Do hedge funds exhibit performance persistence? A new approach

Transcription

1 Do hedge funds exhibit performance persistence? A new approach Nicole M. Boyson * October, 2003 Abstract Motivated by prior work that documents a negative relationship between manager experience (tenure) and performance, we design a new approach to detect persistence. While a portfolio of funds selected on past performance alone shows no persistence, a portfolio that is long low tenure/past good performers and short high tenure/past poor performers displays quarterly persistence. Consistently poor performance among high tenure/past poor performers drives this finding, and occurs because young managers are much more likely than old to be terminated for poor performance. This termination/performance asymmetry provides additional support for previous research showing greater risk-taking behavior among young managers. *Purdue University, Krannert Graduate School of Management, West Lafayette, IN nboyson@mgmt.purdue.edu. Phone: (765) I would like to thank Vikas Agarwal, Mike Cooper, Yong Chen, Jean Helwege, Ravi Jagannathan, Andrew Karolyi, Narayan Naik, Karen Wruck, and René Stulz (my advisor). Ashour Yacoub and John Shelbourne provided excellent research assistance. All remaining errors are my own. 1

2 1. Introduction In selecting a hedge fund for investment, is it helpful to consult the manager s prior performance record? If past performance is indicative of future results, there is value to investors in this information. If not, then investors may be better off selecting a manager based on his reputation, investment style, or trading costs. Recent research regarding this issue finds consistent results: there is some evidence of short term (one to three month) persistence among individual hedge funds. (See Agarwal and Naik (1999, 2000), Bares, Gibson, and Gyger (2002), and Baquero, ter Horst, and Verbeek (2002).) This persistence is not driven by the existence of survivorship bias. At longer time horizons (semi-annual or beyond), however, persistence largely disappears; see e.g., Brown, Goetzmann, and Ibbotson (1999), and Brown and Goetzmann (2001). With a more rigorous approach that controls for common risk and style factors in hedge fund returns, this paper finds no evidence of persistence (short or long-term) when funds are selected based on past performance alone. Style factors explain the previous findings of short-term persistence, consistent with the work of Brown and Goetzmann (2001) who show that certain styles perform well in certain periods; in other periods, these same style do not perform as well. Thus, controlling for style is important in an analysis of performance persistence among hedge funds. However, while controlling for style casts doubt upon the previous findings of persistence, there is another important factor that should be considered in constructing a test of performance persistence manager tenure. Boyson (2003) shows that less experienced managers (hereafter referred to as young or low tenure managers) significantly outperform more experienced managers (hereafter referred to as old or high tenure managers). Specifically, in a sample of hedge funds for the period 1994 to 2000, she finds that after controlling for common risk and style factors, the annual difference in performance between young and old managers drops by about 0.75% for each year of experience. That is, a manager who is 52 years of age has annual performance about 4% lower than the average (47 year old) manager in the sample. Her results suggest the following: since low tenure managers are better, then a bad return for a low tenure manager is more likely to be due to bad luck than for a high tenure manager. Likewise, a good return for a high tenure manager is 2

3 more likely to be due to good luck than for a low tenure manager. In other words, good (bad) returns for low tenure managers are likely to be due to superior manager skill (bad luck); good (bad) returns for high tenure managers are likely to be due to good luck (lack of manager skill). Thus, properly accounting for manager tenure when performing a persistence analysis should detect performance persistence among the young versus the old hedge fund managers. The remainder of the paper designs a more powerful test of performance persistence, taking into account manager tenure. I construct a portfolio that takes a long position in low tenure/past good performers and a short position in high tenure/past poor performers, which by design, should maximize the likelihood of finding persistence. And, this portfolio demonstrates quarterly persistence: controlling for risk and style factors, the excess performance is about 9% annually, which is both economically and statistically significant. This result is driven primarily by persistent underperformance among old, past poor performers. Next, we explain the concentration of persistence among old past performers with the following hypothesis: that the termination relationship is more performance-sensitive for young managers. If this is the case, then old, poor performers have a low probability of being terminated and thus are more likely to persist in the next period. This hypothesis is motivated by theoretical literature that suggests that young managers will be punished more severely for poor performance than are old. 1 It is also motivated by empirical results for mutual fund managers and security analysts. (See Chevalier and Ellison (1999b) and Hong, Kubik, and Solomon (2000)). A conditional survival analysis documents the following results: conditional on having been a poor past performer (in the bottom third of returns), young managers are significantly more likely to be terminated than old. Also, conditional upon having been a middle performer (in the middle third of performance) young managers are still significantly more likely to be terminated than old. Only when we condition upon having been a past good, performer (top third of returns) is there no difference in survival rates 1 See, for example, Zwiebel (1995) and Holmstrom (1999) who describe the process by which investors find out about managers with a learning model. Each period, investors observe a new performance outcome (in this case, a monthly return) by which they learn about manager ability. Since there are more observations for older managers than young, this implies that the sensitivity of a manager s reputation is less dependent on the most recent observation. Hence, old managers are less likely to be assessed as inferior based on a recent bad outcome than are young. 3

