Mutual Fund Performance Examining the Predictive Power of Fund Characteristics

Transcription

1 Mutual Fund Performance Examining the Predictive Power of Fund Characteristics Roland Lu Northwestern University MMSS Senior Thesis 2015 Advisor: Dimitris Papanikolaou Instructor: Joseph Ferrie

2 Acknowledgements I would like to thank Professor Dimitris Papanikolaou for serving as my thesis advisor, providing continued guidance throughout this process. I would also like to thank German Bet for his invaluable assistance in performing my analyses; Professor Joseph Ferrie as the MMSS Thesis Coordinator; and Sarah Muir Ferrer and Professor William Rogerson for their leadership of and support for the MMSS program. Lastly, I would like to thank all of my family, friends and other faculty who have made it an incredible four years here at Northwestern. Abstract This paper investigates the predictive power of various mutual fund characteristics on raw performance for equity mutual funds from January 1970 to December Specifically, the study explores the effects of variability in historical returns, expense ratio, turnover ratio and assets under management on future returns. The regression analyses indicate that higher expense ratios are associated with lower raw performance and that this effect persists even when examining individual decades and when accounting for variability in exogenous factors through a fixed effects model. The results for turnover ratio and assets under management lack the statistical significance and magnitude respectively to be considered meaningful in the context of this study. The results for historical returns provide conflicting results, suggesting the existence of negative performance persistence in the standard model and positive performance persistence when exogenous macroeconomic variability is removed in the fixed effects model. However, both models produce further inconclusive results when examined on a decade specific basis, where we observe consistent variability in signs on the coefficients for historical returns.

3 Contents 1. Introduction The Prediction of Mutual Fund Performance The Model Data and Methodology Results Conclusion References... 31

4 1. Introduction The birth of the mutual fund industry in the U.S. is most commonly traced to the creation of the Massachusetts Investors Trust (MIT), the first U.S. open-ended mutual fund, in 1924 with $50,000 in assets under management (Fink 2008). From these early beginnings, the industry has experienced extraordinary growth and today accounts for over $17 trillion in assets managed as of year-end 2013 a behemoth within the modern investment management universe. Investor and public interest in mutual funds has steadily increased from the 1950s to present day, driven in large parts by innovations in the industry such as money market funds and exchange traded funds as well as the emergence of new distribution channels such as 401(k) and IRA retirement plans. The proliferation of mutual funds, in both number and size, coincides with an increased reliance on mutual funds as a primary destination for investment assets by American households. According to the Investment Company Institute, mutual funds or registered investment companies managed over 22% of U.S. household financial assets at the end of 2013 compared to just 2% of financial assets at the end of 1990 (ICI 2014). In addition, mutual funds have become the cornerstone of retirement plans for most Americans, representing 60% of all invested assets in Defined Contribution (DC) plans including 401(k)s and 45% of all invested assets in Individual Retirement Accounts (IRAs) as of 2013 (ICI 2014). Given the ubiquity and ever-increasing role that mutual funds play with respect to the maintenance and growth of household wealth and retirement plans in the U.S., the performance generated by these mutual fund investments is more important and relevant than ever. Furthermore, the current landscape of struggling and distressed municipal governments plagued by the widespread underfunding of pension obligations necessitates strong returns from the capital invested in mutual funds and other assets. According to The Pew Center on the States, the

5 gap between the states assets and their obligations for public sector retirement benefits totaled over $1.38 trillion in 2010 (Pew 2012) and has only grown since then. Recent high profile municipal bankruptcies include those of Detroit, Michigan and Stockton, California, both of which involved large underfunded pensions. The combination of these factors illustrates the immense value attributable to understanding mutual fund performance and, by extension, the explanatory variables and characteristics that affect performance. This paper seeks to build on previous academic literature that has explored the relationship between various explanatory variables and fund returns (Gruber 1995, Carhart 1997, Shukla 2000, Wermers 2000). The study aims to contribute to the existing body of work in three primary ways. First, it uses updated and expanded Center for Research in Security Prices (CRSP) data from 1970 to 2010 for a larger universe of equity mutual funds, producing more robust results and commentary relevant to market and general environment faced by mutual fund investors today. In addition, it reexamines the explanatory factors put forward by previous academic literature and further scrutinizes their efficacy in the context of revised sample. Lastly, it takes into consideration the potential for omitted variable bias due to heterogeneity in general macroeconomic or market characteristics through the addition of a Fixed Effects model. The remainder of the paper is structured as follows. Section 2 discusses in detail earlier works examining mutual fund returns. Section 3 lays out the construction of the predictive model and selection of factors, while Section 4 walks through the manipulation of the dataset and the steps taken to perform the regression analysis. Finally, Section 5 discusses the results and implications of the regression analysis and Section 6 provides concluding remarks.

