2006 as funds join the

platform and then oscillates between 50 and 60 for the remainder of the sample as funds

list and delist from the platform. Delisting funds tend to outnumber newly listed funds

between January 2007 and spring 2009 when net new listings resume for the remainder

of the sample.

To illustrate the methodology, Exhibit 3 plots the estimated t-statistics for alpha and the

betas for fund #6 (indicating fund #6 in the data base). This fund has a track record of

slightly more than 3 years (170 weeks) from the launch of the trading platform until fund

#6 delisted on June 25, 2008. Using this sample, we obtain t-statistics for 144 weeks,

20
There are seven funds on the platform with a track record of less than 26 weeks.
11
using a rolling window of 26 weeks. Exhibit 3 shows that over the entire sample period,

fund #6 never achieved a significant alpha. Fund #6 generally had positive exposure to

carry and trend and negative exposure to value and volatility.21 However, these

exposures were not consistently significant throughout the entire sample period, i.e. the

t-statistics of the factor loadings were not constantly above 1.96 (or below -1.96).22 For

example, the exposure to value was most of the time not significant, but there were

periods (at the beginning of 2006 and towards the middle of 2007) when this manager

exhibited strong contrarian value positioning, i.e. the t-statistics of the value factor were

between -2 and -3. Thus, manager #6 appears to have discretionary trading authority,

tracking Value at some times and not at others, and taking other positions not

significantly related to the Carry and Trend factors.

Crowdedness

Using t-values from equation (1), we estimate crowdedness using equation (2) for three

of the four factors, i.e. carry, value and trend. As the fourth risk factor does not represent

return on a trading strategy, but simply the first difference of the implied foreign

exchange volatility, we do not estimate crowdedness for volatility.

Carry Crowdedness

Exhibit 4 plots our measure for carry crowdedness between September 28, 2005 and

June 30, 2010. We also plot acarry,t and bcarry,t representing the percentage of the funds

21
This positive carry exposure might explain his delisting from the platform during the period of massive
underperformance of carry trades.

12
with significant positive exposure to Carry and the percentage of the funds with

significant negative exposure to Carry (the contrarians) and include the performance of

the carry strategy.

Exhibit 4 suggests an interesting story. At the beginning of our sample, Carry

Crowdedness was minimal (around 5%) as only about 10% of the funds in our analysis

were significantly exposed to Carry and the “contrarians” were about 5%. As Carry

started to exhibit very strong performance between mid-2006 and mid-2007, the number

of both carry managers and contrarians increased. The first group appeared to be chasing

the good performance of the carry strategy; while the second group was betting that

carry was overdone. As the first group was only slightly larger than the second, Carry

Crowdedness increased steadily to about 15%. In the summer 2007, the contrarians

started to “die-out” as the performance of the carry strategy accelerated.23 As a result

Carry Crowdedness reached a peak at 32% in early April 2008 as the contrarians either

gave up or were forced out of the market. Interestingly, the carry strategy exhibits a

substantial decline just a few months later. While the popular press attributes the

liquidation of the carry trade to the credit crunch and the decline of the equity markets, a

possible reason behind the rapid liquidation of carry trades might be that this strategy

had become crowded. This result is consistent with the “liquidity spiral” story suggested

22
We are referring here to the results of the rolling regressions. PL (2008b) show that manager #6 exhibits
significant positive exposure to carry and trend and no significant exposure to value and volatility over a 3
year period, from April 6, 2005 until March 26, 2008.
23
PL (2008b) show that managers who did not survive had as a group significant negative exposure to
Carry between April 2005 and March 2008. Ironically, although the liquidation of the carry trade might
have hurt carry managers, the strong performance of the carry strategy until the credit crunch was
devastating for managers betting too early on liquidation of carry trades.
13
by Pedersen (2009) and the shrinking hedge fund asset base discussed in Jylhä and

Suominen (2009).

With the Lehman Brothers bankruptcy in September 2008 and the ensuing global

financial crisis, managers unwound carry trades and Carry Crowdedness collapsed. A

flight to quality led managers into relative safe, low interest rate assets. After a decline

of nearly 30%, by spring 2009 the performance of Carry resumed an upward trend and

crowdedness in the carry trade advanced again to 32% in late 2009. Crowdedness

subsided in early 2010 but dropped precipitously (along with performance) after the

“flash crash” in May 2010.