4 among young and old managers. Thus, being in the bottom two-thirds of performance significantly hurts young managers relative to old. The second survival analysis (this time, conditional upon manager tenure) establishes the following result: conditional on being young, past poor performers are more likely to be terminated than past good performers. However, conditional on being old, there is no difference in survival rates between past poor and past good performers. These findings support each other are broadly consistent with the idea that investors are more likely to tolerate poor performance from managers with moreestablished reputations (i.e., more experienced managers). Thus, this finding helps to explain the continued poor performance among old, past poor performers. This paper makes two contributions to the literature. First, it takes advantage of the empirical result that young managers outperform old to design a test that detects risk- and style-adjusted performance persistence at the quarterly level. While selecting funds based on past performance alone results in a finding of no performance persistence, the more powerful approach of choosing funds based on both past performance and manager tenure does result in a finding of persistence. This persistence is mostly concentrated among the old, past poor performers. To our knowledge, this is the first paper in the literature to test for performance persistence in this manner. Second, this paper explains this finding of persistence (notably, that it is concentrated among the old, past poor performers) as being driven by differences in termination rates among young and old managers. Specifically, there is an interesting asymmetry in the shape of the termination and age/performance relationship: the termination process is much more performance-sensitive for young managers than for old. At first glance, this relationship appears similar to that in the mutual fund industry: Chevalier and Ellison (1999b) also find that young mutual fund managers are more likely to be terminated than old for poor performance. However, there is a key difference between their results and the results of this paper. They find that for young mutual fund managers, the probability of termination decreases steeply with performance when managers have negative excess returns, but it is fairly insensitive to performance differences at positive excess return levels. 2 As long as a young manager s returns are positive, his probability of failure does not differ from that of an 2 Chevalier and Ellison (1999b), page

5 older manager. By contrast, in the hedge fund industry the performance threshold is much higher, and is measured relative to other managers rather than based on absolute performance: unless a young hedge fund manager s returns are in the top third of managers, his probability of failure is significantly higher than that of an older manager. In other words, young hedge fund managers have to beat two-thirds of other managers to reduce their probability of failure to the same as that of older managers. Clearly, this high threshold of performance in the hedge fund industry sets up a different incentive structure for young mutual fund versus young hedge fund managers: while young mutual fund managers concerned about survival need only avoid posting a negative excess return (which gives them an incentive to play it safe and avoid idiosyncratic risk), young hedge fund managers that are concerned with survival need to post returns in the top third of all performers (which gives them an incentive to make bold investment decisions relative to other managers.) The empirical evidence for hedge fund managers is completely consistent with this implied incentive: Boyson (2003) shows that young managers herd less and take on more idiosyncratic risk than old. While this paper contributes to a relatively small and recent literature in the hedge fund industry, researchers have been studying persistence in the mutual fund and pension fund industries for many years, with mixed results. An early study by Jensen (1968) finds no support for persistence. Papers supporting persistence over five to ten year periods include Grinblatt and Titman (1992), Elton, Gruber, Das and Hlavka (1993), and Elton, Gruber, Das and Blake (1996), who attribute this persistence to manager stock-picking ability. Support for shorter-term (one to three year) persistence comes from Hendricks, Patel, and Zeckhauser (1993), Goetzmann and Ibbotson (1994), Brown and Goetzmann (1995), and Wermers (1999). Carhart (1997) shows that the one-year momentum effect of Jegadeesh and Titman (1993) accounts for much of the performance persistence found by Hendricks, Patel, and Zeckhauser (1993), and that differences in mutual fund expenses and trading costs can explain nearly all of the remaining persistence. Christopherson, Ferson, and Glassman (1998) apply conditional performance evaluation techniques to a sample of pension funds, and show that the conditional approach is better able to detect persistence and predict future performance than unconditional (linear) methods. This persistence is mostly concentrated among the 5

6 worst performers. More recently, using a Bayesian approach with daily mutual fund data, Bollen and Busse (2002), and Busse and Irvine (2002) find evidence of quarterly performance persistence that is not explained by momentum. This paper is organized as follows. Section 2 describes the data. Section 3 describes the performance measures and portfolio formation process used in Section 4. Section 4 performs the first analysis of persistence, and shows that when common risk and style factors are properly accounted for, there is no evidence of quarterly persistence when funds are selected based on past performance only. Section 5 motivates and designs a more powerful test of persistence that incorporates manager tenure, and shows evidence of persistence at the quarterly level. Section 6 performs a detailed survival analysis to explain the patterns in persistence found in Section 5. Section 7 discusses the results of the survival analysis in light of the career concerns literature. Section 8 concludes. 2. Data Data was provided by Tremont Advisory Shareholders Services (TASS). TASS has been collecting hedge fund data directly from managers since the late 1980's, and currently has over 2,400 funds in their database, both living and dead. 3 The database includes monthly net-of-fee returns, as well as expenses, fees, size, terms, age, and style of the funds. For the quarterly persistence tests, we require that each fund have at least 6 months of consecutive returns for inclusion in the sample. For the semi-annual persistence tests, each fund must have at least 12 consecutive months of returns for inclusion, and for the annual persistence tests, each fund must have at least 24 months of consecutive returns. For all time frames (quarterly, semi-annual, and annual) each fund must have at least $5 million in assets during the period January, 1994 to December, In constructing the sample, an important issue must be considered: backfilling or instant history bias (see Edwards and Park (1996). On the date that TASS adds a new fund to their database, they backfill historical returns. Typically, a hedge fund manager will start his fund with a limited amount of personal capital before selling shares to the public. He hopes to compile a good track record so as to eventually attract outside investors. Thus, most 3 TASS has maintained data on dead funds since

