Suspicious Patterns in Hedge Fund Returns

and the Risk of Fraud

Nicolas P. B. Bollen
Vanderbilt University

Veronika K. Pool

Downloaded from [Link] by Trina Holloway user on 20 March 2019

Indiana University

Recent cases of hedge fund fraud have caused large losses for investors and have fueled the
debate regarding the ability of regulators to oversee the industry. This article proposes
a set of performance flags, based on suspicious patterns in returns, as indicators of a
heightened risk of fraud. We collect a sample of hedge funds charged with legal or
regulatory violations and find that funds charged with misappropriation, overvaluation,
misrepresentation, or Ponzi schemes trigger the performance flags at a higher frequency
than other funds. (JEL G20, G23)

Recent episodes of fraud in a variety of markets raise important concerns about

the reliability of financial reports when managers have unchecked control of
the inputs. The scandals at Enron, WorldCom, Tyco, and other U.S. companies
in the late 1990s and early 2000s, for example, were caused in large part by
exploitative managerial discretion over accounting choices involving revenue
recognition, mark-to-market valuation, and the use of special-purpose entities
to hide debt. Similarly, many fiascos in the hedge fund industry, including
those at the Bayou Hedge Fund Group, the Lancer Group, and Madoff’s
Ponzi scheme, were enabled by the ability of managers to misreport returns
by inflating portfolio values, or in some cases by outright fabrication, in an
environment of little oversight. The goal of this article is to better understand
the link between managerial incentives, misreported returns, and the incidence
of hedge fund fraud.
We hypothesize that investor preferences for specific properties of hedge fund
returns, combined with performance-based managerial compensation, provide

Bollen is the E. Bronson Ingram Professor of Finance at the Owen Graduate School of Management, Vanderbilt
University. Pool is an Assistant Professor at the Kelley School of Business, Indiana University. The authors thank
seminar participants at American University, the Chinese University of Hong Kong, Hong Kong University,
Jiao Tong University, Rice University, Penn State University, SUNY Buffalo, the University of Oxford, the
University of Warwick, Virginia Tech University, the CFTC, and SIFR. George Aragon, Utpal Bhattacharya,
Stephen Brown, Stephen Dimmock, Bing Liang, Gerry Martin, Jacob Sagi, Randall Thomas, Marno Verbeek,
and Harvey Westbrook also provided useful comments. The Financial Markets Research Center at Vanderbilt
University supplied generous support. Send correspondence to Nicolas P. B. Bollen, Owen Graduate School of
Management, 401 21st Avenue South, Vanderbilt University, Nashville, TN 37203; telephone: (615) 343-5029.
E-mail: [Link]@[Link].

© The Author 2012. Published by Oxford University Press on behalf of The Society for Financial Studies.
All rights reserved. For permissions, please e-mail: [Link]@[Link].
doi:10.1093/rfs/hhs085

[13:45 28/7/2012 [Link]] Page: 2673 2673–2702

 The Review of Financial Studies / v 25 n 9 2012

an incentive for some managers to exploit their discretion over the valuation of a
fund’s portfolio. As a consequence, misreported fund returns likely feature one
or more detectable patterns that are consistent with the desired attributes. Using
this insight, we assess the ability of low-cost quantitative algorithms to identify
a heightened risk of fraud using only a fund’s history of reported returns.
Our approach is motivated by prior research. Previous studies argue that, due
to the structure of managerial incentives, certain outcomes, such as superior
performance around fiscal year-ends, may be especially desirable for hedge
fund managers. For instance, Bollen and Pool (2009) and Agarwal, Daniel,

Downloaded from [Link] by Trina Holloway user on 20 March 2019

and Naik (2011) show that investors direct more capital to funds that feature a
higher number of months with positive returns. Since managerial compensation
is a direct function of the quantity of assets under management, some managers
may misreport returns in order to attract and maintain their investor base. Ponzi
schemes are an extreme example in which the only way for managers to attract
capital is through misreported returns. Gregoriou and Lhabitant (2009) describe
how Fairfield Sentry, a Madoff feeder fund, reported only 10 months of losses
in a 215-month history. Under reasonable assumptions about the distribution of
returns, as we show later, the probability of observing this paucity of negative
returns is unimaginably small. This example illustrates how peculiarities in
hedge fund returns, driven by managerial incentives, can leave a statistical
footprint.
We derive a family of performance flags based on suspicious patterns in
a fund’s reported returns that are consistent with fraudulent activity. These
include: (1) a discontinuity in the distribution of hedge fund returns, (2) three
measures of low correlation between hedge fund returns and the returns of
style factors, (3) serial correlation in hedge fund returns, (4) conditional serial
correlation in hedge fund returns, and (5) a family of six data-quality indicators.
These data-quality indicators include measures such as the percentage of
negative returns, the percentage of repeat returns, and the number of returns
exactly equal to zero. In each case, we design an appropriate statistical test for
significance using a range of analytic and simulation techniques.
To study whether performance flags can help identify cases of hedge fund
fraud, we construct a sample of funds that have been the subject of regulatory
enforcement actions or investor lawsuits, which we label problem funds,
following the terminology of Brown et al. (2008). We classify each problem
fund as either a reporting violation or a trading violation based on the nature of
the alleged offense. In no case is a fund charged with both types. A reporting
violation is defined as misappropriation, overvaluation, misrepresentation, or
running a Ponzi scheme. These actions are deceitful and more likely to be
reflected in patterns in self-reported returns than other offenses. A trading
violation involves short-sale rules, insider trading, or any other violation, none
of which imply that reported returns are inaccurate. We compare problem
funds to all other funds in a broad sample drawn from the TASS and CISDM
databases, which we label non-problem funds. Nine of the 12 performance

2674

[13:45 28/7/2012 [Link]] Page: 2674 2673–2702

 Suspicious Patterns in Hedge Fund Returns and the Risk of Fraud

flags are triggered at a significantly higher rate for reporting violations than for
non-problem funds. Six of the flags, including low correlation with style factor
returns and too few negative returns, have rejection frequencies over 40% in the
sample of reporting violations. In contrast, none of the performance flags are
triggered more often in the sample of trading violations. These results indicate
that our returns-based indicators are naturally more informative for violations
in which managers are attempting to deceive their investors.
We conduct a probit analysis to test whether the performance flags provide
incremental information for identifying the incidence of violations in a

Downloaded from [Link] by Trina Holloway user on 20 March 2019

multivariate setting. Existing studies examine the relation between violations
and measures of hedge fund operational risk, defined in Brown et al. (2008)
as “the risks of failure of the internal operational, control, and accounting
systems; failure of the compliance and internal audit systems; and failure of
personnel oversight systems, that is, employee fraud and misconduct.” Brown
et al. (2008), Brown et al. (2009), and Dimmock and Gerken (forthcoming) find
an association between operational risk and SEC violations; consequently, we
include as one of the independent variables in our probit analysis a proxy for
operational risk based on Brown et al. (2009). We find that several performance
flags, in addition to the operational risk proxy, are statistically and economically
related to the likelihood of regulatory and legal violations. This result indicates
that the information contained in our collection of suspicious patterns in returns
is complementary to the predictive power of operational risk measures.
The coefficient estimates from the probit analysis provide a means for
constructing a single metric (f -score) for each fund that represents the
conditional probability of a subsequent violation. We find that roughly 50%
of the reporting violations feature a conditional probability that exceeds the
90th percentile for non-problem funds. This result illustrates how a regulator or
institutional investor could compute a scalar measure of the risk of subsequent
trouble.
We examine the response of investors to performance in problem funds and
non-problem funds to test whether investors are already aware of the heightened
risk of a subsequent violation. If investors are already able to identify funds with
a higher risk of future legal trouble, through due diligence efforts that focus on
operational risk, for example, we might expect good performance by problem
funds to attract less capital than good performance by non-problem funds.
Our flow-performance analysis shows that investors do not seem to be able to
differentiate between problem and non-problem funds: good performance by
both types triggers a similar flow response. This implies that our performance
flags may represent a new source of information for hedge fund investors
regarding the risk of regulatory or legal violations.
Our article makes two main contributions. First, the hand-collected data set
of enforcement actions and lawsuits against hedge funds is a rich, natural
laboratory for studying the behavior of fund managers. We demonstrate for
the first time a significant statistical association between a wide variety of

2675

[13:45 28/7/2012 [Link]] Page: 2675 2673–2702

 The Review of Financial Studies / v 25 n 9 2012

suspicious patterns in returns and the incidence of fraudulent activity and

other regulatory violations. While prior research documents the existence of
anomalies in reported returns, our study offers a comprehensive analysis of the
frequency with which they precede legal or regulatory trouble. Brown et al.
(2008), Brown et al. (2009), and Dimmock and Gerken (forthcoming) study
violations self-reported on SEC Form ADV. While valuable, this information
is available only for funds registered as investment advisers, and, as noted by
Dimmock and Gerken (forthcoming), is not generally available historically.1
Furthermore, Brown et al. (2012) find evidence that funds filing Form ADV

Downloaded from [Link] by Trina Holloway user on 20 March 2019

indicate prior violations at a much lower frequency than a sample of funds
examined by a due diligence firm. This result suggests that the historical Form
ADV sample may be populated with funds with little to hide.
Second, our success in detecting the risk of fraud has important regulatory
implications. By their nature, performance flags can be checked quickly and
at low cost in large-scale databases using quantitative algorithms. We show
that targeting funds for examination using risk-based screens can be effective,
since funds that trigger performance flags are much more likely to ultimately be
charged with legal or regulatory violations. In the spirit of Becker (1968), anti-
fraud provisions of the securities statutes are a stronger deterrent to unlawful
behavior when the probability of getting caught is higher. Our results therefore
complement the findings of Brown et al. (2008), Brown et al. (2009), and
Dimmock and Gerken (forthcoming), who show that operational flags contain
information related to the risk of fraud; in total, this evidence suggests that
closer supervision of the hedge fund industry is feasible.
Performance flags can help the SEC determine which funds to examine, in
the same way that the IRS uses red flags in tax returns to determine which filers
to audit. Similarly, they can help due diligence firms and institutional investors
decide whether a particular fund poses a heightened risk of fraud. Cox, Thomas,
and Kiku (2003) study determinants of SEC enforcement actions and find little
evidence that objective, quantifiable predictors of fraud are associated with
subsequent actions, suggesting that SEC enforcement activity is triggered by
investor complaints and adverse press coverage. However, the SEC has been
widely criticized for failing to recognize the Madoff Ponzi scheme earlier than
it did, prompting reforms to its enforcement division, including an emphasis
on risk-based screens to trigger examinations.2 Following the passage of the
Dodd-Frank Act, the SEC will collect information from hedge funds and private
equity firms, including a record of monthly returns, thereby permitting the use
of the performance flags we examine.

1 Title IV of the Dodd-Frank Act changes exemption rules for the Investment Advisers Act, including eliminating
the private adviser exemption. Going forward, advisers managing at least $100 million are required to register
with the SEC and file a Form ADV.
2 See [Link]/spotlight/[Link].