Trend Crowdedness

Exhibit 5 plots our measure for Trend Crowdedness. In contrast to the Carry

Crowdedness, Trend was a relative crowded strategy at the beginning of our sample

period. The percentage of the funds which had significant positive trend exposure was

between 25% and 35%, with only a very small percentage of the funds being

“contrarians”. As most of the currency research in the 1990s (see for example Levich

and Thomas, 1993) advocated trend-following strategies, this is not a surprise. However,

as Trend failed to deliver returns, crowdedness declined to near zero or slightly negative

(contrarian) by May 2008. This change did not result from a rise in the numbers of

contrarians, but rather that trend-followers appeared to be “giving up.” Ironically, the

trend strategy began to deliver excellent performance a few months later in the fall and

winter 2008. Crowdedness in Trend increased in the midst of this favorable

14
performance, before returning to single digit levels by the start of 2009, actually ahead

of a 10% correction in Trend through spring 2009.

Crowdedness in Trend returned reaching 21.6% in November 2009 and after following a

jagged course, returned to near zero or slightly negative at the end of our sample.

Value Crowdedness

Exhibit 6 plots our measure for Value Crowdedness. The pattern is different, but the

main story bears a strong similarity to our interpretation for Carry Crowdedness. The

percentage of the funds which exhibited positive significant exposure to Value was

relatively small and constant around 10%. On the other hand, the percentage of the

contrarians (funds with significant negative value exposure) was rising steadily through

the spring 2008, peaking at 32%. Thus the contrarian value trade became progressively

more crowded reaching 28.3% in April 2008. A few months later in the summer 2008,

the financial crisis intensified and undervalued currencies rose, causing substantial losses

to contrarian value traders, who had crowded into this position. The contrarians closed

down their positions until Value reached a small positive crowdedness level of 7.8% in

May 2009. The performance of Value was stunning with the factor rising from 90 (in the

summer 2008) to over 120 (in the summer 2009).

The contrarian Value trade re-emerged in the summer 2009 and into the fall, peaking at

-24.5% in January 2010 and into March. Concern over Greek external debt in spring

2010 coincided with a flight toward undervalued currencies. The Value trade earned a

15
quick 8% return, leading Value contrarians to exit their positions. Once again,

crowdedness in a trading strategy proved to be unrewarding for those holding the

relatively popular trading position. At the end of the sample, pro-Value and contrarian

traders were small both a small percentage of managers on the platform, and

crowdedness was nearly zero.

b. Determinants of Crowdedness

As we can see from Exhibits 4, 5 and 6, our measure of crowdedness can vary

considerably. For example, Carry Crowdedness varied between a low of -10% to a high

of 32% reached on two dates separated by nearly two years. Trend Crowdedness ranged

between -3% and 34%. And Value Crowdedness varied between about 12% early in the

sample to about -28%. Selected extreme values of crowdedness for each of the trading

strategies are summarized in the first column of Table 1.

In this section we consider the question of what drives crowdedness. There are two

channels which impact the crowdedness of a trading strategy: (1) Through existing

managers adopting or abandoning a strategy, and (2) by managers entering or exiting the

trading platform which determines the number of funds in our data sample. Table 1

summarizes the composition of our universe of managers at the peaks and troughs in

crowdedness for each style.

Carry Crowdedness was at a trough on December 28, 2005. At that point, 41 funds were

active on the platform (only 16 of these funds (40%) survived until the end of our

16
sample).24 Of the 41 funds, 2 had significant carry exposure, 5 were betting against carry

and 34 had no significant carry exposure. Carry Crowdedness reached an interim peak

on April 9, 2008. At that point, 53 funds were active on the platform (with a track record

of at least 26 weeks). From these 53 funds, 28 were active as of December 28, 2005 and

25 were new funds.

Of the new funds, 12 (or 48%) had significant carry exposure. Furthermore, a significant

proportion from the existing funds with no carry exposure (9 funds or 22%) converted to

having positive carry exposure. Thus, only one of the carry managers at the peak of

Carry Crowdedness was a carry manager when Carry Crowdedness was at its low. The

increase in the Carry Crowdedness seems to have been driven by 1) many new funds

with positive carry exposure joining the platform, and 2) a large number of the existing

funds with no carry exposure, adopting a carry style.

In the next cycle, Carry Crowdedness drops to -10.5% on November 5, 2008 reflecting

that 9 funds (out of 57) hold contrarian styles while only 3 fund have positive carry

exposure and fully 45 funds have no exposure. The increase in funds with no exposure,

from 26 to 45 over the interval, is largely by 21 funds that switched their style betas, and

only 4 new funds whose returns also showed a zero style beta.