7 funds arrive in the database with a history of strong performance which was never available to the public, which biases returns upward. This difference is often large -- using the TASS database, Fung and Hsieh (2000) calculate the bias as about 3.6% per year. The average incubation period in our sample is about one year; thus, to control for this bias we drop the incubation period for each fund. The final sample used to perform the quarterly persistence tests includes 1,659 funds with at least 6 months of returns, $5 million in assets, and all of the fund characteristic variables. Table 1a includes summary statistics of the return and fund characteristic variables. The final sample for the semi-annual persistence tests has 1,503 funds, and the final sample for the annual persistence tests has 982 funds. Summary statistics for these samples do not differ materially from the quarterly sample. As a brief summary, the average fund in the sample is about three years old, has about $80 million in net assets, and annualized excess returns of about 8%. More than half the managers use leverage and have personal capital invested, and the styles of US Equity, Relative Value, Event Driven, and Emerging Markets make up over 60% of the funds in the sample. 4 About 30% of the sample consists of funds that failed at some time during the sample period. With respect to failed funds, a number of researchers have emphasized the importance of including dead as well as extant funds in an analysis of performance. 5 Not including defunct funds in the sample can bias returns upwards. In an earlier study of offshore hedge funds, Brown, Goetzmann, and Ibbotson (1999) find that not including defunct funds in the sample biases returns upwards by 3% per year. Liang (2000) compares two major hedge fund databases (Hedge Fund Research (HFR) and TASS), and finds that the TASS database tends to contain more dead funds, and thus, should be more appropriate for the type of study in this paper. He finds survivorship bias in the TASS data of about 2% per year, which is approximately what we calculate using the sample of 1,659 funds. 4 See Appendix B for a description of the fund styles. 5 In the mutual fund literature, see, for example, Brown, Goetzmann, Ibbotson, and Ross (1992), Wermers (1996), and Carhart (1997). 7

8 3. Performance measures and portfolio formation process 3.1. Performance measures This paper uses a multi-factor model to control for common risk factors in hedge fund performance. Since hedge funds have exposures to a number of markets, and can engage in dynamic trading strategies (such as using options, futures, and leverage), using a broad set of indices is appropriate (see Fung and Hsieh (1997) and Ben Dor, Jagannathan, and Meier (2003) for further discussion). The passive indices are obtained from Datastream and include: the US Trade Weighted Dollar index to capture currency risk, gold and commodity indices, and the Lehman Brothers 30-year Treasury bond and US aggregate bond indices. For stock market risk, the Value-Weighted CRSP index (obtained from WRDS) is used. Additionally included are the Fama-French (1992,1993) SMB (a zero-investment portfolio constructed by subtracting the returns of large market capitalization firms from the returns of small capitalization firms) and HML (a zero investment portfolio constructed by subtracting the returns of low book to market ratio stocks from the returns of high book to market ratio stocks) factors, as well as Jegadeesh and Titman's (1993) momentum factor (MOM) a zero-investment portfolio constructed as the spread between the performance of stocks which were in the top 30% of returns in the prior twelve months and those which were in the bottom 30%. 6 Table 1b reports summary statistics for the passive indices and other risk factors (HML, SMB, and MOM) used in the paper. Since a number of researchers have stressed the importance of considering style in a study of hedge fund performance (see, e.g., Fung and Hsieh (1997), Brown, Goetzmann, and Ibbotson (1999), Ibbotson and Patel (2002), and Ben Dor, Jagannathan, and Meier (2003)), the model includes style as well as common risk factors. These style factors are hedge fund indices are published jointly by Credit Suisse First Boston (CSFB) and TASS, and represent a number of hedge fund trading strategies or styles. They are constructed so as to minimize survivorship bias. For further detail about the indices used, see Appendix A. Table 1b reports summary statistics for all hedge fund style indices used in the paper. The model is as follows: 6 These returns were obtained from the website of Kenneth French. 8