2676

[13:45 28/7/2012 [Link]] Page: 2676 2673–2702

 Suspicious Patterns in Hedge Fund Returns and the Risk of Fraud

The rest of the article is organized as follows. In Section 1, we discuss the

set of performance flags used in the study and provide implementation details.
Section 2 describes the data, with special attention paid to the collection of fraud
cases used in our study. Section 3 reports the frequency at which performance
flags are triggered, and measures the incremental information they provide for
identifying problem funds. Section 4 offers concluding remarks.

1. Quantitative Screens for Hedge Fund Fraud

Downloaded from [Link] by Trina Holloway user on 20 March 2019

This section describes the five categories of performance flags we use to indicate
a heightened risk of fraud. Details of the statistical tests we use to determine
whether the flag is triggered are available in an Online Appendix.

1.1 Discontinuity at zero

Hedge funds are sometimes described as absolute-return vehicles that strive
to deliver positive returns regardless of market conditions, as in Waring and
Siegel (2006). Investors appear to value this property. As shown by Bollen
and Pool (2009) and Agarwal, Daniel, and Naik (2011), hedge fund investors
direct capital toward managers who have reported a higher number of positive
monthly returns, even after controlling for the level of cumulative return. If
managers are able to deliver mostly positive returns through skill, then counting
the number of losses would be a reasonable way to identify ability. However,
evidence suggests that loss-avoidance might be evidence of misreporting.
Bollen and Pool (2009) show that the distribution of hedge fund returns is
discontinuous at zero. The discontinuity disappears when returns are computed
at the bimonthly frequency instead, indicating that some returns are inflated
to avoid reporting losses, and that these overstatements are subsequently
reversed. Thus, our first performance flag, which we label Kink, is triggered
when the distribution of a hedge fund’s reported returns features a significant
discontinuity at zero.
The statistical test for a discontinuity in Bollen and Pool (2009) is data-
intensive and infeasible for testing individual funds. Instead, we adopt the
histogram approach of Burgstahler and Dichev (1997). We count the number
of return observations that fall in three adjacent bins, two to the left of zero and
one to the right. Under the null hypothesis of a smooth distribution, the number
of observations that fall in the middle-bin should be approximately equal to
the average of the surrounding two bins. We interpret a significant shortfall in
the middle-bin observations as evidence that some negative returns have been
purposefully inflated above zero.

1.2 Low correlation with other assets

One purported benefit of hedge fund investment is diversification due to a low
correlation with standard asset classes. Furthermore, as argued by Sun, Wang,
and Zheng (2012), managers with skill and informational advantages likely

2677

[13:45 28/7/2012 [Link]] Page: 2677 2673–2702

 The Review of Financial Studies / v 25 n 9 2012

pursue profitable trading strategies that are different from those followed by
other managers; hence, low correlation with funds in the same category could
be a predictor of abnormal performance. Indeed, Titman and Tiu (2011) show
that hedge funds in the lowest quartile, as ranked by the adjusted R-squared
of factor model regressions, have the highest subsequent Sharpe ratios. Note,
however, that if a manager misreports returns, his fund may also feature low
correlation with standard asset classes, hedge fund style factors, and even funds
in the same category. For example, as described by Harry Markopolos in his
2009 testimony to the U.S. House of Representatives, a manager such as Madoff

Downloaded from [Link] by Trina Holloway user on 20 March 2019

who reports a random positive return every month will artificially generate a
correlation between the fund’s returns and any other time series very close to
zero. Indeed, Madoff’s returns had a correlation of only 0.06 with the S&P 500,
whereas his supposed split-strike conversion strategy should have featured a
correlation close to 0.50. Thus, we use low correlation with a set of style factors
as an indicator of fraud.
We measure the uniqueness of a fund’s return series three ways. First, we
regress fund returns on the subset of hedge fund style factors that maximizes
the regression’s adjusted R-squared, which we label Maxrsq. Then we assess
whether the Maxrsq is significantly different from zero using a bootstrap
simulation. The Maxrsq flag is triggered if a fund’s Maxrsq is smaller than
the 90th-percentile using the randomly generated data. Second, we repeat the
R-squared test but allow the factor loadings to switch once during each fund’s
history, adopting the methodology of Bollen and Whaley (2009), who argue that
fund-specific changes in factor loadings can significantly boost a regression’s
explanatory power. We label a fund’s R-squared with switches in factor loadings
the Switchrsq, with an associated flag that is triggered if a fund’s Switchrsq is
smaller than the 90th-percentile critical value. Third, in the spirit of Sun, Wang,
and Zheng (2012), we avoid the selection of appropriate factors and instead
compare a fund to the other funds in the same reported style. For each style, we
create an equally weighted index using all funds in the style in a given month.
We then regress a fund’s returns on the appropriate index, after modifying
the index to eliminate the returns of the fund in question. A fund triggers the
Indexrsq flag if the index slope coefficient is statistically insignificant at the
10% level.
Some fund managers may trigger one or more of our tests for low correlation
by delivering on what they promise: an alternative investment. If so, then
triggering one of these flags could be viewed as generating a false positive (i.e., a
false indicator of the risk of fraud). The extent of false positives will be measured
when we compare the rejection rates of problem funds and non-problem funds.

1.3 Unconditional serial correlation

Getmansky, Lo, and Makarov (2004) argue that when hedge funds are invested
in illiquid securities, returns can feature artificially low volatility and positive
serial correlation. Funds invested in more illiquid securities, such as emerging

2678

[13:45 28/7/2012 [Link]] Page: 2678 2673–2702

 Suspicious Patterns in Hedge Fund Returns and the Risk of Fraud

market debt, feature higher serial correlation than other funds. One explanation
for this result is that when market prices are not available, fund returns reflect
new information gradually over time as managers conservatively update model
values when computing an NAV. Another explanation is that some managers
purposely smooth returns by reporting moving averages of current and lagged
portfolio returns. Working (1960) shows how moving averages feature lower
volatility than raw observations, and will possess serial correlation even when
raw observations have none.
We regress fund returns on their first lag to test for unconditional serial

Downloaded from [Link] by Trina Holloway user on 20 March 2019

correlation as follows:
RtO = a +bRt−1O
+εt , (1)
where RtO , a fund’s observed return at date t, is potentially different from Rt ,
the actual return of the fund. We label this the AR(1) flag. The AR(1) flag is
triggered if the b coefficient is positive and significant at the 10% level.

1.4 Conditional serial correlation

The ability of managers to misreport by smoothing is increasing in the illiquidity
of the assets they hold, since the opportunity to exercise discretion exists
only when recent trade prices are not available. Thus, Getmansky, Lo, and
Makarov (2004) are careful to point out that since marking-to-model and
deliberate smoothing generate identical time-series properties, it is difficult
to make a statement about a manager’s intent without additional information.
Asness, Krail, and Liew (2001) and Bollen and Pool (2008) suggest additional
econometric techniques that attempt to distinguish innocuous behavior from
purposeful misreporting. We adopt the procedure in Bollen and Pool (2008),
which is based on the premise that managers have an incentive to smooth losses
to delay reporting poor performance, and an incentive to fully report gains in
their competition for investor capital. As a consequence, the amount of serial
correlation varies conditional on the magnitude of lagged returns.
When testing for conditional serial correlation, the distinction between a
fund’s observed return RtO on date t and the actual return of the fund’s portfolio
Rt is again important. We make the assumption that the degree of smoothing,
and hence serial correlation, is a function of the actual lagged return, Rt−1 . Of
course, the actual return of a fund is unobservable, hence we proxy for it by
using the fitted value of the optimal factor model constructed in the Maxrsq test.
The fitted values can be interpreted as the component of fund returns generated
by exposure to liquid assets. As in Bollen and Pool (2008), we regress observed
returns on their lag with an interaction term as follows:

RtO = a +b+ Rt−1

O
+b− (1−It−1 )Rt−1
O
+εt , (2)

where It−1 = 1 if the fitted value of observed returns in month t −1 is greater

than its mean and zero otherwise. Thus, b− measures the incremental serial
correlation following poor returns; a positive b− coefficient means that serial

2679

[13:45 28/7/2012 [Link]] Page: 2679 2673–2702

 The Review of Financial Studies / v 25 n 9 2012

correlation is higher, which is consistent with an avoidance of reporting losses.

We label this the CAR(1) flag. The CAR(1) flag is triggered if the b− coefficient
is positive and significant at the 10% level.

1.5 Data quality

Straumann (2008) investigates the quality of hedge fund return data, in the
spirit of Liang (2003), without taking a strong stand on the motives behind any
“man-made” patterns that are revealed by his analysis. He examines each fund’s
return history and attempts to detect five patterns: (1) too many returns exactly

Downloaded from [Link] by Trina Holloway user on 20 March 2019

equal to zero, (2) too few unique returns, (3) too long a string of identical
returns, (4) too many recurring blocks of length two, and (5) a distribution
of the last digit that rejects the null of uniform. The fund’s quality score is
then recorded as the number of patterns present, with a higher score indicating
lower quality. In addition to these five metrics, we include in our study of data
quality the percentage of fund returns that are negative, since if returns are
simply fabrications, as in Ponzi schemes, then managers naturally will report
few losses. While these patterns are not necessarily cause for concern, their
presence could be an indicator of more harmful actions.
Critical values for three of the six data-quality tests, the uniformity of the last
digit, the number of returns equal to zero, and the percentage of negative returns,
are obtainable analytically. As described by Straumann (2008), the uniformity
test compares the percentage of observations ending in each digit 0 through
9 to its expected value of 10% under the null of a uniform distribution, and
aggregates these differences in a goodness-of-fit test statistic. A fund triggers
the Uniform flag when the probability of a more extreme difference from the
uniform distribution is less than 10%. A fund triggers the # Zero flag when
the probability of generating the observed number of zero returns or more is
less than 10%. A fund triggers the % Negative flag when the probability of
generating the observed number of negative returns or fewer is less than 10%.
Obtaining critical values for the other tests, the percentage of repeated returns,
the maximum length of a repeated return string, and the number of pairs of
repeated returns, is challenging, especially given the impact of rounding, and
so we establish fund-specific critical values through Monte Carlo simulation.
To develop intuition for the magnitudes of these tests, Table 1 shows two
sets of percentiles of the six test statistics from 10,000 simulations. Panel A
shows results for a 60-month history, which is the median history length of the
funds in our sample. Returns are generated from a scaled t-distribution with
excess kurtosis of 3, annual mean of 10%, and volatility 15%, and reported to
the nearest 0.0001%. Panel B shows the results for a 120-month history, with
returns that are generated from a scaled t-distribution with excess kurtosis of
3, annual mean of 5%, and volatility of 10%, and reported to the nearest 0.1%.
In Panel A, the high precision of the reported returns, along with the assumed
mean and volatility, imply an extremely low probability of observing a return
exactly equal to zero, hence there is less than a 5% chance of observing any