6 2. The Prediction of Mutual Fund Performance Mutual fund investors today are continually inundated with recommendations by the financial media as well as from friends and financial advisors, all advocating the attractiveness and value of one mutual fund or another. The proliferation of advice and opinions on mutual fund selection, in combination with the intrinsic importance of identifying mutual funds that generate strong returns, has resulted in the contribution of a significant amount of academic research to the discussion in recent decades. These studies have analyzed a large multitude of fund characteristics and tested their efficacy as explanatory variables in predicting mutual fund returns. Among the characteristics that have been examined are various fees and expenses, trailing or historical returns of various time horizons, turnover ratios and fund size via assets under management. Several of the studies aggregate multiple characteristics in their analysis in an attempt to more comprehensively explain variation in mutual fund performance. A. Expenses The fees charged by mutual funds pay for fund maintenance as well as the various services the fund provides, most importantly portfolio management. In the process of selecting funds, the total expense ratio is often one of the most important characteristics for prospective investors. Fittingly, it has been a particular focus of existing academic literature with regards to its relation to mutual fund performance. Carhart s paper On Persistence in Mutual Fund Performance (1997) examines the effects of heterogeneity in total expense ratios, among other fund characteristics, and its relation to mutual fund performance. Specifically, Carhart demonstrates that expenses have at least a one-for-one negative impact on fund performance as evidenced by the average coefficient on contemporaneous expense ratio with respect to

7 abnormal returns in a univariate cross-sectional regression (Carhart 1997). Carhart s findings present statistically significant evidence of a strong negative relationship between expense ratios and performance and suggest the possibility expenses may be used as a predictive variable for future returns. Gruber explores the usage of expense ratios as a predictive variable for future returns in his paper Another Puzzle: The Growth in Actively Managed Mutual Funds (1996). He ranks the mutual funds in his sample into deciles according to several attributes including expense ratios and calculates the absolute and risk-adjusted returns generated by these decile portfolios on a 1 year and 3 year basis. Gruber finds statistical significance in both the Spearman Rank Coefficients of the deciles between the selection and performance periods as well as in the outperformance of lower expense ratio deciles compared to higher expense ratio deciles (Gruber 1996). These results support the idea that investors are better off buying funds with low expense ratios, though Gruber notes that the magnitude of the differences between deciles (portfolios) reported in the table is not as large as it is for the other forecasters and that even buying the top decile portfolio of funds with the lowest expense ratios would not generated a positive risk adjusted return (Gruber 1996). Shukla similarly tests the predictive ability of expense ratios with respect to performance in his Identifying Superior Performing Mutual Funds (2000). In the study, he performs regressions of raw monthly fund returns on trailing expense ratios with a four quarter lag for equity mutual funds from 1993 to Unlike Gruber, Shukla finds that expense ratios are the single most important characteristic when it comes to future mutual fund returns (Shukla 2000). He draws this conclusion from the 10.68% average R 2 in the 19 quarters for which the single factor regression model is estimated and additionally from the fact that the coefficient in the