9 K r = α + b F + b H + ε pt pt pkt kt pdt dt t k= 1 d= 1 D (1) where r pt is portfolio p s return in month t in excess of the risk-free rate (t = 1 to T months), the F kt s are each of the passive index returns and HML, SMB, and MOM (k = 1 to K) in month t, and the H dt s are each of the hedge fund index returns (d = 1 to D) in month t Portfolio formation process This section describes the methodology by which hedge funds are sorted into portfolios to evaluate performance persistence. Similar to HPZ (1993), and Carhart (1997), we form portfolios on lagged hedge fund returns and test for both short-term and longer-term persistence. 7 As noted earlier, quarterly (but not longer) persistence in hedge funds has been documented by Agarwal and Naik (2000), Baquero, ter Horst, and Verbeek (2002), and Bares, Gibson, and Gyger (2002). Following Carhart (1997), funds are sorted into decile portfolios based on lagged returns, which are net of fees and expenses and in excess of the risk-free rate. They span the period January, 1994 to December, 2000 for a maximum time series per fund of 84 months. Not all funds have all 84 months of returns available. Some funds fail before the end of the sample period, and other funds do not begin until some time after January, 1994; however, as long as a fund has at least 6 consecutive monthly returns, it is included in the sample. 8 Funds are sorted into portfolios based on three different time frames: three months, six months, and one year. 9 This initial period is called the formation period, and for the persistence analysis, each portfolio is held for a length of time equal to its formation 7 When we form portfolios based on lagged returns, we do this in two different ways. The first method is to form portfolios based on lagged excess-of-risk-free rate returns. The second method forms portfolios based on their returns in excess of their style average. For both methods, the raw returns of the portfolios are then regressed against a number of common risk factors AND style factors as in Equation (1). We find that the results in the paper are robust to whether the portfolios are formed based on excess-of-risk-free-rate or excess-ofstyle-category returns. Thus, only the results from the excess-of-risk-free rate are reported throughout. 8 We require six months of returns for quarterly tests, 12 months for semi-annual tests, and 24 months for annual tests. 9 Since in all the tests performed, funds never exhibit persistence at the semi-annual or annual level, this paper only reports results for quarterly examination periods. Results for other time horizons are available from the author by request. 9

10 period. For example, at the three month time horizon, funds are first sorted into portfolios using the prior quarter's return. These portfolios are held for three months, and equalweighted portfolio returns are calculated for each of the three months. Every three months, portfolios are re-formed. This yields a time-series of 81 monthly returns for each portfolio (the first three months are used in the initial formation period). For the six-month formation period, there are 78 monthly portfolio returns, and for the one-year formation period, there are 72 monthly portfolio returns. As noted above, the process of using portfolios of funds to study persistence has been used by HPZ (1993) and Carhart (1997). Another common methodology examines persistence using individual funds (rather than portfolios of funds). See, for example, Brown and Goetzmann (1992), Goetzmann and Ibbotson (1994), and Agarwal and Naik (2000). While the individual fund approach is appropriate in certain cases, this paper uses the portfolio approach, since this approach is better-suited to a study of hedge funds. The main reason is that the portfolio approach allows for a risk- and style- adjusted analysis of persistence while using a minimum number of time-series observations. As described above, in the portfolio approach, each period a portfolio of funds is created for which an average return is calculated, resulting in a long time-series of portfolio returns which may then be adjusted for common risk and style factors. In this case, only the returns used in the initial formation period are not analyzed for persistence, and persistence may be examined for a larger number of funds over any time frame (e.g., monthly, quarterly, semiannually, etc.). By contrast, when persistence among individual funds is studied, properly adjusting for risk and style factors is typically accomplished by calculating intercepts (alphas) from time-series regressions over the period being analyzed. These alphas are then compared to each other on a period-over-period basis to test for persistence. If one wishes to use a multi-factor model to control for a number of risk and style factors, this necessarily lengthens the time frame over which persistence may be analyzed. For example, this paper uses an model with eighteen (18) independent variables. A study of persistence among individual funds that uses an 18-factor model will not be able to test for persistence at the quarterly, semiannual, or even annual level, since alphas could not be calculated for periods shorter than 18 months due to the degrees of freedom constraint. Additionally, funds would be required to have a much longer 10