In important ways, the subsequent cycles of Carry Crowdedness – peaking at 32.1% on

January 13, 2010 and then reaching a trough of 1.6% on June 16, 2010 – mimic the prior

24
This low survivorship rate highlights the importance of including dead funds in the analysis.
17
two descriptions. The rise in crowdedness is the result of a few new funds (3) that follow

carry joining the platform, and a larger number of existing funds (16) switching to a

positive carry strategy. The decline in crowdedness is the result also of a few new funds

(7) with no exposure to carry joining the platform, and a larger number of existing funds

(18) switching to a neutral carry strategy.

Trend Crowdedness emulates many of the same patterns. Trend Crowdedness was at its

peak of 33.9% on December 6, 2006. At that point, 59 funds were active on the platform

(only 23 of these funds (39%) did not survive until the end of our time horizon). Of these

59 funds, 21 had significant trend exposure, 1 was positioned against trend and 37 had

no significant trend exposure. Trend Crowdedness reached an interim low value (-1.6%)

on May 14, 2008. At that point, 56 funds were active on the platform (with a track

record of at least 26 weeks). Of these 56 funds, 40 were active as of December 28, 2005

and 16 were new funds. Of the new funds, 13 (or 81%) had no significant trend

exposure. Furthermore, of the 21 funds with positive trend exposure, 16 funds (or 76%)

exited the trend style, i.e. had no exposure to trend as of the end of the time horizon.

From the 21 managers with trend following exposure at the time of the peak in Trend

Crowdedness, only 1 manager exhibited trend exposure at the time of the low in the

Trend Crowdedness. Thus, the decline in the Trend Crowdedness seems to be driven by

1) new funds joining the platform with no trend exposure, and 2) a large number of the

initial trend-followers “giving up” on trend.

18
In the subsequent cycles of Trend Crowdedness which reaches a peak of 21.6% on

November 4, 2009 and then a trough of 3.4% on June 9, 2010 the changes in

crowdedness seem driven by a similar pattern. The rise (or fall) in crowdedness is the

result of a large number of existing firms switching to follow (or desist following) the

trend strategy, and a relatively smaller number of funds.

Further emulating this pattern, Value Crowdedness was at a peak on January 18, 2006.

At that point, 41 funds were active on the platform and only 16 of these funds (or 39%)

survived until the end of our time horizon. Of these 41 funds, 6 had significant value

exposure, 1 was betting against value and 34 had no significant value exposure. From

the 6 funds with positive value exposure, 4 funds left the value style and 2 funds exited

the platform. None of the value funds as of January 18, 2006 remained positively

exposed to value as of April 9, 2008 when Value Crowdedness reached an interim

through. At that point, 53 funds were active on the platform (with a track record of at

least 26 weeks). From these 53 funds, 28 were active as of December 28, 2005 and 25

were new funds. Of the new funds, eight (32%) had significant negative value exposure

(contrarian). The decline in Value Crowdedness seems to be driven by a) new funds

joining the platform betting against value, and b) a large number of the existing funds

(10 funds) converting to a value contrarian strategy.

In the subsequent cycles of Value Crowdedness which reaches a peak of 7.8.6% on May

20, 2009 and then a trough of -24.5% on January 13, 2010 the changes in crowdedness

seem driven by a similar pattern. The rise (or fall) in crowdedness is again largely the

19
result of a large number of existing firms switching to follow (or desist following) the

trend strategy, and a relatively smaller number of funds.

A general conclusion we can draw from Table 1 is that the change in crowdedness across

the different styles is driven by the change in styles of the existing managers, but also in

the different styles characteristics of the new managers on the platform. What may be

behind these shifts?

In theory, managers should be attracted by expected returns. As expected returns on a

strategy rise, the desired portfolio allocation to that strategy rises also. However,

specifying the formation of expected returns is always problematic. We consider two

possibilities. First, managers could form expected returns based on the logic of each

trading strategy. For carry trades, as the interest rate differential widens, the expected

return (conditional on a given exchange rate change) rises. For value trades, as

deviations from PPP widen, the expected return rises. We found only weak evidence that

crowdedness in Carry and Value responded to expected returns modelled in this way.25

A second possibility is to model the expected returns on a strategy directly as a function

of past returns on that strategy. If managers form expectations in this way, we would

expect to observe herding in the sense that positive returns on a strategy attracts

newcomers, and negative returns on a strategy encourage managers to abandon a

25
For trend, most simple trend following rules provide only an indication of the future trend, and not the
magnitude, of future exchange rate developments. With trend, therefore, we cannot readily test whether
conditions are more or less favourable for managers to shift in or out of this style.
20
strategy. As we measure crowdedness over 26 weeks, we cumulate the performance of

the different strategies also over 26 weeks. Our methodology (of confirming whether or

not a manager is following some strategy) relies on estimating betas, and needs a number

of weeks before we can determine whether or not a manager has shifted his allocations

in response to higher expected returns in any strategy. So there is a lag of 26 weeks

between when returns on a strategy first appear and when we (the researcher) can

identify a statistically significant relationship, or style beta. Therefore, we have to lag the

cumulative past performance of the strategies by 26 weeks to explore the linkage with

our measure of crowdedness.