8 regression is statistically significant for 18 out of the 19 quarters. Both the R 2 and number of statistically significant coefficients rank first amongst all of the single factor models he tests. In addition, Shukla estimates a multi factor model from which he produces statistically significant coefficients for expense ratios ranging from to over the 19 quarters, suggesting that lower expenses result in improved returns (Shukla 2000). The evidence of a negative relation between the expenses charged by a mutual fund and its performance relative to peers may seem counterintuitive for investors. Gil-Bazo and Ruiz- Verdu attempt to explain this relationship in their study The Relationship between Price and Performance in the Mutual Fund Industry (2009). They propose two explanations for this apparent anomaly. First is that the negative relation is a consequence of factors, which are omitted in univariate regressions, that are both positively correlated with returns and negative correlated with funds operating costs, and thus also with fund fees (Gil-Bazo and Ruiz-Verdu 2009). However, the two find that the negative relationship between fees and performance persists even after controlling for cost determinants, ruling out factors related to operating costs. The second explanation is that funds strategically set fees as a function of their past or expected performance and thus funds with lower expected performance charge higher fees due a combination of (i) having a more inelastic investor base, (ii) targeting performance insensitive investors who accept higher fees, and (iii) increased marketing expenses to attract investors despite poor performance (Gil-Bazo and Ruiz-Verdu 2009). Gil-Bazo and Ruiz-Verdu find that the empirical analysis performed in their study gives credence to all three of the strategic fee pricing alternatives in explaining the negative relationship between fees and performance. B. Historical Returns and Performance Persistence

9 The existence of positive persistence in mutual fund performance over various time periods has been a fiercely debated phenomenon in mutual fund academic literature. Studies have been conducted on raw and risk adjusted returns that have produced evidence both for and against this persistence in performance. In particular, Grinblatt and Titman (1992) and Gruber (1996) perform studies that demonstrate performance persistence is both statistically significant and readily observable. In contrast, Phelps and Detzel (1997) and Jensen (1969) find no comprehensive evidence for persistence in their studies. Irrespective of whether performance persistence truly exists, it seems apparent that historical returns merit additional examination as a predictive factor. Grinblatt and Titman s paper The Persistence of Mutual Fund Performance (1992) analyzes how future performance of mutual funds relates to past performance. In particular, they conduct a modified time series regression on excess returns or alpha of 279 mutual funds existing from 1974 to Grinblatt and Titman find a statistically significant slope coefficient of in the regression of excess returns in the last five years on the returns from the first five years with an R 2 of 0.06 (Grinblatt and Titman 1992). They conclude that there is positive persistence in mutual fund performance and that this persistence cannot be explained by inefficiencies in the benchmark that are related to firm size, dividend yield, past returns, skewness, interest rate sensitivity, or CAPM beta (Grinblatt and Titman 1992). Gruber (1996) similarly studies persistence but does so with respect to both raw and excess returns. Using the decile ranking system described in the Expenses section, he finds statistical significance in the Spearman Rank Coefficients of the deciles between the selection and performance periods for excess returns. Additionally, Gruber finds statistical significance in the differences in performance between top and bottom decile portfolios for both raw and excess returns (Gruber

10 1996). He concludes that future performance is in part predictable from past performance and that excess returns represent the most effective predictive variable in determining both future raw and excess returns (Gruber 1996). Though Phelps and Detzel also initially produce results indicating some degree of positive performance persistence in their paper The Nonpersistence of Mutual Fund Performance (1997), they find that persistence disappears in both the latter periods of their study and in general once they control for additional risk factors (Phelps and Detzel 1997). They use monthly fund return data from 1983 to 1994 for 87 randomly selected debt and equity funds and analyze repeat winners and losers in two year windows using two way tables. Overwhelmingly, the two way tables provide little evidence for positive performance persistence and Phelps and Detzel conclude that they find no reliable evidence of positive performance (Phelps and Detzel 1997). Jensen also finds no evidence of positive persistence performance in his paper Risk, The Pricing of Capital Assets, and the Evaluation of Investment Portfolios (1969). He examines mutual fund performance calculated gross of expenses for the ten year period from Jensen finds that the conditional probabilities of observing positive returns given a run of positive returns in prior years are not significantly different from what would be expected under assumption of independence and thus concludes that there is no striking evidence of superior forecasting ability (Jensen 1969). Jensen further observes that general positive correlation in fund performance observable can be attributed to the fact that poor performing funds continue to perform poorly which is reasonable given the existence of expenses and fees (Jensen).