11 time-series of returns than in the portfolio approach, which would reduce the sample size significantly. Thus, while the individual fund approach is well-suited to mutual fund studies (which have much longer time series of returns and typically have exposures to a smaller number of risk and style factors), it is less suited to a study of hedge funds, with their short time-series of returns and exposures to a large number of risk and style factors. Hence, the portfolio approach is used in the analyses that follow. 4. Do hedge funds exhibit risk and style-adjusted persistence? In this section, we test for performance persistence when controlling for fund exposure to common risk and style factors. Controlling for style is accomplished by including hedge fund style indices as independent variables, using Equation (1) from Section 3.1. A slightly different way to control for the effect of style on performance would be to model the return process for each fund style. Fung and Hsieh (2001) refer to this approach as developing asset-based style factors. For example, Fung and Hsieh (1997) show that the returns of the trend-following style (which is probably most closely related to the Global Macro style in this paper) can be modeled as a look-back straddle on the S&P 500 index. Mitchell and Pulvino (2001) show that the returns to merger arbitrage hedge funds closely resemble short positions in put options. Finally, Agarwal and Naik (2003) show that a good deal of variation in hedge fund returns can be explained with simple option buying/writing strategies. While these approaches are very helpful in understanding the return processes for each hedge fund style, there are not yet asset-based factors developed for each fund style. Since we wish to control for as much of the style exposure as possible, we use-reported styles as regressors. 10 As described in Section 3.2 above, decile portfolios are formed based on a fund's lagged quarterly returns in excess of that fund s style s average return. Then for each portfolio, equally-weighted monthly returns are calculated and regressed against a number of 10 In unreported results, we include the option-based returns developed by Agarwal and Naik (2003) as additional independent variables. When these returns are included separately (without including the hedge fund style index returns), the portfolios load significantly on these factors, and the results are consistent with Agarwal and Naik (2003) in that the returns of hedge funds are similar to short positions in put options. However, when the option returns are included in addition to the hedge fund style index returns, they never receive significant loadings. This result occurs because the option returns are fairly highly correlated with a number of the hedge fund styles. Regardless of whether the option returns are included, the intercepts from the regressions (in which we are most interested) are quite similar. 11

12 passive indices, the HML, SMB, and MOM factors, and a number of hedge fund indices (style factors) using Equation (1), above. Results from this regression are in Table 2. The first column shows the monthly excess-of-risk-free rate returns and standard deviations for the formation period portfolios. The second column shows the monthly excess-of-risk-free rate returns and standard deviations for the lagged decile portfolios. For the lagged decile portfolios, average monthly returns are fairly monotonic, increasing from -0.66% for decile 1 (worst) to 0.49% for decile 10 (best). Standard deviation is highest among the best and worst deciles and lowest in the middle deciles. Examining the alphas (intercepts) from the regressions, the intercept on the best minus worst (10-1) portfolio is positive, but not statistically significant. In addition to the alphas, the coefficients from the most statistically significant independent variables are shown. This analysis provides evidence that once common risk and style factors are considered, there is no evidence of quarterly persistence. This result is in direct contrast with the results of other hedge fund studies, which find some persistence at the quarterly level. There are at least two reasons why the results of this paper contrast with theirs. First, most previous studies examine individual fund performance and define persistence as a fund s being in the top half of returns for two consecutive periods. This paper sets a more difficult standard for persistence, requiring that funds be in the top 10% of performers (rather than the top 50%) for two consecutive periods. Second, while the other studies control for common risk factors, they do not control for style in the same way as in this paper. To control for style effects, other studies compare fund returns in excess of style average, but do not adjust these net returns for exposures to other style indices. To examine the incremental effect that controlling for style indices has on the ability to find persistence, I re-perform the above analysis controlling only for common risk factors (and not style indices). With this approach (in unreported results), I do find evidence of performance persistence at the quarterly (but not longer) time horizon. Thus, this paper provides evidence that style factors account for much of the persistence found in prior studies. 12

13 5. Manager tenure as a predictor of persistence While the results of the previous section indicate that prior research findings of quarterly performance persistence can be largely explained by omitted style factors, there is another factor systematically linked to performance that has been ignored in tests of performance persistence. This factor is manager tenure the length of time that a manager has been overseeing his fund. Boyson (2003) shows that, controlling for common risk and style factors, manager tenure is related to performance. Specifically, more experienced (e.g., older or higher tenure) managers underperform less experienced (e.g., younger or lower tenure) managers by approximately -0.75% for each year of tenure. This difference is both economically and statistically significant. If young managers are more skilled than old, this result should be of use in designing a more powerful test of performance persistence. The idea is that young managers with good returns likely achieved those returns due to skill, while old managers with good returns have a higher probability of having achieved those returns due to good luck. The converse also should hold: young managers with poor returns likely experienced bad luck, while old managers with poor returns are likely to have experienced those returns due to lack of skill. Thus, a persistence test that selects funds based both on past performance and on manager tenure should be able to detect persistence. Thus, this result is used to design the following test of performance persistence. Funds are sorted into thirds based on two factors: first, they are sorted into thirds based on prior period returns, and then these portfolios are again sorted into thirds based on manager tenure. These sorts result in nine portfolios, ranging from old, past poor performers (portfolio 1, the worst portfolio) to young, past good performers (portfolio 9, the best portfolio). As before, a portfolio that is long the best portfolio (young, past good) and short the worst portfolio (old, past bad), is formed (this is referred to as the 9-1 portfolio). Again as before, these portfolios are held for the three months (six months, one year), and equal-weighted portfolio returns are calculated for each of the three months (six months, one year). Every three months, portfolios are re-formed. This yields a time-series of 81 (78, 72) monthly returns for each portfolio (the first three (six, twelve) months are used in the initial formation period). 13