2680

[13:45 28/7/2012 [Link]] Page: 2680 2673–2702

 Suspicious Patterns in Hedge Fund Returns and the Risk of Fraud

Table 1
Critical values

Percentile

Flag 5% 10% 25% 50% 75% 90% 95%

Panel A: 60-Month History

# Zero 0 0 0 0 0 0 0
% Repeats 0% 0% 0% 0% 0% 0% 0%
Uniform 3.33 4.17 5.90 8.34 11.39 14.68 16.92
String 1 1 1 1 1 1 1

Downloaded from [Link] by Trina Holloway user on 20 March 2019

# Pairs 0 0 0 0 0 0 0
% Negative 32% 33% 37% 42% 45% 50% 52%

Panel B: 120-Month History

# Zero 0 0 1 2 3 4 4
% Repeats 36% 38% 39% 42% 43% 45% 47%
Uniform 3.33 4.17 5.90 8.34 11.39 14.68 16.92
String 1 1 1 2 2 2 2
# Pairs 0 0 0 1 2 3 3
% Negative 35% 37% 39% 43% 46% 48% 50%
Listed are percentiles of six tests for data quality: # Zero is the number of returns exactly equal to zero, % Repeat
is the percentage of returns that are repeated at least once, Uniform is a chi-squared statistic testing whether
the last reported digit is uniformly distributed, String is the maximum length of a sequence of identical returns,
# Pairs is the number of pairs of repeated returns, and % Negative is the percentage of returns below zero. Panel
A shows results for a 60-month history, with returns that are generated from a scaled t -distribution with excess
kurtosis of 3, annual mean of 10%, and volatility 15%, and reported to the nearest 0.0001%. Panel B shows
results for a 120-month history, with returns that are generated from a scaled t -distribution with excess kurtosis
of 3, annual mean of 5%, and volatility of 10%, and reported to the nearest 0.1%. For # Zero, Uniform, and %
Negative, the percentiles are computed analytically. For % Repeat, String, and # Pairs, bootstrap percentiles are
established by simulating 10,000 series of returns, which are then rounded to the appropriate precision.

zeros in a 60-month history. In Panel B, the extra rounding, and changes to

the mean and volatility, raise the probability of a zero return so that there is a
75% chance of observing at least one zero in a 120-month history. Similarly,
the maximum string length and expected number of pairs of repeated returns
are larger in Panel B. Note that in both panels there is only about a 5% chance
of observing less than one-third negative returns.3

2. Data
The hedge fund data used in our analysis are drawn from the Center for
International Securities and Derivatives Markets (CISDM) and the Lipper
TASS (TASS) databases. The sample period is from January 1994 through
December 2008. Returns are net of all management and performance-based
fees. Both databases include live and defunct hedge funds, funds of funds, and

3 As mentioned previously, Gregoriou and Lhabitant (2009) describe how a Madoff feeder fund, Fairfield Sentry,
reported only 10 months of losses in a 215-month history. Using the fund’s sample mean and standard deviation,
the probability of observing 10 or fewer negative returns is approximately 0.03%. Fairfield Sentry’s reported
returns have a monthly mean of 0.9% and a monthly volatility of 0.8%, which seems impossibly low. If we use
instead a monthly volatility of 4.0%, which is approximately the average volatility in our sample, the probability
of 10 or fewer negative returns in a series of 215 months is 6.5×10−33 %.

2681

[13:45 28/7/2012 [Link]] Page: 2681 2673–2702

 The Review of Financial Studies / v 25 n 9 2012

commodity trading advisers. We focus on live and defunct individual hedge

funds, since these constitute the vast majority of enforcement actions. A fund
must have at least 24 contiguous monthly observations of returns that overlap
with the sample period for the factor returns, January 1994 through September
2008, to be included in our analysis.4 Some fund managers report returns to
both databases, so we check for matches and delete duplicates. First, potential
matched pairs are formed by examining the names of funds in the two databases.
Since these potential matched pairs have names that are often not identical,
though very similar, we then compute the correlation between the return series

Downloaded from [Link] by Trina Holloway user on 20 March 2019

of each pair of funds. Pairs with correlation above 99% are deemed matches,
and the series with the shorter history is discarded.
We collect a set of 13 style factors that are used in the existing hedge fund
literature to proxy for the trading strategies employed by hedge fund managers.
These factors are drawn from four sources. The excess return of the market and
the returns of size, value, and momentum portfolios are from Kenneth French’s
website. Five trend-following factors, which are the returns of portfolios of
options on bonds, foreign currencies, commodities, short-term interest rates,
and stock indexes, are obtained from David Hsieh’s website. The change in the
yield of a ten-year Treasury note and the change in the credit spread, defined as
the yield on ten-year BAA corporate bonds less the yield of a ten-year Treasury
note, are obtained from the U.S. Federal Reserve’s website. The Agarwal and
Naik (2004) OTM index call and OTM index put factors, designed to capture
non-linear exposure to equities generated by derivatives or dynamic trading,
were provided by Vikas Agarwal. To estimate the factor model, we also require
a risk-free rate, and for this we use the one-month T-Bill rate from Kenneth
French’s website.
Our primary research question is whether funds with return series that trigger
performance flags are more likely to be subsequently prosecuted by the SEC or
the subject of investor lawsuits.5 We hand-collect information regarding claims
of regulatory and legal violations from three types of sources.
First, we search the litigation section of the SEC’s website using the
keyword “hedge fund.” The search uncovers 785 SEC documents, including
administrative proceedings, litigation releases, and court complaints. From
these, we construct a sample of 167 unique prosecution cases involving
477 unique hedge funds. We validate this sample using a database of SEC
examinations from a consultancy. The consultancy’s database complements our
website search because in a significant number of the cases the SEC document
does not explicitly identify the fund as a hedge fund.

4 We also discard 16 TASS funds and 7 CISDM funds with at least one monthly return observation greater than
200%, likely the result of database errors.
5 We do not distinguish between successful and unsuccessful prosecutions. This means fund managers who were
charged, truly guilty, but not convicted remain in our sample, in addition to fund managers who were innocent
and wrongly charged. The latter group adds noise to our analysis and should bias toward a non-result.

2682

[13:45 28/7/2012 [Link]] Page: 2682 2673–2702

 Suspicious Patterns in Hedge Fund Returns and the Risk of Fraud

Second, we search the websites of the Department of Justice and the

Commodity Futures Trading Commission (CFTC) for additional releases on
regulatory violations by hedge funds, the majority of which overlap with the
cases already examined. We also use the Stanford Law School’s Securities
Class Action Lawsuit Clearinghouse website to collect information on investor
lawsuits against hedge funds. Our search sets “industry classification” to
“financial.” We manually check the resulting 562 matches and find 14 new
cases, several of which are lawsuits targeting funds associated with the Bernard
Madoff scandal. Although the managers of these funds are not directly involved

Downloaded from [Link] by Trina Holloway user on 20 March 2019

in fraudulent activities, the performance of these funds is directly connected
to Madoff’s Ponzi scheme. Therefore, we include them in our database. To
complete our sample of Madoff-related funds, we use the list of Madoff victims
assembled by The Wall Street Journal and find six additional cases involving
16 unique funds.
Third, we obtained a list of over 330 unique fraud cases involving over 528
unique funds from a leading due diligence firm. Most of these cases are captured
in our sample of SEC and CFTC documents and, as a result, overlap with our
previous list.
After merging the hedge fund cases, our final list contains 1,069 unique
fund names. These funds are associated with a wide range of fund problems,
from relatively small regulatory violations, such as a failure to file proper
documentation with the SEC, to much more serious offenses, such as
misappropriation or a Ponzi scheme. We manually match our list of problem
funds to our CISDM/TASS database, resulting in 340 matches. We were able
to identify filing dates for 317 of the 340 cases. Filing years range from 1997
through 2009, with about 75% occurring in the last three years of the sample.
The number increases over time, consistent with the growth of the industry
and an increased emphasis on enforcement at the SEC. In addition, anecdotal
evidence suggests that frauds are more likely to be discovered during market
downturns such as the financial crisis at the end of our sample period. Cases
involving overvaluation or Ponzi schemes, for example, are likely to fall apart
following an increase in withdrawal requests, since at that time it becomes clear
that assets in the fund have been expropriated or valuations have been inflated.
For each fund in our sample of 340 funds in the CISDM/TASS data that are
the subject of legal or regulatory proceedings, we discard all returns observed
after the filing date of proceedings against them, since managers may change
their reporting behavior once proceedings have commenced.6 We require 24
remaining observations, and 191 have sufficient return data to be in our final
sample. We label these problem funds. The remaining 8,575 funds in our return
database are labeled non-problem funds.

6 For robustness, we also conduct our analyses discarding observations that occur one year prior to the filing date
or after. This procedure results in additional funds being dropped from the sample since we are left with fewer
than 24 observations. However, our main findings are qualitatively unchanged.

2683

[13:45 28/7/2012 [Link]] Page: 2683 2673–2702

 The Review of Financial Studies / v 25 n 9 2012

Table 2
Violation types
Cases in Cases in
Violation All Cases Sample All Cases Sample

Misappropriation 31 19 7% 7%
Overvaluation 81 52 18% 20%
Misrepresent 72 34 16% 13%
Ponzi 82 24 19% 9%
Short-sale Violation 58 34 13% 13%
Insider Trading 38 29 9% 11%
Other 79 62 18% 24%

Downloaded from [Link] by Trina Holloway user on 20 March 2019

Total 441 254 100% 100%
Listed are the number of SEC enforcement actions and lawsuits that involve each type of violation. A case can
involve more than one violation. All Cases includes the 340 funds named in the SEC enforcement actions and
lawsuits collected for our analysis that also appear in the CISDM/TASS hedge fund databases. Cases in Sample
includes a subset of 191 funds that have at least 24 months of contiguous returns.