11 C. Turnover Ratio Underlying the argument in favor of actively managed mutual funds is an implicit assumption that the higher levels of trading activity vis-à-vis passive index funds is borne out of superior stock selection ability. If this is true, we would expect higher turnover ratios to generate higher gross performance. However, the heightened transaction expenses related to higher trading volume reduce net performance and must also be factored into any comprehensive analysis of turnover ratio. Carhart (1997) and Wermers (2000) both attempt to measure the cumulative effect of fund turnover on performance in their respective studies. Carhart estimates monthly univariate cross-sectional regressions of turnover ratios on contemporaneous abnormal returns from 1966 to 1993 using data from The Center for Research in Security Prices (CRSP) in his On Persistence in Mutual Fund Performance (1997). He modifies CRSP reported turnover by adding one-half of the percentage change in total net assets (TNA) adjusted for investment returns and mergers to better capture a more accurate value of fund turnover in a given period. Carhart finds an average coefficient of for turnover ratio that is statistically significant in the vast majority of the included months and concludes that the cross-section tests indicate that turnover reduces performance for the average fund (Carhart 1997). These results suggest that overall stock selection ability amongst mutual fund managers is inadequate in overcoming the increased transaction expenses and that high turnover ratios might be a negative predictive variable for future returns. Wermers builds on Carhart s work by studying whether funds that trade more frequently generate better performance in his paper Mutual Fund Performance: An Empirical Decomposition into Stock-Picking Talent, Style, Transaction Costs, and Expenses (2000) using

12 CRSP data for mutual funds from 1975 to He ranks the mutual funds by their prior year turnover ratio and generates fractile portfolios from which he observes differences in raw performance, both gross and net, and excess performance or alpha. Wermers finds that highturnover funds hold stock portfolios that significantly outperform the portfolios of low-turnover funds and that the outperformance in gross returns is 4.3% per year on average and statistically significant (Wermers 2000). High-turnover funds also incur higher transaction costs compared to low-turnover funds with an average annual difference of 2.4% which somewhat reduces the net return advantage of high-turnover funds, though outperformance still exists and is statistically significant on a net basis (Wermers 2000). With regards to excess performance, he finds no statistically significant differences in performance between high-turnover and low-turnover funds and confirms Carhart s findings that high-turnover funds do not generate better excess returns and thus cannot justify higher trading activity (Wermers 2000). D. Fund Size and Assets under Management As the mutual fund industry continues to experience steady asset inflows, so too have individual funds grown to manage increasingly large amounts of capital. Though there are some obvious advantages to increased size in individual funds, concerns over the scalability of certain fund strategies as well as the incentives resultant of existing fee structures have led to increased scrutiny by investors of the various implications of larger fund sizes. In particular, research attention has been drawn towards examining the relationship between fund size and performance. Chen, Hong, Huang, and Kubik set out to comprehensively investigate the effects of scale on performance for mutual funds in their paper Does Size Erode Mutual Fund Performance?

13 The Role of Liquidity and Organization (2004). Using mutual fund data from 1962 to 1999, they estimate regressions of returns on lagged fund size and further adjust for potential bias by estimating additional multi-factor regressions that remove the effects of style heterogeneity, fund age, turnover, and past performance, among other factors. Chen, Hong, Huang, and Kubik find that a two standard deviation shock in the log of a fund s total assets results in 65 to 96 basis point movement in the fund s future annual return and they note that this is both statistically significant and economically important due to the fact that the universe of funds in their sample underperforms the market portfolio by approximately 96 basis points a year after fees and expenses (Chen, Hong, Huang, and Kubik 2004). They identify these results as strong evidence that fund size erodes performance and additionally that this relationship is not driven by heterogeneity in fund styles, fund size being correlated with other observable fund characteristics, or any type of survivorship bias (Chen, Hong, Huang, and Kubik 2004). Instead, they observe that the negative effect on fund size on performance is most pronounced for mutual funds characterized by small cap strategies, suggesting that liquidity may be the key reason why the relationship exists in the first place. 3. The Model The primary inspiration for our predictive model is drawn from prior analyses performed by Carhart (1997) and Shukla (2000). In both studies, the authors estimate cross-sectional multifactor regressions of realized returns on several mutual fund characteristics to generate coefficients for each of the explanatory characteristics. Carhart s model uses excess realized returns, contemporaneous characteristic information and data on equity mutual funds from 1962 to 1993 which is sourced from a precursor to the CRSP Mutual Fund database. Shukla constructs a model with realized raw returns at time t and one year lagged fund characteristics at t-1using