14 The quarterly persistence results are in Table 3. The intercept from the 9-1 portfolio is positive and significant at the 5% level (t-value = 2.05). The annualized excess return from investing in this portfolio is about 9%/year, which is economically significant as well. To conserve space, this table shows coefficients and related t-statistics only for the dependent variables which are statistically significant in at least one of the portfolio regressions. While investing in the 9-1 portfolio results in significant quarterly persistence, there is no persistence at the semi-annual and annual levels (which is consistent with prior research). An examination of the coefficients on the explanatory variables indicates some interesting patterns in the data. First, the worst portfolios load positively and significantly on the value-weighted CRSP index, while the best portfolios do not have significant exposure to this factor. This could be interpreted as herding behavior by the worst (and oldest) managers, which is consistent with the findings of Boyson (2003). Additionally, the best portfolios have positive exposure to the currency and commodity indices, while the worst portfolios have negative exposure to these indices. Also, the worst portfolios have negative exposure to bond indices, while the best have positive exposure. Finally, style plays an important role in explaining the return differences; managed futures appear to have been out of favor during the time frame, while the long-short equity style was very successful during this time. Thus, it appears that the best managers were successful in both short and long positions in the equity market (as evidenced by their negligible exposure to the VW CRSP index) while the worst managers had a more pronounced long exposure (as evidenced by their significant exposure to the VW CRSP index and insignificant exposure to the long/short equity style index). While there is evidence of quarterly performance persistence, it appears that poor performance among old, past bad managers is driving this result. The net annualized return of 9% for the 9-1 portfolio attributes about -5.5% to the poor performance of the old, past bad managers (which is statistically significant at the 10% level) and about +3.5% to the good performance of the young, past good managers (which is not statistically significant at conventional levels). Thus, there appears to be an asymmetry in persistence: old, past bad managers continue to perform quite badly, while young, past good managers continue to perform fairly well, although at levels that are not statistically significant from zero. Due to 14

15 the lack of statistical significance, it is probably most accurate to say that while young, past good managers may not continue their past good performance, the are at least able to avoid future poor performance. The next section investigates the likely cause of this asymmetry in more detail. 6. Why do old, past bad returns persist? The persistence test in Section 5 indicates persistence at the quarterly level when funds are selected for investment based on both manager tenure and past performance. This persistence is concentrated among old, past poor performers. This section investigates the likely cause of this continued poor performance. We consider the following hypothesis: young managers are fired more often than old for poor performance. If this is true, then old managers with past poor performance are less likely than young to fail, and thus are more likely to show (poor) performance persistence. This idea comes from models that relate termination to a learning process where investors learn about a manager s ability over time. (See, for example, Jovanovic (1979), Zwiebel (1995), and Holmstrom (1999)). Early in a manager s career, when his reputation is not well-established, investors put more weight on his most recent performance (in the case of a hedge fund manager, his most recently reported monthly return). Eventually, as his reputation becomes more established, each subsequent monthly return has less and less impact on his assessed reputation. The implication of this process is that the sensitivity of termination to the most recent performance evaluation should decrease over time as managers gain reputation. Another reason is noted by Chevalier and Ellison (1999b): since more experienced managers have survived a selection process, the market s assessment of their abilities may be further away from the threshold level at which it becomes efficient to fire the manager. Thus, we examine whether the termination process is more performance sensitive for young managers. In studying this relationship, we repeat and augment the analysis of Boyson (2003), who performs an unconditional survival test and shows that age and manager ability are both positively related to the likelihood of a manager s survival. Here, we extend her work by performing a conditional survival analysis. 15

16 In this analysis, we follow her approach, which uses a time-varying proportional hazards model to study the relationship between manager termination and a number of factors. Intuitively, this model examines each hedge fund that fails (one per time period) and compares its explanatory variables to the explanatory variables on the set of hedge funds that could have failed during the period but did not. If the values of the explanatory variables for those that failed differ from the values of the explanatory variables for those that survived, the coefficients will be significantly different from zero. Time-varying proportional hazards models (which are a category of the more general hazard functions) have several advantages over the more commonly-used probit and logit models. First, they put fewer distributional assumptions on the data; second, they calculate the conditional rather than the absolute probability of failure (conditional upon not having failed in a prior period); and finally, they do not introduce sample-selection bias into the data. Instead of using annual failure rates, the more flexible proportional-hazards model allows for more frequent failure times which reduces bias and adds precision to the estimates. 11 Table 4 performs a number of specifications of the model. Panel A reports results from unconditional regression specifications that include as dependent variables a number of lagged quarterly returns and the manager tenure variable. For ease of interpretation, the coefficient on the manager tenure variable is annualized. In the table, a negative coefficient implies a positive probability of survival, while a positive coefficient implies a positive probability of failure. The results are consistent with Boyson (2003) who uses a smaller, though similar, sample: both current and past returns, as well as manager tenure, are strongly negatively related to failure. All else equal, good funds and old managers are more likely to survive than bad funds and young managers. Panel B begins the first conditional analysis; in this case, conditional upon past performance. For certain of the portfolios described in Section 5, dummy variables are created as follows. The first regression (Column 1) models the probability of survival by tenure, given that the fund s past performance was in the bottom third of returns. This 11 For details on the model and a thorough description of the estimation process, see Boyson (2003). For technical details, see Cox (1972) and Kalbfleisch and Prentice (1980). Finally, for related finance/economics literature that uses this methodology, see Helwege (1996), Lunde, Timmerman, and Blake (1999), and Brown, Goetzmann, and Park (2001). 16