Table 2 summarizes the correlations between our measure of crowdedness and the

lagged performance of the trading strategies. Panel A contains results over the whole

sample period. Panels B and C show results for the first and second halves of the sample.

Table 2 suggests some herding in the carry strategy: good past performance attracts

newcomers. There is weaker support for herding in the value strategy, but no support in

our sample for herding in the trend strategy.

Several factors may influence this weak evidence. First, not all managers have

discretionary authority to allocate toward a given currency strategy, even when it

appears to be profitable. For example, a fund that specializes in trend following, and has

stated so in an investment mandate, cannot shift and take positions in carry trades even

when they appear likely to generate profits. Only discretionary managers can shift their

trading style in response to a new market environment. So only a small number of

21
managers in the sample have the ability to shift. We should not expect that 100% of all

the managers in any sample will follow carry (for example) when carry is profitable,

because some of those managers are trend followers or value managers, by design or

choice. Second, managers might be constrained in joining the platform; a fund needs 18

months of track record to list on DB FX Select. Therefore, even if a new carry manager

might be keen to join the platform (as he expects future carry returns to be high), he

would have to wait for the appropriate track record before to join. Finally, the past return

of a strategy might not be the best proxy of what managers think regarding the future

expected return of a strategy.

c. Robustness Checks

As a robustness check, we calculate several alternative crowdedness measures that are

based on the difference in the percentages of those funds with betas above a certain

positive cut-off minus those with betas below a certain negative cut-off, regardless of

whether the t-statistics are significant or not. Thus, we define an alternative measure of

crowdedness of style F at time t (C*F,t) as the percentage of the funds with F-beta greater

than X minus the percentage of the funds with F-beta less than -X.

C *F , t = a *F , t − b *F , t (3)

where

a*F,t is the percentage of funds with beta to risk factor F greater than X over the period t-25
through t, i.e. we use rolling windows of 26 weeks to estimate the exposures to the risk factors
with equation (1).

b*F,t is the percentage of funds with beta to risk factor F less than – X over the period t-25
through t-25, i.e. we use rolling windows of 26 weeks to estimate the exposures to the risk factors
with equation (1).
22
Exhibit 7 plots our original measure of crowdedness for Carry, Trend and Value along

with the alternative measure of crowdedness for a 0.50 cut-off (X=0.50). There is a high

correlation between the two measures for Carry and Value, i.e. 78% and 81%,

respectively. For Trend the correlation is smaller at 68%, but still positive and

significant.26 We calculate the alternative measure of crowdedness for different cut-off

values X=0.25, 0.50, 0.75 and 1.0. Table 3 summarizes the correlations between our

original and alternative measures of crowdedness.

The graphs in Exhibit 7 show that both the original and alternative measures of

crowdedness behave quite similarly over our sample. The relationship is very strong for

Carry and Value while there is an apparent break for Trend during parts of 2007. During

January –May 2007, Trend Crowdedness declined from 17% to 9%, while the alternative

measure increased from 19% to 36%. However, for most of the remainder of the sample,

the generally close association between C and C* for Trend reasserts itself.

Overall, the alternative measures (C*) show similar patterns as our original measure (C)

over much of the sample, with only a small number of instances where the measures

move in opposite directions for an extended period.

26
Testing the significance of the correlation coefficient between C and C* shows that all correlations are
highly significant with p-values < 0.00001.
23
4. Policy and Investment Implications

The financial crisis of 2008 highlights the importance of detecting crowded trades due to

the risks they pose to the stability of the financial system and to the global economy.

However, there is a perception that crowded trades are difficult to identify.27 To date, no

single measure to capture the crowdedness of a trade or a trading style has developed.