For the majority of problem funds, the litigation releases provide sufficient
detail, including a timeline, for a more granular categorization. Based on
the description of events, we place the problem funds into seven categories:
misappropriation, overvaluation, misrepresentation, Ponzi scheme, short-sale
violation, insider trading, and other. We place funds in the other category if
they do not fit any of the first six groups. Though an individual case often
involves several different offenses, we place each fund in a single category that
best captures the primary reason for the charges, which generally corresponds
to the most egregious offense. Not all of the violations involve fraud; for
instance, some of our problem funds are charged with illegal short-selling or
trading on inside information. For these reasons, we further separate our sample
into two groups: funds that are charged with misappropriation, overvaluation,
misrepresentation, or a Ponzi scheme, which we refer to as reporting violations,
and funds that are charged with only short-sale violations, insider trading, or
any other offense, which we refer to as trading violations. One might expect
that our performance flags will more successfully identify reporting violations,
as these are related to the accuracy of reported returns.
Table 2 lists the number of fraud cases, and associated funds in our
final sample, which fall into each category. The totals exceed 340 and 191,
respectively, because some funds are charged with multiple offenses. The
proportion of cases in each category is similar in the full sample and the
subset with sufficient return data, with the exception of Ponzi schemes. These
constitute 19% of the full sample but only 9% of the funds with return data.
As shown in Getmansky, Lo, and Makarov (2004), the returns of illiquid fund
types often feature different time-series properties than the returns of liquid
fund types, including positive serial correlation. This result suggests that the
opportunity for fraud may also vary across fund types; hence, we extract fund
strategies from the CISDM and TASS databases. Table 3 lists 13 categories
of styles created from the two databases, as well as the number of funds in
each category, for non-problem funds, problem funds, reporting violations,
and trading violations. There are problem funds in almost all styles, in rough

2684

[13:45 28/7/2012 [Link]] Page: 2684 2673–2702

 Suspicious Patterns in Hedge Fund Returns and the Risk of Fraud

Table 3
Hedge fund styles
Non-Problem Problem Reporting Trading
Strategy Funds Funds Violations Violations

Equity Long/Short 3,764 92 27 65

Multi-Strategy 856 32 14 18
Event Driven 754 7 3 4
Emerging Markets 679 1 0 1
Equity Market Neutral 594 16 11 5
Fixed Income Arbitrage 558 15 14 1
Global Macro 527 8 2 6
Convertible Arbitrage 327 7 3 4

Downloaded from [Link] by Trina Holloway user on 20 March 2019

Sector 205 8 4 4
Other 105 2 1 1
Distressed Securities 89 1 0 1
Single Strategy 60 0 0 0
Short Bias 57 2 0 2
Total 8,575 191 79 112
Listed are the names of 13 hedge fund styles describing the strategies of the hedge funds in our sample. The list is
a combination of the categories used in the TASS and CISDM hedge fund databases. Also listed, by style, are the
number of hedge funds in our sample that are not the subject of SEC enforcement actions or lawsuits, labeled non-
problem funds, the number of hedge funds that are the subject of SEC enforcement actions or lawsuits, labeled
problem funds, the number of those that are charged with reporting violations, and the number that are charged
solely with trading violations. Reporting violations include misappropriation, overvaluation, misrepresentation,
and Ponzi schemes. Trading violations include violating short-sale rules, insider trading, or any other offense.

proportion to the size of each style. Note that more than half of the trading
violations are in the Equity Long/Short category, but only one-third of the
reporting violations are, consistent with the lack of discretion in valuation for
equity-based funds. In contrast, of the 15 problem funds in the Fixed Income
Arbitrage category, 14 are reporting violations.
Table 4 lists summary statistics of the 191 problem funds (Panel A) and
the 8,575 non-problem funds (Panel B). Listed are the number of funds and
equally weighted cross-sectional averages of each fund’s monthly average
return, standard deviation, Sharpe ratio, skewness, excess kurtosis, expected
shortfall, and first-order serial correlation. Also listed are the number of funds
with positive serial correlation coefficients significant at the 5% two-sided level.
In both Panel A and Panel B in Table 4, live funds feature substantially
lower Sharpe ratios than defunct funds. Defunct problem funds have average
Sharpe ratios of 0.5901 versus 0.3091 for live funds; defunct non-problem
funds have average Sharpe ratios of 0.1688 versus 0.0964 for live funds. This
is at first surprising, as evidence suggests that many hedge funds ultimately
shut down following periods of poor performance, so that defunct funds have
lower returns. Liang and Park (2010), for example, find that prior-year return
is negatively related to the probability of fund failure. Note though that in our
sample the live funds’statistics are heavily influenced by their poor performance
in 2008, whereas most defunct funds had left the databases by then. Given the
historic growth of the industry, any sample of live hedge funds will be largely
populated by new funds; hence, the average return of live funds will be heavily
influenced by the most recent period.

2685

[13:45 28/7/2012 [Link]] Page: 2685 2673–2702

[13:45 28/7/2012 [Link]]

Downloaded from [Link] by Trina Holloway user on 20 March 2019

Table 4
Summary statistics
Type No. of funds μ σ SR Skew Kurt ES AR(1) % >>0
2686

The Review of Financial Studies / v 25 n 9 2012

Panel A: Problem Funds

Live 50 0.0116 0.0337 0.3091 0.5929 4.8155 −0.0567 0.1400 0.2200

Defunct 141 0.0098 0.0396 0.5901 −0.8430 13.2612 −0.0868 0.1520 0.2411
All 191 0.0102 0.0381 0.5165 −0.4671 11.0503 −0.0789 0.1489 0.2356

Panel B: Non-Problem Funds

Live 3,967 0.0052∗∗∗ 0.0398 0.0964∗∗∗ −0.6697∗∗∗ 5.1427 −0.0979∗∗∗ 0.2166∗∗ 0.4495∗∗∗
Defunct 4,608 0.0073∗∗ 0.0419 0.1688∗∗∗ −0.2045∗∗ 3.9661∗∗∗ −0.0907 0.1433 0.2554
All 8,575 0.0064∗∗∗ 0.0409 0.1353∗∗∗ −0.4197 4.5104∗∗∗ −0.0941∗ 0.1772 0.3452∗∗∗

Panel C: Reporting Violations

Live 8 0.0204 0.0377 0.5610 1.0723 12.0918 −0.0400 0.1773 0.5000

Defunct 71 0.0085 0.0417 0.8929 −2.0368 23.2371 −0.1039 0.1805 0.2394
All 79 0.0097 0.0413 0.8593 −1.7220 22.1084 −0.0975 0.1802 0.2658

Panel D: Control Group for Reporting Violations

Live 75 0.0072 0.0367 0.1374∗∗∗ −0.4252 7.2138 −0.0850∗ 0.2620 0.5600

Defunct 560 0.0084 0.0446 0.1915∗∗ −0.2560∗∗∗ 3.9253∗∗∗ −0.0951 0.1705 0.2911∗
All 635 0.0083 0.0437 0.1851∗∗∗ −0.2760∗∗∗ 4.3137∗∗∗ −0.0939 0.1813 0.3228∗∗

Panel E: Trading Violations

Live 42 0.0099 0.0329 0.2611 0.5016 3.4295 −0.0599 0.1329 0.1667

Defunct 70 0.0111 0.0375 0.2829 0.3678 3.1429 −0.0695 0.1231 0.2429
All 112 0.0106 0.0358 0.2748 0.4180 3.2504 −0.0659 0.1268 0.2143

Panel F: Control Group for Trading Violations

Live 250 0.0056∗∗∗ 0.0404∗∗∗ 0.0648∗∗∗ −0.4759∗∗∗ 2.9520 −0.0931∗∗∗ 0.2232∗∗∗ 0.4800∗∗∗
0.0085∗∗ 0.1885∗∗∗ −0.1094∗∗∗ −0.0879∗ 0.1697∗∗ 0.3185∗∗
Page: 2686

Defunct 518 0.0425 3.0872

All 768 0.0075∗∗∗ 0.0418∗∗ 0.1483∗∗∗ −0.2287∗∗∗ 3.0432 −0.0896∗∗∗ 0.1871∗∗∗ 0.3711∗∗∗
Listed are summary statistics of hedge funds in our sample. The summary statistics are the equally weighted cross-sectional averages of the mean monthly return, μ; the standard deviation
of monthly returns, σ ; the Sharpe ratio, SR; the skewness, Skew; the excess kurtosis, Kurt; the expected short fall based on a 5% threshold, ES; the first-order serial correlation coefficient,
2673–2702

AR(1); and the percentage of funds with a positive and statistically significant first-order serial correlation coefficient, %>0. Panel A shows results for the subsample of funds that are the
subject of SEC enforcement actions or lawsuits. Panel B, shows the results for all other funds. In Panel B, the symbols “***,” “**,” and “*” indicate significance at the 1%, 5%, and 10%
levels, respectively, for a test that the corresponding averages in Panels A and B are different. Panels C and E show results for reporting violations and trading violations, respectively, with
results from their control groups in Panels D and F. In Panels D and F, the symbols “***,” “**,” and “*” indicate a significance difference between the violations and their control group.
 Suspicious Patterns in Hedge Fund Returns and the Risk of Fraud

More importantly, problem funds feature substantially higher average returns

and Sharpe ratios than other funds. Overall, the problem funds deliver 1.02%
per month versus 0.64% for all other funds. Sharpe ratios tell the same story,
with the problem funds averaging 0.5165 versus 0.1353 for all other funds.
The differences are statistically significant using a standard two-sample t-test,
as indicated by the asterisks in Panel B in Table 4. These results are perhaps
no surprise, as many problem funds are charged with overstating performance.
All categories feature substantial excess kurtosis, consistent with option-like
payoffs in the strategies they employ, motivating Fung and Hsieh (2001, 2002,

Downloaded from [Link] by Trina Holloway user on 20 March 2019

2004) to use baskets of traded options to mimic strategy returns. However,
the defunct problem funds have dramatically higher excess kurtosis than the
defunct non-problem funds, 13.2612 versus 3.9661. Furthermore, while live
problem funds feature positive skewness of 0.5929 versus –0.6697 for live non-
problem funds, defunct problem funds feature negative skewness of –0.8430
versus –0.2045 for defunct non-problem funds. These results suggest that when
problem funds become defunct, they tend to feature large, negative returns
relative to all other funds.
We also compute expected shortfall, defined as the average monthly return
below the 5th percentile of a fund’s reported returns. Liang and Park (2010)
find that downside risk measures, including expected shortfall, are a significant
determinant of fund failure. In our sample, the live problem funds feature
expected shortfall of –5.67% compared to –9.79% for live non-problem funds.
For live funds, then, the problem funds appear superior to the non-problem
funds. This result indicates that some of the problem funds report returns that
may be too good to be true; hence, the summary statistics will appear more
attractive than for non-problem funds.
The last two columns of Table 4 reveal one other difference between the
groups of funds: the live problem funds feature less serial correlation than the
live funds in the other two groups. The average serial correlation coefficient is
0.1400 for live problem funds, for example, versus 0.2166 for live non-problem
funds. This result may seem counter-intuitive because serial correlation could
be driven by purposeful misreporting in an effort to dampen the perceived
volatility of a fund. We explore this phenomenon in greater detail in the other
panels of Table 4, which show the results for reporting violations and trading
violations separately.
In Panels C and E in Table 4, we report results for the reporting violations and
trading violations, respectively, and in Panels D and F we report the results for
the corresponding control groups. We construct the control groups to account for
other determinants of performance by matching each problem fund with 10 non-
problem funds. First, for each problem fund, we determine potential controls by
matching on live/defunct status, style, maximum assets under management, and
time period. Funds are classified as equity if they are in the Equity Long/Short or
Equity Market Neutral styles; equity (non-equity) problem funds are matched
with equity (non-equity) non-problem funds. The time periods of two funds

2687

[13:45 28/7/2012 [Link]] Page: 2687 2673–2702

 The Review of Financial Studies / v 25 n 9 2012

match if their first dates are within 24 months of each other, and if their last
dates are as well. Next, we compute pairwise correlations with all potential
matches, and pick the 10 funds with the highest correlation.
In Panels C and E in Table 4, both subsets of the problem funds feature
higher Sharpe ratios than the funds in their respective control groups.
However, the reporting violations have more extreme performance, with
average Sharpe ratios of 0.8593 versus 0.1851 for their control group. The
reporting violations also feature dramatic levels of negative skewness and
excess kurtosis, completely driven by the defunct funds. With regard to serial

Downloaded from [Link] by Trina Holloway user on 20 March 2019

correlation, the reporting violations are similar to their control group, whereas
the trading violations feature significantly less serial correlation. One possible
explanation for this result is that managers charged with trading violations
are typically involved with insider trading or violating short-selling rules. As
a consequence, their fund returns may be generated by capturing short-term
profits more so than other funds, and as a result there is not the same type of
mechanical serial correlation present that arises from conservatively marking
illiquid positions to market.
These preliminary results show that the reported returns of problem funds
exhibit pronounced differences relative to those of non-problem funds. Further,
those funds charged with reporting violations feature return distributions that
are dramatically different from funds charged with trading violations. The next
section determines whether they differ as well in their tendency to trigger
performance flags.