17 category corresponds to portfolios 1,2, and 3 from Table 3. If a fund is in portfolio 1 (high tenure managers with poor past performance) it is assigned a value of one (1). If a fund is in portfolios 2 or 3 (middle tenure managers with poor past performance, and short tenure managers with poor past performance, respectively), it is assigned a value of zero (0). Since this analysis is focusing on past poor performers, portfolios 4-9 are excluded. Hence, the proportional hazards model is comparing the probability of failure for managers in portfolio 1 (old and bad) to the probability of failure for managers in portfolios 2 and 3 (middle tenure and bad, and young tenure and bad, respectively). The negative and statistically significant coefficient on the dummy variable indicates that old and bad funds have a 26.5% lower probability of failure due to poor performance than do young and middle tenure bad funds. The next column repeats the analysis of the first column, this time modeling the probability of failure for managers in portfolio 4 from Table 3 (middle third of performance with high tenure) against the probability of failure for managers in portfolios 5 and 6 (middle third of performance with middle tenure, and middle third of performance with low tenure, respectively). This time, portfolios 1-3 and 7-9 are excluded from the analysis. The results are similar to those in column 1: the negative and statistically significant coefficient on this variable indicates that old and middle-third performing funds have a 36% lower probability of failure than do young and middle tenure, middle-third performing funds. The final analysis is performed in column 3, which models the probability of failure for managers in portfolio 7 from Table 3 (top third of performance with high tenure) against the probability of failure for managers in portfolios 8 and 9 (top third of performance with middle tenure, and top third of performance with short tenure, respectively). In this case, the coefficient is positive but not statistically significant. This indicates that for good managers, failure probabilities do not vary systematically by manager tenure. To summarize the results of Panel B, termination is much more performance-sensitive for young than for old managers. Young managers must perform in the top third of all managers to significantly reduce their likelihood of being terminated. Next, Panel C of Table 4 estimates another set of survival functions, this time conditional upon manager tenure. From Panel B, it is clear that the termination relationship is highly performance-sensitive for young managers. Panel C provides additional evidence that 17

18 supports this result. Specifically, column 1 examines the relationship between termination and performance, conditional upon being a young manager. In this case, the managers in column 9 (the young, past good performers) are assigned a dummy variable value of one (1), while the managers in columns 3 and 6 (the young, past poor performers and the young, past middle third performers) are assigned a dummy variable value of zero (0). This analysis models the probability of fund failure for young managers in the top third of performance against the probability of failure for young managers in the bottom two-thirds of performance. The negative and significant coefficient indicates that young, past good performers are about 45% more likely to survive than young managers in the bottom two-thirds of performance. The last analysis in Panel C examines only the old managers (columns 1, 4, and 7 of Table 3) for differences in termination probability that are related to performance. In this case, managers in column 7 (old tenure and good past performance) are assigned a dummy variable of one (1), while managers in columns 1 and 4 (old tenure and bottom two-thirds of performance) are assigned a dummy variable of zero (0)). The statistically insignificant coefficient on this variable indicates that there are not differences in termination probabilities related to performance for older managers. Thus, the results of Panel C support the results of Panel B above: young managers can increase their probability of survival significantly by being in the top third of performers, while among old managers performance is unrelated to the likelihood of survival. As noted above, these results provide evidence that fund survival is much more performance-sensitive for young than for old managers. This asymmetry in the terminationperformance relationship drives the result from the previous section that performance persistence is concentrated among old, past poor performers. Old managers survive more often, regardless of performance. Thus, there is a greater likelihood of seeing persistence among old, past poor performers (since they are unlikely to drop out of the sample) than among young, past good performers (since these managers have to continue to have very strong performance in order to survive). 18