Using a unique data base of professional foreign exchange manager returns, we propose

and estimate a new measure for style crowdedness. Our measures of crowdedness offer

more perspective on events in currency markets over the financial crisis period. In the

first quarter of 2008, the data show that a higher than usual percentage of the funds were

significantly exposed to Carry, and these funds suffered during the market turbulence in

the last quarter of 2008 when Carry collapsed. Similarly, in the first quarter of 2008 a

high percentage of the funds were significantly betting against Value. However, later in

2008 Value delivered strong performance, resulting in substantial losses for the

contrarians who were caught wrong-footed. The story for Trend is different: Trend was a

crowded strategy at the beginning of our sample period, but this crowdedness led simply

to flat performance for the trend strategy during this period. After managers gave up on

the trend strategy, Trend delivered strong performance, leading to opportunity costs, but

no actual losses.

27
For example, in attempting to measure the extent of carry trade activity, Galati, Heath and McGuire
(2007) analyze various banking and capital flow data. The authors do not offer numerical estimates. They
conclude that “growth in carry trades funded in yen and Swiss francs has probably contributed to increased
activity in these currencies” but that “the available data do not allow for a more refined measurement of
the size of carry trade positions.” And on the same theme, in their analysis of carry trading and currency
movements in 2008, McCauley and McGuire (2009) conclude that “Carry trades always defy
measurement.” It is worth stressing that our approach does not provide a quantitative estimate of the
volume of carry trades outstanding.
24
Updating the sample from March 2008 until June 2010 when data became available,

confirmed our hypotheses. Following a strong performance of carry in 2009, carry

became a crowded strategy once again, only to experience a strong reversal during the

European sovereign debt crises in the spring of 2010. The patterns for Trend and Value

also reinforced the earlier findings.

Our results suggest that our measure of crowdedness deserves closer monitoring. In our

short sample period, the anecdotal evidence shows that crowdedness may provide useful

signals regarding the future performance of a given strategy. While our sample period is

too short for more formal statistical tests, our analysis suggests that there may be an

adverse relationship between crowdedness and style performance, in particular in the

carry and value styles. This will hopefully stimulate some future research on this subject.

As more and more funds attempt to exploit market timing strategies by switching among

trading styles in order to deliver alpha and not simply beta, crowdedness may again

become a significant element of market dynamics. Indeed as US dollar interest rates

remained close to zero during 2009, commentators alleged that a surging US dollar

based carry trade had developed that will have dire consequences once it begins to

unwind.28 This was realized during the European sovereign debt crises in spring 2010

with the US dollar index surging 15% between January and June 2010. Additional data

28
Nouriel Roubini, "Mother of all carry trades faces an inevitable bust," Financial Times, November 1
2009.
25
will allow researchers to track when our measure of crowdedness reveals any unwinding

and whether changes in crowdedness correlate with exchange rate movements.29

Hearings held by the U.S. House Financial Services Committee in 2009 considered

proposals for a “systematic risk regulator” who could take into account, among other

things, that crowded trades elevate the risk to the financial system because crowding is

itself a source of instability. But as some observers have noted, “the sad truth [is] that

crowded trades are difficult for the government to identify.”30 Our methodology may

offer useful insights regarding the popularity of certain trades – in currencies, gold, or

other assets – among hedge funds and provide regulators with another tool for

monitoring markets. Although, we apply our approach to currencies, it could be easily

extended to other asset classes. Further research in this area could be relevant for

investors, managers, and regulators.

29
Currency managers’ returns are usually available on a daily basis to plan sponsors. Thus, some
institutional investors could update and follow our measure of crowdedness on a daily basis.
30
Sebastian Mallaby, "A Risky 'Systemic' Watchdog," Washington Post, March 2, 2009.
26
References

Deutsche Bank. “DB FX Positioning Indices,” May 18, 2009.

Brunnermeier, Markus K., Stefan Nagel and Lasse H. Pedersen. (2008) “Carry Trades
and Currency Crashes,” NBER Working Paper No. 14473, November.

Evans, Martin D. D. and Richard K. Lyons. (2002) “Order Flow and Exchange Rate
Dynamics,” Journal of Political Economy, Vol. 110, no. 1 (Feb): 170-80.

Froot, K. A., and T. Ramadorai. "Currency Returns, Intrinsic Value, and Institutional
Investor Flows." Journal of Finance 60, no. 3 (June 2005): 1535-1566.

Froot, K. A., and T. Ramadorai. "Institutional Portfolio Flows and International

Investments." Review of Financial Studies 21, no. 2 (April 2008): 937-971.

Froot, K. and R. Thaler. 1990. “Anomalies: Foreign Exchange,” Journal of Economic

Perspectives, vol. 4, no. 3 (Summer): 179-92.

Fung, W. and D. Hsieh. 2002. “Asset-Based Style Factors for Hedge Funds,” Financial
Analysts Journal, vol. 58, no. 5 (Sept-Oct): 16-27.