14 quarterly data for equity mutual funds from 1993 to 1998 retrieved from the Morningstar Principia Database. We note that both Carhart and Shukla s models are simplistic by design; Shukla emphasizes that we want to construct a model that is parsimonious and utilizes readily available information (Shukla 2000). Our model is constructed with similar intent, namely with an emphasis on practicality and accessibility for retail investors weighing their options for investing in mutual funds. We believe that the coefficients estimated for lagged characteristics regressed on fund returns provide more informative value to potential investors than contemporaneous characteristics. Accordingly, we estimate a predictive model using a cross-sectional regression of realized raw returns on several one year lagged characteristics. The general econometric model which we estimate is as follows: 1,,,,,,,,,,, where, is the realized raw return of fund j in year t and,, is the one-year lagged value of attribute k on fund j. The coefficient, s, which reveal the nature and strength of the relationship between each individual characteristic k and raw returns, are estimated from the regression model. We also give consideration to the potential for bias resultant of heterogeneity in general macroeconomic and stock market characteristics and its substantial influence on the performance of equity mutual funds. We address the possibility for omitted variable bias by conducting further analyses in which we generate dummy variables for each year in the sample and estimate a Fixed Effects regression. The Fixed Effects regression produces characteristic coefficients that

15 are unaffected by year to year variability in factors beyond the fund attributes that we are attempting to examine. The fixed effects model that we estimate is given by: 2,,,,,,,,,,,, 1970, 1971, 1972, 2008, 2009 where the variables are as described in equation (1) with the addition of the dummy variables for years 1970 through 2009 and their corresponding coefficients,, which represent the effect of year-to-year heterogeneity in macroeconomic and market environment on fund returns. 4. Data and Methodology A. Data The source for all of our mutual fund data is the Center for Research in Security Prices (CRSP) at The University of Chicago Booth School of Business. In particular, we employ use of the CRSP Survivor-Bias-Free US Mutual Fund Database, also used in Carhart (1997) and Wermers (2000), which provides extensive historical information for approximately 30,000 US open-ended funds from 1962 to 2014 and is generally considered the most comprehensive database for both active and inactive mutual funds today. Included are mutual funds of all strategies, including equity funds, taxable and municipal bond funds, international funds, and money market funds, among others. For each of these mutual funds, the CRSP database contains data points on a broad range of characteristics such as fund name histories, investment strategies, fee structures, holdings, asset allocation, daily and monthly net returns, daily and monthly net asset values (NAVs), monthly total net assets (TNA), and dividends and distributions. All of the

16 data items begin at varying times between 1962 and 2008 depending on variability and are subject to ongoing efforts to check and improve data quality by CRSP. We choose to narrow the CRSP data used in our study from January 1970 to December 2009, which represents all of the data available for complete decades. In addition, we limit the mutual funds examined to exclusively equity mutual funds as defined by the CRSP Style Code. Given that no single source for investment style exists for the full time period from 1962 to present, the CRSP Style Code aggregates style and objective designations from three distinctive and non-overlapping classifications: Wiesenberger Objective Codes ( ), Strategic Insight Objective Codes ( ), and Lipper Objective Codes (1998 Present). By only using equity mutual fund data, we are able to address the likely differences in fund returns and predictive fund characteristics between mutual funds with different asset allocations and styles namely between equity funds and fixed income or money market funds. The dataset used in our study therefore consists of returns and fund characteristics for U.S. equity mutual funds from 1970 to B. Analysis Methodology In conducting our analysis, we first choose and, if necessary, construct the relevant fund characteristics necessary to estimate the return models (1) and (2) detailed in Section 3. Given the frequency of attribute data provided by CRSP, it is necessary that we use annual information for both returns and fund characteristics in estimating the regression. Characteristic selection for testing and further examination is driven by prior research examining the effects of various characteristics on mutual fund performance discussed in Section 2. Specifically, we identify