19 7. Relationship of termination and performance persistence to managerial career concerns This section relates the results of Sections 3.5 and 3.6 to previous hedge fund and mutual fund literature regarding the effect of a manager s reputational concerns (or his career concerns ) on his behavior and ultimately, on his performance. The career concerns literature discusses the idea that a manager s concern for keeping his current job or obtaining a better job can mitigate potential agency problems which occur as a result of misaligned incentives between managers and investors. See Fama (1980). It is reasonable to think that these concerns will change over a manager s career and affect his behavior, specifically his propensity to take on risk. For example, Chevalier and Ellison (1999b) study the behavior of mutual fund managers, and show that these managers increase risk-taking behavior as their careers progress (they tend to increase idiosyncratic risk and mimic other managers less (or herd less)). Their explanation is that there are implicit incentives in the mutual fund industry which relate to a manager s likelihood of losing his job that cause young managers to be more risk-averse than old. As evidence, they model the termination/performance relationship and show that for excess performance below zero, termination is more likely for young than for old managers. However, for excess performance above zero, termination rates do not differ among young and old managers. Additionally, the termination/performance relationship is fairly flat at all levels of return for old managers: termination is much less performance dependent for old managers. Thus, the implicit incentives are clear: a young manager that wishes to avoid termination will avoid risks that could lead to negative excess performance. That is, he will avoid unsystematic risk and herd with other managers. And, since termination is not dependent on performance for old managers, these managers will take on more risk (they will herd less) in order to increase the possibility that their returns will end up in positive territory. By contrast, Boyson (2003) argues the opposing case for hedge fund managers. She shows that hedge fund managers reduce volatility risk and herd more as they age. She attributes this behavior to career concerns that increase over time in the hedge fund industry (by contrast, Chevalier and Ellison (1999b) argue that career concerns decrease over time in 19

20 the mutual fund industry). Specifically, she argues that hedge fund managers have more to lose in terms of reputation, personal capital, and current income than do hedge fund managers, should their funds fail. One piece of evidence that supports this argument is the empirical regularity cited in Brown, Goetzmann, and Park (2001) that a failed hedge fund manager is unlikely to start a successful hedge fund in the future. By contrast, mutual fund managers that fail often are hired by other fund companies. Other support for her argument is that hedge fund managers have tremendously high salaries (often 10 or more times that of a comparable mutual fund manager), and have their own capital invested in their funds. These factors all imply that failure would be much less desirable for old than for young managers: their funds tend to be larger, so their incomes and personal capital invested are higher, and finding another job would result in a much larger pay cut than for a younger manager with a smaller fund. Next, Boyson (2003) shows that, all else equal, risk-taking behavior leads to termination. Both the existence of high career concerns and the higher probability of termination for risk-taking behavior motivate old managers to reduce their risk-taking behavior, which they do. Finally, she links this reduction in risk-taking behavior to lower returns for old hedge fund managers: since old managers take on less volatility risk and herd more, this behavior results in lower returns for these managers. Thus, the different results of Boyson (2003) and Chevalier and Ellison (1999b) can be reconciled by differences in career concerns in the hedge fund and mutual fund industries. These results are also consistent with the finding of Section 5 of this essay that old, past poor performers have persistently worse returns than young, past good performers. However, there is one finding of this current essay that is on the surface, puzzling. In Section 6 we show that young managers are more likely to be terminated for poor performance than are old. At first glance, this result looks very similar to that of Chevalier and Ellison (1999b): they also find that termination is more performance-sensitive for young than for old managers. But a closer look at these results shows a very important difference: while for young mutual fund managers, termination is only more performance-sensitive for excess returns below zero, for young hedge fund managers, termination is more performance-sensitive at a larger range of returns: specifically, if a young manager s returns are not in the top third of all returns, he is 20

21 significantly more likely to fail than an old manager. Thus, the bar is set much higher in the hedge fund industry: young managers must strive to be in the top third of returns in order to continue in the industry. This finding suggests the following story, which is completely consistent with Boyson (2003): for young hedge fund managers who wish to continue in the industry, they must achieve returns higher than one-third of all managers. To maximize their probability of doing this, they must take on more volatility risk and herd less than other managers. Boyson (2003) shows that this type of risk-taking behavior is indeed associated with higher returns. Old hedge fund managers who wish to survive in the industry face a much lower hurdle. Since survival rates for old managers do not vary much with performance (see Section 6), they have strong incentives to herd and reduce risk so as to avoid the possibility of a very bad return that would increase their probability of failure (and would likely result in a loss of personal assets as well). Thus, young managers strive for high returns by taking on more risk, while old managers strive for average (or even below average) returns by taking on less risk. 8. Conclusions This paper designs a more powerful test for performance persistence that is able to detect quarterly persistence among hedge fund managers. This test is motivated by the results of Boyson (2003), who shows that young managers outperform old, on average, which implies that for young managers, good results are driven by skill, while for old managers, good results are driven by luck. If young managers are more skilled, they should show persistence. And this is indeed true: at the quarterly time horizon, young, past good managers outperform old, past poor managers by about 9%/year. This result is driven primarily by the propensity of old, past poor managers to continue underperforming. Additionally, we perform a survival analysis to investigate the result that old, past poor managers have persistence underperformance, and find the following asymmetric relationship. Young managers must perform in the top third of all managers to have survival probabilities that are the same as those of old managers. That is, young managers are punished (by fund failure) significantly more often than old if their returns are in the bottom 21