Galati, Gabriele, Alexandra Heath and Patrick McGuire. 2007. “Evidence of Carry Trade
Activity,” BIS Quarterly Review, (September): 27-41.

Gross, W. 2005. “Consistent Alpha Generation through Structure,” Financial Analysts

Journal, vol. 61, no. 5 (Sept-Oct): 40-43.

Jylhä, Petri and M. Suominen. (2009) "Speculative Capital and Currency Carry Trade
Returns," working paper, Helsinki School of Economics.

National Association of College and University Business Officers (2008), 2007

NACUBO Endowment Study, Washington, D.C.

McCauley, Robert and Patrick McGuire. 2009. "Dollar Appreciation in 2008: Safe
Haven, Carry Trades, Dollar Shortage and Overhedging," BIS Quarterly Review,
(December): 85-93.

Meese, Richard A., and Kenneth Rogoff, 1983, “Empirical Exchange Rate Models of the
Seventies: Do They Fit Out of Sample?” Journal of International Economics, vol. 14,
February, pp. 3–24.

Levich, R.M., and L.R. Thomas, III. 1993. “The Significance of Technical Trading-Rule
Profits in the Foreign Exchange Market: A Bootstrap Approach.” Journal of
International Money and Finance, vol. 12, no. 5 (October):451–474.

27
Pedersen, Lasse Heje. (2009) “When Everyone Runs for the Exit,” NBER working
paper, No. 15297, August 2009.

Pojarliev, M. and R.M. Levich. 2008a. “Do Professional Currency Managers Beat the
Benchmark?” Financial Analysts Journal, (Sep/Oct): vol. 64, no. 5.

Pojarliev, M. and R.M. Levich. 2008b. “Trades of the Living Dead: Style Differences,
Style Persistence and Performance of Currency Fund Managers?” NBER working paper
14355, September 2008. (Journal of International Money and Finance, forthcoming)

Rogoff, K. and V. Stavrakeva (2008), “The Continuing Puzzle of Short Horizon

Exchange Rate Forecasting”, NBER Working paper 14071.

28
Exhibit 1: Rolling Yearly Correlation of Returns on Two Cross Rates

Weekly Data, 01/04/1991 – 07/30/2010

80.0%

60.0%

40.0%

20.0%

0.0%
4‐Jan‐91 4‐Jan‐94 4‐Jan‐97 4‐Jan‐00 4‐Jan‐03 4‐Jan‐06 4‐Jan‐09

‐20.0%

‐40.0%

Correlation between GBP/CHF and NZD/JPY

Data source: Bloomberg and authors’ calculations.

Note: Correlation is computed by using a rolling sample of 52 weekly observations. The

first correlation measure is for January 4, 1991. Currency returns are computed from
January 12, 1990 until July 30, 2010.

29
Exhibit 2: Number of Funds on DB FX Select Platform, Number Used to
Estimate Crowdedness, Number Newly Listed and Delisted.

Weekly data: 4/06/2005 - 6/30/2010

90 20

80 18

16
70

14
60
12
50
10
40
8
30
6

20
4

10 2

0 0
4/6/2005 4/6/2006 4/6/2007 4/6/2008 4/6/2009 4/6/2010
No. on Platform (lhs) No. to estimate Crowdedness (lhs)
Newly listed (rhs) Delisted (rhs)

Source: Deutsche Bank and authors’ calculations

Number of active funds on the platform between week t and t-25 that are used to
estimate Crowdedness on week t. For example, 22 funds were active between 4/6/2005
and 09/28/2005. These 22 were used to estimate Crowdedness on 09/28/2005. Funds
with a track record of less than 26 weeks (1/2 year) are not used for estimating of
Crowdedness.

30
Exhibit 3: Estimated t-values for alpha and beta coefficients for manager #6

Rolling regression results for Rt = α + ∑i β i , Fi ,t + ε t where R are the returns of manager

#6; i = Carry, Trend, Value and Volatility; t = 1, …26 weekly observations.

The first regression is estimated with 26 weekly observations from 4/06/2005 until
6/25/2008 (when manager #6 left the platform). The last regression is estimated with 26
weekly observations from 1/02/2008 until 6/25/2008. The sample contains 144 rolling
windows.

5.00

4.00

3.00

2.00

1.00

0.00
28‐Sep‐05 28‐Jan‐06 28‐May‐06 28‐Sep‐06 28‐Jan‐07 28‐May‐07 28‐Sep‐07 28‐Jan‐08 28‐May‐08
‐1.00

‐2.00

‐3.00

‐4.00

Alpha Carry Trend Value Volatility

31
Exhibit 4: Carry Crowdedness

Rolling regression results for R j ,t = α j + ∑i β i , j Fi ,t + ε j ,t for manager j active on the

platform at least from week t-25 onwards. The number of managers varies according to
Exhibit 2.