3. Results
3.1 Frequency of performance flags
Our main results are displayed in Table 5. Listed are the percentages of non-
problem funds, reporting violations, and trading violations that trigger each
performance flag. Also listed are p-values of tests for a difference between
the rejection rate of non-problem funds and the other two groups. An ideal flag
would be triggered for a high percentage of problem funds, thereby minimizing
Type I error, and a low percentage of non-problem funds, resulting in a low
Type II error as well. As described in Section 1, our tests are established with a
10% significance level, so that the expected rejection rate for a sample of funds
that report returns accurately is 10%.
Panel A of Table 5 shows results when using all available observations for
each fund. Most of the flags are triggered at a rate substantially higher than
10% for the non-problem funds. The highest rejection rate is 41.7% for the
AR(1) test, and two other tests trigger the flags in over 30% of the funds.
These results can be interpreted two ways. A test may reject at a frequency
above the significance level in the non-problem funds for innocuous reasons.
For example, as argued by Getmansky, Lo, and Makarov (2004), a fund may
feature positive serial correlation because illiquid assets in the portfolio are

2688

[13:45 28/7/2012 [Link]] Page: 2688 2673–2702

 Suspicious Patterns in Hedge Fund Returns and the Risk of Fraud

Table 5
Flag frequencies
Panel A: Full History

Non-Problem Problem Funds

Funds Reporting Violations Trading Violations

Flag (N = 8,575) (N = 79) p-value (N = 112) p-value

Downloaded from [Link] by Trina Holloway user on 20 March 2019

# Zero 18.3% 13.9% 0.3165 14.3% 0.2748
% Repeats 37.4% 60.8%∗∗∗ 0.0000 37.5% 0.9785
Uniform 33.6% 30.4% 0.5423 26.8% 0.1273
String 11.3% 20.3%∗∗ 0.0122 8.0% 0.2820
# Pairs 13.5% 21.5%∗∗ 0.0392 12.5% 0.7519
% Negative 26.0% 46.8%∗∗∗ 0.0000 8.9%∗∗∗ 0.0000
AR(1) 41.7% 30.4%∗∗ 0.0415 27.7%∗∗∗ 0.0027
CAR(1) 7.4% 13.9%∗∗ 0.0290 1.8%∗∗ 0.0230
Maxrsq 25.5% 54.4%∗∗∗ 0.0000 21.4% 0.3226
Switchrsq 22.7% 45.6%∗∗∗ 0.0000 21.4% 0.7529
Indexrsq 17.2% 41.8%∗∗∗ 0.0000 12.5% 0.1936
Kink 17.3% 43.0%∗∗∗ 0.0000 10.7%∗ 0.0668

Panel B: Backfill Test

Non-Problem Problem Funds

Funds Reporting Violations Trading Violations

Flag (N = 6,753) (N = 58) p-value (N = 96) p-value

# Zero 15.9% 17.2% 0.7889 10.4% 0.1408

% Repeats 35.3% 65.5%∗∗∗ 0.0000 25.0%∗∗ 0.0350
Uniform 33.5% 37.9% 0.4808 26.0% 0.1220
String 11.0% 19.0%∗ 0.0543 8.3% 0.4059
# Pairs 13.3% 24.1%∗∗ 0.0156 10.4% 0.4107
% Negative 26.3% 46.6%∗∗∗ 0.0005 4.2%∗∗∗ 0.0000
AR(1) 41.6% 22.4%∗∗∗ 0.0031 24.0%∗∗∗ 0.0005
CAR(1) 7.3% 8.6% 0.7075 3.1% 0.1151
Maxrsq 23.1% 51.7%∗∗∗ 0.0000 22.9% 0.9579
Switchrsq 20.9% 36.2%∗∗∗ 0.0045 21.9% 0.8201
Indexrsq 15.7% 43.1%∗∗∗ 0.0000 10.4% 0.1561
Kink 17.4% 48.3%∗∗∗ 0.0000 13.5% 0.3267
Listed are the percentage of hedge funds that trigger each performance flag at the 10% significance level. Results
are shown for 8,575 non-problem funds; 79 funds with reporting violations, consisting of misappropriation,
overvaluation, misrepresentation, or Ponzi schemes; and 112 funds with trading violations, involving short-sale
rules, insider trading, or any other offense. Panel A includes all available return observations, whereas Panel B
drops the first 12 for each fund. # Zero is triggered by a high number of returns exactly equal to zero. % Repeat
is triggered by a high number of returns that are repeated. Uniform is triggered by a distribution of the last
digit of returns that is significantly different from a uniform distribution. String is triggered by a long string of
repeated returns. # Pairs is triggered by a high number of pairs of repeated returns. % Negative is triggered by
a low number of negative returns. AR(1) is triggered by a statistically significant and positive first-order serial
correlation coefficient. CAR(1) is triggered by a larger serial correlation conditioned on a negative lagged fitted
value from a regression involving an optimal set of style factors. Maxrsq and Switchrsq are triggered by an
adjusted R-squared that is not significantly different from zero. Indexrsq is triggered by an insignificant relation
between a fund and its category peers. Kink is triggered by a discontinuity at zero in the distribution of a hedge
fund’s returns. The p-values are from tests for a difference between the rejection frequencies of non-problem
funds and the other groups. “***,” “**,” and “*” indicate significance at the 1%, 5%, and 10% levels, respectively.

2689

[13:45 28/7/2012 [Link]] Page: 2689 2673–2702

 The Review of Financial Studies / v 25 n 9 2012

revalued conservatively by the manager. Alternatively, a large number of funds

that have not been charged with a violation may be engaging in questionable
reporting behavior. Bollen and Pool (2009) report that a substantial number
of managers round up returns from negative to positive, consistent with the
rejection rates in our non-problem fund sample of 26.0% for the % Negative
flag and 17.3% for the Kink flag. In other words, our performance flags may
be indicating that some of the non-problem funds are in fact at higher risk of
fraud but have not yet been charged with any violations.
Turning next to the funds with reporting violations, the rejection rates are

Downloaded from [Link] by Trina Holloway user on 20 March 2019

substantially higher than the non-problem funds in nine of the 12 performance
flags, and in six cases the rejection rate is above 40%. The biggest difference
between the rejection rates of non-problem funds and reporting violations is
in the Maxrsq flag, which is triggered in 54.4% of reporting violations versus
25.5% of non-problem funds, followed by the Kink flag, which is triggered
in 43.0% of the reporting violations and only 17.3% of the non-problem
funds. Thus, funds with returns that are unexplained by style factors, or return
distributions with a pronounced discontinuity at zero, are much more likely
to be engaged in activity resulting in a reporting violation than other funds.
This result suggests that performance flags may be used to identify funds with
higher risk of fraud.
Other performance flags are triggered at a higher rate for the reporting
violations than non-problem funds, including the % Negative, # Pairs, and
String flags. The highest rejection rate of the other flags is 60.8% for the
% Repeat flag, compared to 37.4% in the non-problem funds. The multiplicity
of potential indicators suggests that a multivariate comparison of the two groups
of funds may be warranted, as pursued below.
For the trading violations, the rejection rates are significantly different from
those of the non-problem funds for four of the 12 performance flags, but in
all cases the trading violations reject at lower rates. This result indicates that
the performance flags are associated with activities that are related to fraud as
opposed to violations of trading rules.
Prior hedge fund research has established a variety of important biases
inherent in the self-reported returns contained in the commercially available
databases, including the CISDM and TASS data we employ. Two of the most
important are backfill bias and survivorship bias. We include defunct funds
in our study so survivorship bias is not a concern. Ordinarily backfill bias is
mitigated by dropping the first 12 or 18 observations for a given fund, as these
may reflect unusually good performance. Our study focuses on peculiarities
in reported returns, and so we include as many observations as possible.
Nonetheless, it is interesting to see whether the patterns we identify are present
in the full history of a fund as opposed to just its initial stage; hence, we re-
compute performance flags after dropping the first 12 observations of each
fund’s history. This reduces the number of funds that satisfy our minimum
data requirements. Panel B shows the results for the backfill-corrected sample.

2690

[13:45 28/7/2012 [Link]] Page: 2690 2673–2702

 Suspicious Patterns in Hedge Fund Returns and the Risk of Fraud

Table 6
Cross-sectional distribution of flags
Percentile
Sample 5% 10% 25% 50% 75% 90% 95%

Panel A: Kink
Non-Problem Funds (N = 8,575) −3.10 −2.32 −1.20 −0.24 0.64 1.41 1.84
Reporting Violation (N = 79) −16.76 −11.52 −3.51 −1.29 0.00 0.51 1.15
Trading Violations (N = 112) −2.25 −1.80 −0.87 −0.22 0.70 1.61 1.82

Panel B: Maxrsq

Downloaded from [Link] by Trina Holloway user on 20 March 2019

Non-Problem Funds 7.9% 12.1% 21.5% 35.1% 50.9% 64.2% 70.8%
Reporting Violation 3.6% 5.9% 9.0% 20.4% 39.6% 48.2% 63.7%
Trading Violations 11.2% 13.0% 21.8% 32.4% 46.0% 58.3% 62.6%

Panel C: % Negative
Non-Problem Funds 13.3% 20.0% 28.6% 36.1% 42.5% 48.4% 52.6%
Reporting Violation 0.0% 0.0% 5.9% 10.5% 23.5% 37.1% 41.5%
Trading Violations 11.0% 14.4% 27.2% 33.8% 38.2% 43.6% 48.8%

Panel D: % Repeat
Non-Problem Funds 0.0% 0.0% 0.0% 2.1% 6.1% 12.5% 18.9%
Reporting Violation 0.0% 0.0% 0.0% 5.8% 14.8% 23.8% 41.3%
Trading Violations 0.0% 0.0% 0.0% 3.0% 7.7% 13.4% 18.6%

Listed are percentiles of the cross-sectional distributions of four patterns in returns underlying performance
flags for three samples of hedge funds. Results are shown for 8,575 non-problem funds; 79 funds with reporting
violations, consisting of misappropriation, overvaluation, misrepresentation, or Ponzi schemes; and 112 funds
with trading violations, involving short-sale rules, insider trading, or any other offense. Panel A shows the
percentiles of a standard normal test statistic for a discontinuity at zero in the distribution of a hedge fund’s
returns. Panel B shows the percentiles of the maximum adjusted R-squared obtained by searching over subsets
of 13 explanatory variables. Panel C shows the percentiles of the percentage of negative returns. Panel D shows
the percentiles of the percentage of repeat returns.