17 expense ratio, trailing returns, turnover ratio and assets under management as fund characteristics that might affect returns. For expense ratio, we use exp_ratio as reported by the CRSP database, which is calculated as the ratio of total investment that shareholders pay for the fund s operating expenses including 12b-1 fees and any waivers or reimbursements for the most recent fiscal year. We generate two new variables for trailing returns, namely one year annual return and three year annual compound return. We calculate annualized returns using the monthly returns provided by the CRSP US Mutual Fund Database and name the variable annreturn. Whereas many previous studies focus on performance persistence on a one year basis (Jensen 1969, Greenblatt and Titman 1992, Phelps and Detzel 1992, Carhart 1997), several others also test for longer term persistence (Gruber 1996, Shukla 2000). Both Gruber and Shukla use three year trailing returns as a proxy for longer term trailing performance and we proceed in similar fashion with our study. Rather than use the three year average annual return (AAR) for mutual funds sometimes reported by various financial resources, we calculate the three year annual compound return using a geometric mean and name the variable annreturn_3y. The calculation is as follows: Like expense ratio, turnover ratio is also taken directly from the CRSP database. turn_ratio is defined as the minimum of aggregated sales or aggregated purchase of securities divided by the average 12-month Total Net Assets (TNA) of the fund. Lastly, we describe assets under management using the natural log of Total Net Assets (TNA), which is defined as the total value of a fund s portfolio less any accrued liabilities. The CRSP database includes mtna, the Total Net Assets of a fund at month s end; we use the December month end values of mtna as

18 the annual TNA. Given the large nominal values and positive skew of the variable, we take the natural log of CRSP reported mtna to produce our variable, ln_tna. We note that Carhart (1997) similarly uses the natural log of TNA in estimating cross-sectional regressions of fund returns in his analysis. With these characteristic variables, we proceed to estimate the cross-sectional regression models (1) and (2) detailed in Section 3. First, we estimate single factor regressions of annual returns on one year lagged values of each of the individual characteristics. These univariate regressions provide us with a simplified interpretation of the effects each of the five characteristics one year return, three year compound annual return, expense ratio, turnover ratio, and total net assets on one year forward fund returns. Next, we estimate several variations of the multivariate cross sectional regression model (1) described in Section 3. Specifically, we estimate the following regressions: (4A),,,,,,,,,,,, _,, (4B),,, 3,,,,,,,,, _,, (4C),,,,,, 3,,,,,,,, (5A),,,,,, 3,,,,,,,,, _,,

19 Model 4A represents a four factor model using only trailing one year annual returns to test performance persistence, which is the most common method in the prior research discussed in Section 2. Model 4B is a similar four factor model that uses trailing three year compound annual returns to test longer term performance persistence. Model 4C includes both trailing one year annual returns and trailing three year compound annual returns to measure the effects of both shorter and longer term persistence but omits total assets, which several previous studies including Carhart (1997) and Shukla (2000) find to be insignificantly related to returns. Finally, model 5A combines one year lagged values of all of the characteristic variables in a five factor model. All of the regressions are estimated with annual returns as the dependent variable. We note here that the existing literature summarized in Section 3 generally analyzes mutual fund datasets with return and characteristic information for ten to twenty years (Jensen 1969, Greenblatt and Titman 1992, Gruber 1996, Phelps and Detzel 1997, Shukla 2000, Wermers 2000) and further that the time period of these datasets varies widely. Given that our dataset spans four decades from 1970 to 2009, we decide to further estimate decade specific regressions for model 5A to check for any heterogeneity in predicted coefficients between the decades. From the results produced by our estimates of decade specific regressions which are detailed in Section 5 below, we elect to conduct additional analysis and perform further regressions of the four models above adjusted to reflect the effect of changes in general macroeconomic conditions and market dynamics on fund returns. We do so by estimating the fixed effects model (2) with individual dummy variables for each year from 1970 to 2009 as detailed in Section 3. The dummy variables allow us to control for changes in market returns, liquidity, and macroeconomic trends and more accurately isolate the effects of the relevant fund