Carry crowdedness is defined as in Equation #2. The first measure for crowdedness is
estimated as of 9/28/2005 with 26 weekly observations from 4/06/2005 until 9/28/2005.
The last measure of crowdedness is estimated as of 6/30/2010 with 26 weekly
observations from 1/06/2010 until 6/30/2010. The sample contains 249 rolling windows.

45% 120

40%
115

35%
110
30%
105
25%

100
20%

15% 95

10%
90

5%
85
0%
80
-5%

75
-10%

-15% 70
28-Sep-05 29-Mar-06 27-Sep-06 28-Mar-07 26-Sep-07 26-Mar-08 24-Sep-08 25-Mar-09 23-Sep-09 24-Mar-10

Crowdedness Pro Carry Contrarian Performance of Carry

Crowdedness measures are on left-hand scale and Performance measure is on the right-
hand scale.

32
Exhibit 5: Trend Crowdedness

Rolling regression results for R j ,t = α j + ∑i β i , j Fi ,t + ε j ,t for manager j active on the

platform at least from week t-25 onwards. The number of managers varies according to
Exhibit 2.

Trend crowdedness is defined as in Equation #2. The first measure for crowdedness is
estimated as of 9/28/2005 with 26 weekly observations from 4/06/2005 until 9/28/2005.
The last measure of crowdedness is estimated as of 6/30/2010 with 26 weekly
observations from 1/06/2010 until 6/30/2010. The sample contains 249 rolling windows.

40% 115

35%

110
30%

25%

105
20%

15%
100

10%

5%
95

0%

-5% 90
28-Sep-05 29-Mar-06 27-Sep-06 28-Mar-07 26-Sep-07 26-Mar-08 24-Sep-08 25-Mar-09 23-Sep-09 24-Mar-10

Crowdedness Pro Trend Contrarian Performance of Trend

Crowdedness measures are on left-hand scale and Performance measure is on the right-
hand scale.

33
Exhibit 6: Value Crowdedness

Rolling regression results for R j ,t = α j + ∑i β i , j Fi ,t + ε j ,t for manager j active on the

platform at least from week t-25 onwards. The number of managers varies according to
Exhibit 2.

Value crowdedness is defined as in Equation #2. The first measure for crowdedness is
estimated as of 9/28/2005 with 26 weekly observations from 4/06/2005 until 9/28/2005.
The last measure of crowdedness is estimated as of 6/30/2010 with 26 weekly
observations from 1/26/2010 until 6/30/2010. The sample contains 249 rolling windows.

40% 125

30% 120

20% 115

10% 110

0% 105

-10% 100

-20% 95

-30% 90
28-Sep-05 29-Mar-06 27-Sep-06 28-Mar-07 26-Sep-07 26-Mar-08 24-Sep-08 25-Mar-09 23-Sep-09 24-Mar-10

Crowdedness Pro Value Contrarian Performance of Value

Crowdedness measures are on left-hand scale and Performance measure is on the right-
hand scale.

34
 Exhibit 7: Original Crowdedness Measure (C) and Alternative Crowdedness
Measure (C*) for X=0.50

Rolling regression results for R j ,t = α j + ∑β i i, j Fi ,t + ε j ,t for manager j active on the platform at

least from week t-25 onwards. The number of managers varies according to Exhibit 2. Crowdedness
measure (C) is defined as in Equation 2. Details of the calculation are in the notes to Exhibits 4, 5, and 6.
Crowdedness measure (C*) is defined as in Equation 3. We use the identical rolling regression method
based on active managers. C* tracks the number of funds with β>0.5 less the number of funds with β<-
0.5 regardless of whether or not the β coefficients are significant.