The results are qualitatively identical to those in Panel A, indicating that funds
with reporting violations feature suspicious patterns beyond the first year of
their histories.
Table 6 reports the cross-sectional distribution of four of the patterns in
returns underlying the performance flags with the highest rejection rates for the
reporting violations.7 Panel A reports various percentiles of the Kink statistic,
which is distributed standard normal under the null hypothesis. The median
non-problem fund has a Kink statistic of –0.24 compared to –0.22 for trading
violations, whereas the median reporting violation has a Kink statistic of –1.29.
At the 5th percentile, non-problem funds have a Kink of –3.10 versus −2.25
for trading violations and –16.76 for reporting violations. Thus, it appears that
the cross-sectional distribution of the Kink statistic is substantially shifted to
the left for reporting violations compared to that of all other funds. Figure 1
depicts the empirical distribution of returns for the three samples of funds. In all
cases, there is a statistically significant discontinuity in the distribution at zero,

7 We do not show the results for the Switchrsq flag, given its close relation with the Maxrsq flag.

2691

[13:45 28/7/2012 [Link]] Page: 2691 2673–2702

 The Review of Financial Studies / v 25 n 9 2012

A 10%
8%

6%

4%

2%

0%

+10

+15

+20

+25

+30

+35

+40

+45

+50
-50

-45

-40

-35

-30

-25

-20

-15

-10

+5
-5

Downloaded from [Link] by Trina Holloway user on 20 March 2019

B 10%
8%

6%

4%

2%

0%
+10

+15

+20

+25

+30

+35

+40

+45

+50
-50

-45

-40

-35

-30

-25

-20

-15

-10

+5
-5

C 10%
8%

6%

4%

2%

0%
+10

+15

+20

+25

+30

+35

+40

+45

+50
-50

-45

-40

-35

-30

-25

-20

-15

-10

+5
-5

Figure 1
Pooled return distributions
Depicted are three histograms of monthly hedge fund returns using 101 bins centered on the first bin to the
right of zero and labeled “0.” Bold vertical bars indicate the two bins bracketing zero. Figure 1A includes
returns of 8,575 funds, which are not in a sample of funds subject to SEC examinations or lawsuits. Figure 1B
includes returns of 79 funds charged with misappropriation, overvaluation, misrepresentation, or Ponzi schemes.
Figure 1C includes returns of 112 funds charged with violating short-sale rules, insider trading, or other offenses.
Bin sizes are 17 basis points (bp), 16 bp, and 35 bp, respectively.

consistent with the results in Bollen and Pool (2009). However, the reporting
violations sample features a much sharper break.
Panel B of Table 6 shows percentiles of the Maxrsq. At the median, non-
problem funds have a Maxrsq of 35.1% compared to just 20.4% for reporting
violations. This is consistent with the result in Table 5 that over 50% of
the reporting violations have a Maxrsq that is indistinguishable from zero.
Percentiles from the trading violations are in all cases similar to the non-problem
funds. Panel C reports the distribution of the percentage of negative returns.
Here again there are substantial differences. Reporting violations have only
10.5% negative returns at the median, for example, compared to 36.1% for the

2692

[13:45 28/7/2012 [Link]] Page: 2692 2673–2702

 Suspicious Patterns in Hedge Fund Returns and the Risk of Fraud

non-problem funds. Panel D shows that the reporting violations have a larger
percentage of repeat returns than other funds, 5.8% at the median, for example,
compared to 2.1% for non-problem funds. Taken together, these results suggest
that the performance flags contain information that distinguishes the reporting
violations from the rest of the sample.

3.2 Probability of a violation

We use probit analysis to assess the relation between performance flags and

Downloaded from [Link] by Trina Holloway user on 20 March 2019

the incidence of subsequent SEC examinations or lawsuits. The dependent
variable equals one for a violation and zero otherwise. The probit analysis
provides three benefits. First, the multivariate setting allows us to determine
which of the performance flags are more informative for measuring the risk of
fraud. Second, the probit is a convenient method for assessing the additional
explanatory power of the performance flags as a group relative to the operational
risk variables studied in Brown et al. (2008) and Brown et al. (2009). Third,
we can construct a scalar measure of the risk of fraud using the fitted value
of the probit, taking into account the empirical relevance of each performance
flag relative to the others.
The independent variables include the performance flags listed in Table 5
except for Switchrsq, # Pairs, and % Negative, each of which is highly
correlated with at least one other flag. The # Zero, % Repeat, and String flags
equal one when their test statistics reject the null at the 10% level and zero
otherwise. For these variables, the marginal effect, which we label  prob,
equals the increase in probability when the flag equals one. Maxrsq is the
fund’s maximum adjusted R-squared. Kink is the z-score of a discontinuity
test. Indexrsq is the p-value of the slope coefficient from a regression of
fund returns on a corresponding index. AR(1) and CAR(1) are z-scores from
serial correlation coefficients. Uniform is the p-value of a test for a uniform
distribution in the last digit of a fund’s returns. For these variables,  prob
equals the increase in probability for a one-cross-sectional-standard-deviation
increase in the independent variable. We also include each fund’s average
monthly return, volatility, and the natural log of a fund’s maximum assets
under management to determine whether the performance flags contribute
incremental information beyond these summary statistics.8
From the univariate frequency analysis in Table 5, we know that most of the
flags are triggered at a higher rate for reporting violations than for non-problem
funds, and none of the flags are triggered at a higher rate for trading violations.
Thus, we expect ex ante that more of the performance flags are significant in
the probit analysis for the reporting violations than the trading violations.

8 The control group is formed by matching on fund attributes including size. However, since the matching criteria
still permit substantial variation in size between a problem fund and its matches, including size as an independent
variable is not redundant.

2693

[13:45 28/7/2012 [Link]] Page: 2693 2673–2702

 The Review of Financial Studies / v 25 n 9 2012

In addition to the independent variables described above, we include a

univariate measure of operational risk labeled Omega based on the results
of Brown et al. (2008) and Brown et al. (2009). Brown et al. (2008) collect
proxies for operational risk from the mandatory 2006 SEC Form ADV filings.
Following the elimination of the private adviser exemption to the Investment
AdvisersAct contained in the Dodd-Frank legislation, FormADV filings may be
more widely available in the future. However, research on fraud and operational
risk is limited by the lack of historical filings for many advisers. Brown et al.
(2009) argue that it is possible to construct an instrument for operational risk

Downloaded from [Link] by Trina Holloway user on 20 March 2019

using more commonly accessible information from hedge fund databases such
as TASS and CISDM. Their instrument is a linear combination of the following
variables: average monthly returns, standard deviation, fund size, fund age,
whether the fund reports total assets, incentive fee, margin, whether the fund
is audited, personal capital, an onshore indicator, whether the fund is open
to investors, and whether the fund accepts managed accounts. We use these
variables in a first-stage probit model to construct an estimate of a fund’s
operational risk. Omega, our operational risk proxy, equals a fund’s predicted
probability from this model scaled by the unconditional probability of being a
problem fund.
Results are displayed in Table 7. We conduct the analysis four ways. In all
cases, we compare the non-problem funds to reporting violations and trading
violations separately. In Panel A, the violations are compared to the full sample
of non-problem funds, whereas in Panel B the violations are compared to the
control groups. Listed are coefficient estimates, their p-values, and marginal
effects. Both the number of funds and the number of violations are reduced by
the data limitations for constructing the Omega operational risk proxy.
For the reporting violations in Panel A in Table 7, three of the performance
flags are significant at the 10% level. The two most highly significant are
Kink and Indexrsq, for which a one-standard-deviation change would result
in an increase in probability of a reporting violation by 0.22% and 0.32%,
respectively, from an unconditional probability of 0.91%. For the trading
violations in Panel A, three of the performance flags are statistically significant,
but are the wrong sign. For these funds, Avg Return and Volatility are also
significant. One explanation for this result is that funds engaging in insider
trading can generate superior returns. For both the reporting violations and
trading violations, Omega is statistically and economically significant: a one-
standard-deviation change increases the probability of a violation by 0.21%
and 0.18%, respectively. This result indicates that the information provided
by performance flags is complementary to the information contained in the
operational risk proxy. Note also that Omega is significant for both types of
violations, consistent with the idea that weak internal controls would raise the
probability of all types of violations. In contrast, the performance flags are
only informative for reporting violations, since they are measures based on
suspicious patterns in reported returns.

2694

[13:45 28/7/2012 [Link]] Page: 2694 2673–2702

[13:45 28/7/2012 [Link]]

Downloaded from [Link] by Trina Holloway user on 20 March 2019

Suspicious Patterns in Hedge Fund Returns and the Risk of Fraud
Table 7
Probit analysis
Panel A: Full Sample Panel B: Control Group
Reporting Violations Trading Violations Reporting Violations Trading Violations
Flag Estimate p -value  prob Estimate p -value  prob Estimate p -value  prob Estimate p -value  prob

Avg Return −2.1856 0.8035 −0.02% 20.3490∗∗∗ 0.0084 0.30% −3.6664 0.6711 −0.43% 10.1960 0.4148 0.85%
Volatility 1.8233 0.4700 0.06% −6.5592∗∗∗ 0.0079 −0.35% 2.2641 0.4120 0.84% −1.4321 0.6477 −0.42%
Maxrsq −0.7158∗ 0.0554 −0.15% 0.0379 0.8919 0.01% −1.5913∗∗ 0.0115 −3.27% −1.1521∗∗∗ 0.0060 −2.66%
Kink −0.1193∗∗∗ 0.0000 −0.22% 0.0706∗∗ 0.0305 0.22% −0.1350∗∗∗ 0.0000 −3.98% 0.0467 0.4248 0.70%
Indexrsq 0.5095∗∗∗ 0.0003 0.32% −0.2501 0.1310 −0.13% 0.7991∗∗∗ 0.0000 5.04% 0.1747 0.5247 0.54%
AR(1) 0.0155 0.7409 0.03% 0.0008 0.9821 0.00% 0.0299 0.6303 0.44% 0.0394 0.5451 0.52%
CAR(1) 0.0190 0.5295 0.04% −0.0734∗∗ 0.0205 −0.25% 0.0261 0.4124 0.74% −0.1650∗∗∗ 0.0014 −3.26%
# Zero 0.2462 0.1177 0.10% −0.1423 0.2284 −0.10% 0.4962∗ 0.0518 2.12% −0.2339 0.1887 −1.09%
Uniform −0.2792 0.1058 −0.09% 0.0993 0.5394 0.05% −0.5606∗ 0.0546 −1.94% 0.2366 0.3298 0.83%
% Repeats −0.0540 0.7367 −0.03% −0.2604∗ 0.0528 −0.20% 0.1980 0.4201 0.95% −0.2425 0.2392 −1.18%
String −0.1786 0.2560 −0.06% 0.0211 0.8909 0.01% −0.4443∗ 0.0720 −1.61% −0.0548 0.8097 −0.19%
Size −0.0107 0.7466 −0.02% 0.0589 0.1398 0.17% −0.1144∗∗ 0.0170 −2.13% 0.0353 0.5215 0.68%
Omega 0.1225∗∗∗ 0.0000 0.21% 0.0812∗∗∗ 0.0004 0.18% 0.4106∗∗∗ 0.0000 4.00% 0.4447∗∗∗ 0.0000 4.49%
No. of Obs. 6,144 6,144 525 641
No. of Violations 56 54 56 54
Pseudo-R 2 26.10% 8.28% 32.00% 16.90%

Listed are results of four probit analyses relating fund attributes to the incidence of reporting or trading violations. Panel A compares violations to all non-problem funds. Panel B compares
violations to a set of control funds. Violations include 56 funds accused of a reporting violation, consisting of misappropriation, overvaluation, misrepresentation, or Ponzi schemes, and 54 funds
accused of a trading violation, involving short-sale rules, insider trading, or any other offense. The dependent variable equals one for funds with a violation and zero otherwise. The independent
variables are fund attributes and test statistics. Avg Return and Volatility are computed from monthly returns. Maxrsq is the fund’s maximum adjusted R-squared. Kink is the z-score of a test for
a discontinuity at zero in the distribution of a hedge fund’s returns. Indexrsq is the p-value of the slope coefficient from a regression of fund returns on a corresponding style index. AR(1) is the
z-score of a fund’s first-order serial correlation coefficient. CAR(1) is the z-score of a fund’s incremental first-order serial correlation following poor returns. # Zero equals one if a fund has too
many returns exactly equal to zero. Uniform is the p-value of a test for a uniform distribution in a fund’s last digit of returns. % Repeat equals one if a fund has too many returns that are repeated.
Page: 2695

String equals one if a fund has too long a string of repeated returns. Size is the natural log of a fund’s maximum assets under management. Omega measures operational risk. The # Zero, % Repeat,
String, and % Negative tests are evaluated at the 10% significance level. For these tests,  prob equals the increase in probability when the variable equals one. For other independent variables,
 prob equals the increase in probability for a one-standard-deviation increase in the independent variable. “***,” “**,” and “*” indicate significance at the 1%, 5%, and 10% levels, respectively.
2695
2673–2702
 The Review of Financial Studies / v 25 n 9 2012

In Panel B of Table 7, the results are similar when comparing

the two sets of violations to their control groups. For the reporting
violations, the two most significant flags are again Indexrsq and Kink.
Here, the unconditional probability of 10.67% is increased by 5.04% with
a one-standard-deviation increase in Indexrsq and by 3.98% for a one-
standard-deviation decrease in the Kink variable. Proportionately, these
increases are larger than in Panel A, indicating that after controlling
for other fund attributes, the flags are even more informative. For the
trading violations, the only performance flag that is significant and the

Downloaded from [Link] by Trina Holloway user on 20 March 2019

predicted sign is the Maxrsq flag. For both the reporting violations and
trading violations, the Omega variable is again significant, raising the
unconditional probability of a violation by 4.00% and 4.49%, respec-
tively.
The control group results in Panel B in Table 7 are important because
they indicate that some of the performance flags contain information about a
heightened risk of fraud even after controlling for other fund attributes. Hence,
the performance flags, notably Maxrsq, Indexrsq, and Kink, are not simply
proxying for ordinary factors that might lead to greater scrutiny and hence
higher rates of observed violations. Instead, they appear to contain incremental
information regarding the behavior of some fund managers and the likelihood
of subsequent charges.
As discussed in King and Zeng (2001), the statistical properties of the probit
model are heavily influenced by the mean of the dependent variable, which
is the relative frequency of events in the sample. The coefficient bias and
parameter instability normally associated with small samples remain an issue
with much larger sample sizes when events are “rare.” Since the full sample
results in Table 7 involve a roughly 1% unconditional probability, we test for
robustness using King and Zeng’s rare events logistic regression. The results
are qualitatively identical using this rare events procedure.
To create a scalar measure of the risk of fraud for each fund, we compute
the f -score, which is the ratio of a fund’s fitted probability of a violation,
given the coefficient estimates in the probit analysis, to the unconditional
probability. This metric, computed as in Dechow et al. (2011), could be used
by regulators as a single encompassing indicator of the probability of trouble
and the need for an examination of the fund. In this analysis, we first rerun the
probit analysis after excluding the Omega variable to assess the ability of the
performance flags alone to separate the problem funds from the non-problem
funds.
Table 8 shows the percentage of funds with f -scores above three levels:
the 95th and 90th percentiles established using the non-problem funds as well
as 1.00, which occurs when the conditional probability of a violation equals
the unconditional probability. Panel A compares violations to the full sample of
non-problem funds. Using coefficients from the analysis of reporting violations,
the 95th percentile of the f -score is 2.64 for non-problem funds, i.e., 5% of the

2696

[13:45 28/7/2012 [Link]] Page: 2696 2673–2702

 Suspicious Patterns in Hedge Fund Returns and the Risk of Fraud

Table 8
f -scores
Panel A: Full Sample Panel B: Control Group
Reporting Violations Reporting Violations
Problem Non-Problem Problem Non-Problem
f -score Funds Funds f -score Funds Funds

>2.64 34.9% 5.0% >2.87 34.9% 5.0%

>1.94 48.5% 10.0% >1.84 50.0% 10.0%
>1 80.3% 30.6% >1 77.3% 25.1%

Trading Violations Trading Violations

Downloaded from [Link] by Trina Holloway user on 20 March 2019

Problem Non-Problem Problem Non-Problem
f -score Funds Funds f -score Funds Funds

>2.38 20.8% 5.0% >2.21 16.9% 5.0%

>1.91 33.8% 10.0% >1.85 28.6% 10.0%
>1 67.5% 38.4% >1 75.3% 37.7%
Listed are the percentage of hedge funds by f -score, defined as the ratio of conditional probability to unconditional
probability of a subsequent violation. Conditional probabilities are computed for each fund using the fitted
values from a probit analysis. Panel A shows results comparing violations to the full sample of non-problem
funds, whereas Panel B compares violations to a control group. In both cases, reporting violations, including
misappropriation, overvaluation, misrepresentation, or Ponzi schemes, are analyzed separately from trading
violations, involving short-sale rules, insider trading, or any other offense. Three levels of f -score are assessed:
those corresponding to the 95th and 90th percentiles of non-problem funds and 1.00, which occurs when the
conditional probability equals the unconditional probability.

non-problem funds have f -scores above 2.64. In contrast, 34.9% of reporting

violations have f -scores that exceed this level. These results correspond to
the size and power of the indicator based on performance flags: a 5% false
positive rate as compared to a 34.9% successful identification rate.9 Similarly,
10% of the non-problem funds have f -scores above 1.94 compared to 48.5%
of reporting violations. The f -score is therefore able to identify a subset of
the fund universe that has a much higher risk of fraud than the typical fund.
The trading violations have weaker results. Panel B displays results comparing
the two sets of violations to their respective control groups, and the results are
qualitatively identical, again indicating that the performance flags are providing
new information.

3.3 Response of investors

The performance flags appear to help predict which funds subsequently are
prosecuted by the SEC or are sued by investors. A related question is whether
the flags are redundant because investors are already able to spot these
funds through other means, for instance, through due diligence efforts that

9 While the power may seem low, note that in practice the SEC uses a wide variety of variables to select funds
for examination, so the f -score can be viewed as a supplement to the other signals used, such as complaints
from investors or counter-parties. Moreover, combining our performance measures with operational risk proxies
would further increase the magnitude of the f -scores for problem funds, as implied by the significance of Omega
in Table 7.

2697

[13:45 28/7/2012 [Link]] Page: 2697 2673–2702

 The Review of Financial Studies / v 25 n 9 2012

focus on other information, such as the presence of operational flags discussed

in Brown et al. (2008). To answer this question, we examine the response
of investors to performance, and whether this sensitivity differs across the
problem funds and non-problem funds. As argued by Brown et al., if investors
are already able to identify funds with a higher risk of future legal trouble, we
might expect good performance by problem funds to attract less capital than
good performance by non-problem funds.
The flow-performance regressions are similar to those in Sirri and Tufano
(1998), also adopted in Brown et al. (2008), in which annual percentage flow

Downloaded from [Link] by Trina Holloway user on 20 March 2019

is calculated as the change in assets under management after controlling for
fund returns. To measure performance, each year we sort funds into percentiles
using raw returns relative to funds in the same style. We estimate the relation
between performance and flows using a piecewise linear specification with four
performance variables based on fund percentiles. We only include observations
of flows and returns for problem funds in the years prior to the filing date of the
SEC enforcement actions or lawsuits. We also include interaction terms equal
to the product of each performance variable and an indicator that equals one if
a fund is in the problem fund subsample and zero otherwise. These interaction
terms measure the difference between the flow-performance sensitivity of
problem funds and non-problem funds. We use prior-year log fund size, return
volatility, and flow, as well as contemporaneous average flow to the fund’s style
group, incentive fee, and management fee, as additional explanatory variables
in the regression. The flow-performance results are estimated using the Fama
and MacBeth (1973) method.
Table 9 shows the results of the flow-performance analysis. For both
reporting violations and trading violations, flows are negatively related to
size, consistent with prior research and the tendency of successful funds to
increase in size and then close to new investment due to concerns over the
scalability of a fund’s strategy. Flows are also negatively related to volatility,
which is consistent with Brown et al. (2008) and Ding et al. (2008), and
consistent with risk-averse investors. Flows are positively related to prior-
year flows, perhaps driven by multi-year fund growth, and positively related
to contemporaneous category flow, as one might expect. More surprisingly,
we find that flows are positively related to management fees, suggesting that
high fees may indicate superior future fund management.10 The coefficients
on the performance variables are all positive, and all are significant at the 1%
level. Their magnitudes indicate that above-median returns are rewarded with
inflows at a higher rate than below-median returns are punished by outflows:
the best-performing fund has performance-related inflows equal to about 54%

10 This result is consistent with Gregoriou (2002) but in contrast to Brown et al. (2008) and Ding et al. (2008). Our
data extend through 2008, whereas data in the latter two studies end in 2005 and 2004, respectively. Low returns
in 2008 could distort the sensitivity of flows to fees; managers of high-fee funds have more of an incentive to
prohibit outflows by temporarily invoking redemption restrictions, leading to the positive relation between flows
and fees as in Ang and Bollen (2010).