Panel A: Carry Crowdedness

40%

30%

20%

10%

0%

-10%

-20%
28-Sep-05 29-Mar-06 27-Sep-06 28-Mar-07 26-Sep-07 26-Mar-08 24-Sep-08 25-Mar-09 23-Sep-09 24-Mar-10

Carry Crowdedness Carry Crowdedness* (0.50)

Panel B: Trend Crowdedness Panel C: Value Crowdedness

50% 20%

40% 10%

30% 0%

20% -10%

10% -20%

0% -30%

-10% -40%
28-Sep-05 29-Mar-06 27-Sep-06 28-Mar-07 26-Sep-07 26-Mar-08 24-Sep-08 25-Mar-09 23-Sep-09 24-Mar-10 28-Sep-05 29-Mar-06 27-Sep-06 28-Mar-07 26-Sep-07 26-Mar-08 24-Sep-08 25-Mar-09 23-Sep-09 24-Mar-10

Trend Crowdedness Trend Crowdedness* (0.50) Value Crowdedness Value Crowdedness* (0.50)

35
Table 1: Characteristics of Funds at High and Low Points of Crowdedness by
Style of Strategy
# funds with
# funds with significant # funds with
# funds on the
Crowdedness significant negative no significant
platform
(date) exposure exposure exposure
(contrarian)
Panel A: Carry
-7.31% 41 2 5 34
(Dec 28, 2005)
32.08% 53 22 5 26
(Apr 9, 2008) (25, 28, NA) * (12, 1, 9) (2, 2, 1) (11, 13, 2)
-10.52% 57 3 9 45
(Nov 5, 2008) (9, 48, NA) (1, 0, 2) (4, 1, 4) (4, 20, 21)
32.08% 53 21 4 28
(Jan 13, 2010) (7, 46, NA) (3, 2, 16) (1, 1, 2) (3, 22, 3)
1.64% 61 7 6 48
(June 16, 2010) (11, 50, NA) (3, 4, 0) (1, 1, 4) (7, 23, 18)
Panel B: Trend
33.9% 59 21 1 37
(Dec 6, 2006)
-1.64% 56 2 3 51
(May 14, 2008) (16, 40, NA) (1, 0, 1) (2, 0, 1) (13, 22, 16)
21.57% 51 13 2 36
(Nov 4, 2009) (10, 41, NA) (3, 0, 10) (0, 0, 2) (7, 27, 2)
-3.39% 59 4 6 49
(June 9, 2010) (12, 47, NA) (1, 0, 3) (1, 0, 5) (10, 24, 15)
Panel C: Value
12.20% 41 6 1 34
(Jan 18, 2006)
-28.30% 53 2 17 34
(April 9, 2008) (25, 28, NA) (1, 0, 1) (8, 0, 9) (16, 16, 2)
7.84% 51 4 0 47
(May 20, 2009) (11, 40, NA) (1, 0, 3) (10, 24, 13)
-24.53% 53 1 14 38
(Jan 13, 2010) (5, 48, NA) (0, 0, 1) (1, 0, 13) (4, 30, 4)
5.00% 60 5 2 53
(May, 19, 2010) (9, 51, NA) (1, 1, 3) (1, 0, 1) (7, 34, 12)
Each triple of numbers (a, b, c) indicates (a) The number of new funds since the previous date in the table
(upper left column); (b) The number of funds with the same style also at the previous date in the table
(upper left column); and (c) The number of existing funds (active also at the previous date in the table)
which switched style. For example, a fund with no significant carry exposure on one date, but with
significant carry exposure on the following date is counted as having switched style.

36
Table 2: Correlations of Crowdedness Measures with Lagged Performance of
Trading Strategies

Crowdedness for each style factor is defined as in Equation #2. The first measure for
crowdedness is estimated as of 9/28/2005 with 26 weekly observations from 4/06/2005
until 9/28/2005. Performance of each trading strategy is measured over the prior 26-
week period. The first measure for lagged performance is for the period 10/13/2004 until
04/06/2005. Entries in the table are the correlation of crowdedness for each style factor
with its own lagged performance.

Sample Periods Carry- Trend- Value-

Crowdedness Crowdedness Crowdedness

Sept 28, 2005 – June 30, 2010 41% -16% 23%

(249 weekly observations)
Feb 13, 2008 – June 30, 2010 47% -6% 29%
(124 weekly observations)
Sept 28, 2005 – Feb 13, 2008 41% 2% 19%
(125 weekly observations)

37
Table 3: Correlation between Original Measure of Crowdedness (C) and
Alternative Measures of Crowdedness (C*) for Alternative Beta
Values

The original measure of crowdedness (C) is defined in equation #2. An alternative

measures of crowdedness (C*) based on the percentage of managers with styles betas
greater than or equal to a given cut-off value is defined in equation #3. Both measures of
crowdedness are estimated on 26-week periods from 4/6/2005 until 6/30/2010. The
sample contains 249 rolling windows.

Carry Trend Value

Crowdedness Crowdedness Crowdedness
Crowdedness* (0.25) 83% 63% 84%
Crowdedness* (0.50) 78% 68% 81%
Crowdedness* (0.75) 56% 68% 71%
Crowdedness* (1.00) 34% 63% 55%

38