2698

[13:45 28/7/2012 [Link]] Page: 2698 2673–2702

 Suspicious Patterns in Hedge Fund Returns and the Risk of Fraud

Table 9
Flow-performance sensitivity
Panel A: Reporting Violations Panel B: Trading Violations
Variable Estimate p -value Estimate p -value

Violation −0.0584 0.6937 0.3551 0.3363

Lag Size −0.1787∗∗∗ 0.0000 −0.1789∗∗∗ 0.0000
Lag Volatility −2.8544∗∗∗ 0.0011 −2.8013∗∗∗ 0.0013
Lag Flow 0.0748∗∗∗ 0.0000 0.0751∗∗∗ 0.0000
Category Flow 0.6820∗∗∗ 0.0000 0.6829∗∗∗ 0.0000
Management Fee 0.0750∗∗ 0.0212 0.0772∗∗ 0.0191
Incentive Fee −0.0011 0.6326 −0.0013 0.6036

Downloaded from [Link] by Trina Holloway user on 20 March 2019

qrank1 1.0846∗∗∗ 0.0001 1.0788∗∗∗ 0.0001
qrank2 1.3542∗∗∗ 0.0002 1.3492∗∗∗ 0.0002
qrank3 0.5424∗∗ 0.0270 0.5407∗∗ 0.0271
qrank4 0.7667∗∗∗ 0.0003 0.7683∗∗∗ 0.0003
Violation × qrank1 −0.9659 0.1309 −0.5533 0.6981
Violation × qrank2 1.5840 0.3477 −2.6871 0.3203
Violation × qrank3 −2.1922 0.1937 −1.6069 0.1471
Violation × qrank4 0.0192 0.9510 0.1056 0.8383
No. of Obs. 19,110 19,212
R2 12.45% 12.89%

Listed are the results of regressions of annual fund flow, as a percentage of beginning-of-year assets under
management, on lagged performance. Other variables are defined as follows: Violation is an indicator variable
that equals one if a fund is subsequently charged with a violation and zero otherwise, Lag Size is the natural log
of prior-year fund assets under management, Lag Volatility is prior-year return volatility, Lag Flow is prior-year
fund flow, Category Flow is the contemporaneous average percentage flow to funds in the same style category,
Management Fee is the fixed management fee, and Incentive Fee is the fund’s performance bonus. Performance
measures are based on percentiles of raw returns relative to all other funds in a fund’s style category, and are
split by quartile, as defined as follows for fund i in year t −1:
 
qrank1i,t−1 = percentilei,t−1 −0.50  if 0.75  percentilei,t−1  1.00 and 0 otherwise
qrank2i,t−1 = percentilei,t−1 −0.50 if 0.50  percentilei,t−1 < 0.75 and 0 otherwise
qrank3i,t−1 = percentilei,t−1 −0.50 if 0.25  percentilei,t−1 < 0.50 and 0 otherwise
qrank4i,t−1 = percentilei,t−1 −0.50 if 0.00  percentilei,t−1 < 0.25 and 0 otherwise.

Interaction terms are the product of qrank and Violation. Panel A lists results for all non-problem funds and
79 funds with reporting violations, consisting of misappropriation, overvaluation, misrepresentation, or Ponzi
schemes. Panel B lists results for all non-problem funds and 112 funds with trading violations, involving short
sale-rules, insider trading, or any other offense. “***,” “**,” and “*” indicate significance at the 1%, 5%, and
10% levels, respectively.

of fund assets, whereas the worst-performing fund has outflows equal to about
38% of fund assets. These results indicate an approximately linear relation
between performance and subsequent fund flow after controlling for known
determinants.
The main finding of this analysis is that investors do not appear to treat
problem and non-problem funds differently prior to the violations. Performance
coefficients are the same for reporting and trading violations. More importantly,
coefficients on all interaction terms are insignificant. That is, the sensitivity
of investors to performance is the same for non-problem funds and funds
that are subsequently prosecuted by the SEC or the subject of investor
lawsuits. This result suggests that investors either choose to disregard the
heightened risk of subsequent legal trouble, or are unaware of it, indicating
that predictors such as the performance flags developed here contain new
information.

2699

[13:45 28/7/2012 [Link]] Page: 2699 2673–2702

 The Review of Financial Studies / v 25 n 9 2012

4. Conclusion

Recent cases of hedge fund fraud have prompted revisions to the SEC
Investment Advisers Act, requiring more advisers to register. In the past,
opponents of a registration requirement argued that additional regulation would
act as a hollow promise of safety, since agencies such as the SEC do not have
adequate resources to conduct a sufficient number of examinations to prevent
and deter fraudulent activity. The goal of this study is to test whether low-cost
pre-screens based on suspicious patterns in returns are effective in predicting

Downloaded from [Link] by Trina Holloway user on 20 March 2019

which funds are subsequently charged with legal or regulatory violations.
We determine whether each fund in a large sample triggers one or more
performance flags based on the fund’s time series of reported returns.
Funds subsequently prosecuted for reporting violations trigger the flags at a
substantially higher frequency than other funds, suggesting that pre-screens
based on suspicious patterns in returns may be effective predictors of hedge
fund fraud. In a probit analysis, we find a statistically and economically
significant association between the incidence of a reporting violation and
several performance flags. Since we include a proxy for operational risk as
an additional determinant, our results suggest that performance flags may be
a useful complement to operational risk measures, such as those studied by
Brown et al. (2008) and Dimmock and Gerken (forthcoming).
Although we find a strong relation between the existence of suspicious
patterns in a hedge fund’s returns and the probability of a subsequent legal
or regulatory action, the analysis has been limited to a study of past violations.
Naturally, investors and regulators are more concerned about the ability to
predict future trouble. One important caveat regarding the association between
performance flags and fraud risk is related to the Lucas (1976) critique
of macroeconomic policy research. Historical associations may not persist
following the implementation of a new rule or target if behavior is affected.
Performance flags will provide valuable information in the future only if the
reporting behavior we study persists. There are at least two reasons why one
might expect this to occur. First, conversations with SEC staff confirm that they
use a wide variety of information sources, including patterns in returns, but that
the reasons leading to an examination never have to be revealed. Hence, it is
unclear whether or not fund managers will ever learn which performance flags
are relevant. Second, many of the performance flags result from managerial
incentives to shape the distribution of fund returns in order to attract and
maintain investors. Those managers who purposefully misreport to appease
their investor base may not cease even if it raises the chance of an examination—
if they do, then at least the performance flags served the purpose of providing
a deterrent to future abuse.
Our study indicates that the risk of fraud can be successfully measured using
widely available return data. This evidence supports regulatory oversight of the
hedge fund industry. More generally, our results suggest that required disclosure

2700

[13:45 28/7/2012 [Link]] Page: 2700 2673–2702

 Suspicious Patterns in Hedge Fund Returns and the Risk of Fraud

of financial records can aid in the detection of fraud risk using quantitative
methods.

References

Agarwal, V., N. Daniel, and N. Naik. 2011. Do Hedge Fund Managers Manage Their Reported Returns? Review
of Financial Studies 24:3281–3320.

Agarwal, V., and N. Naik. 2004. Risks and Portfolio Decisions Involving Hedge Funds. Review of Financial
Studies 17:63–198.

Downloaded from [Link] by Trina Holloway user on 20 March 2019

Ang,A., and N. Bollen. 2010. Locked Up by a Lockup: Valuing Liquidity as a Real Option. Financial Management
39:1069–95.

Asness, C., R. Krail, and J. Liew. 2001. Do Hedge Funds Hedge? Journal of Portfolio Management 28:6–19.

Becker, G. 1968. Crime and Punishment: An Economic Approach. Journal of Political Economy 76:169–217.

Bollen, N., and V. Pool. 2008. Conditional Return Smoothing in the Hedge Fund Industry. Journal of Financial
and Quantitative Analysis 43:267–98.

——. 2009. Do Hedge Fund Managers Misreport Returns? Evidence from the Pooled Distribution. Journal of
Finance 64:2257–88.

Bollen, N., and R. Whaley. 2009. Hedge Fund Risk Dynamics: Implications for Performance Appraisal. Journal
of Finance 64:987–1037.

Brown, S., W. Goetzmann, B. Liang, and C. Schwarz. 2008. Mandatory Disclosure and Operational Risk:
Evidence from Hedge Fund Registration. Journal of Finance 63:2785–2815.

——. 2009. Estimating Operational Risk for Hedge Funds: The ωtScore. Financial Analysts Journal 65:43–53.

——. 2012. Trust and Delegation. Journal of Financial Economics 103:221–34.

Burgstahler, D., and I. Dichev. 1997. Earnings Management to Avoid Earnings Decreases and Losses. Journal
of Accounting and Economics 24:99–126.

Cox, J., R. Thomas, and D. Kiku. 2003. SEC Enforcement Heuristics: An Empirical Inquiry. Duke Law Journal
53:737–79.

Dechow, P., W. Ge, C. Larson, and R. Sloan. 2011. Predicting Material Accounting Misstatements. Contemporary
Accounting Research 28:17–82.

Dimmock, S., and W. Gerken. Forthcoming. Predicting Fraud by Investment Managers. Journal of Financial
Economics.

Ding, B., M. Getmansky, B. Liang, and R. Wermers. 2008. Investor Flows and Share Restrictions in the Hedge
Fund Industry. Working Paper, SUNY at Albany.

Fama, E., and J. MacBeth. 1973. Risk, Return, and Equilibrium: Empirical Tests. Journal of Political Economy
81:607–36.

Fung, W., and D. Hsieh. 2001. The Risk in Hedge Fund Strategies: Theory and Evidence from Trend Followers.
Review of Financial Studies 14:313–41.

——. 2002. Asset-based Style Factors for Hedge Funds. Financial Analysts Journal 58:16–27.

——. 2004. Hedge Fund Benchmarks: A Risk-based Approach. Financial Analysts Journal 60:65–80.

Getmansky, M., A. Lo, and I. Makarov. 2004. An Econometric Model of Serial Correlation and Illiquidity in
Hedge Fund Returns. Journal of Financial Economics 74:529–609.

Gregoriou, G. 2002. Hedge Fund Survival Lifetimes. Journal of Asset Management 3:237–52.

Gregoriou, G., and F. Lhabitant. 2009. Madoff: A Riot of Red Flags. Journal of Wealth Management 12:89–98.

2701

[13:45 28/7/2012 [Link]] Page: 2701 2673–2702

 The Review of Financial Studies / v 25 n 9 2012

King, G., and L. Zeng. 2001. Logistic Regression in Rare Events Data. Political Analysis 9:137–63.

Liang, B. 2003. The Accuracy of Hedge Fund Returns. Journal of Portfolio Management 29:111–22.

Liang, B., and H. Park. 2010. Predicting Hedge Fund Failure: A Comparison of Risk Measures. Journal of
Financial and Quantitative Analysis 45:199–222.

Lucas, R. 1976. Econometric Policy Evaluation: A Critique. Carnegie-Rochester Conference Series on Public
Policy 1:19–46.

Sirri, E., and P. Tufano. 1998. Costly Search and Mutual Fund Flows. Journal of Finance 53:1589–1622.

Straumann, D. 2008. Measuring the Quality of Hedge Fund Data. RiskMetrics Journal Winter:65–93.

Downloaded from [Link] by Trina Holloway user on 20 March 2019

Sun, Z., A. Wang, and L. Zheng. 2012. The Road Less Traveled: Strategy Distinctiveness and Hedge Fund
Performance. Review of Financial Studies 25:96–143.

Titman, S., and C. Tiu. 2011. Do the Best Hedge Funds Hedge? Review of Financial Studies 24:123–68.

Waring, M., and L. Siegel. 2006. The Myth of the Absolute-return Investor. Financial Analysts Journal 62:14–21.

Working, H. 1960. Note on the Correlation of First Differences of Averages in a Random Chain. Econometrica
28:916–18.

2702

[13:45 28/7/2012 [Link]] Page: 2702 2673–2702