20 characteristics on returns. We estimate a fixed effects version for all four of the multivariate cross sectional models 4A, 4B, 4C, and 5A as well as for the decade specific regressions of model 5A. 5. Results A. Single Factor Regressions on Individual Characteristics The results of the single factor regressions for each fund characteristic are displayed in Table 1. The coefficients on annreturn and annreturn3y are and respectively and both statistically significant at the 1% level. They suggest that a one percentage point increase in returns in the prior year predicts lower performance by roughly 19 basis points while a one percentage point increase in the trailing three year compound annual return predicts lower performance by roughly 51 basis points. The results provide little evidence in support of positive performance persistence, reinforcing the conclusions reached in studies by Jensen (1969) and Phelps and Detzel (1997). Both expratio and turnratio are estimated to have a negative relationship with future returns, which reconciles with prior analysis performed by Carhart (1997) on expense ratios and turnover ratios. However, the coefficients for the two single factor regressions are not statistically significant, even at the 10% level, so we are unable to confidently interpret these results. Lastly, we estimate a coefficient of on ln_tna that is statistically significant at the 1% level, indicating a negative relationship between fund size and future returns. The coefficient suggests that a one percent increase in Total Net Assets reduces returns by approximately 0.29 basis points. This result corroborates the findings in Chen, Hong, Huang and Kubik (2004) who conclude that fund size erodes performance.

21 Table 1. Single Factor Regression Results Dependent Variable: Mutual Fund Annual Return Period: January 1970 December 2009 Variables 1A 1B 1C 1D 1E annreturn *** (0.0030) annreturn3y *** (0.0067) expratio (0.0925) turnratio (0.0004) ln_tna *** (0.0003) Constant *** *** *** *** *** (0.0008) (0.0010) (0.0016) (0.0009) (0.0012) Observations 113,609 82, , , ,160 R squared Note: All independent variables are one year lagged values (t 1) Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 B. Multivariate Regression Models Table 2 shows the regression results for the four multivariate models 4A, 4B, 4C, and 5A described in Section 4. 4A, which uses only one year lagged returns to measure performance persistence, estimates coefficients similar to those generated by the single factor regressions above. Namely, the coefficients on annreturn, expratio, turnratio, and ln_tna have the same directional sign, statistical significance (with the exception of turnratio which is now significant at the 5% level) and similar magnitudes as their single factor regression counterparts. Model 4B replaces one year lagged returns with lagged three year compound annual returns to

22 represent performance persistence, resulting in several significant changes to the estimated coefficients. All of the coefficients estimated in this model are statistically significant at the 1% level besides for ln_tna, which is insignificant even at the 10% level. The directional signs for the coefficients on annreturn3y, expratio, and turnratio are all consistent with those estimated in the single factor regressions and model 4A while the magnitudes on both performance persistence and expense ratios increase considerably. A one percentage point increase in the three year compound annual return or the expense ratio is associated with a 52 basis point and 79 basis point reduction in future annual returns, respectively. The larger estimated coefficient for lagged three year compound returns compared to one year returns reconciles with Gruber s findings of increased predictive power of three year returns, though we note that he uses excess returns rather than raw returns (Gruber 1996). Furthermore, model 4B exhibits larger predictive power than 4A as indicated through its higher R 2 value (0.068 versus 0.036). In model 4C, we incorporate both of the independent variables for performance persistence and drop ln_tna in an attempt to improve upon the results of 4A and 4B. As displayed in Table 2, all of the coefficients estimated by this model are statistically significant at the 1% level. The coefficient values are similar to those in 4B with the negative effect of performance persistence now split between one year returns and three year compound annual returns, as expected. Moreover, model 4C contains even higher predictive power as measured by its R 2 of Lastly, 5A estimates a five factor model incorporating both of the performance persistence variables as well as the three other characteristics of interest. The coefficients are all statistically significant at the 1% level besides ln_tna and are very similar in magnitude to those estimated in model 4C. The model has an R 2 of 0.072, also similar to that of 4C, suggesting very little incremental predictive power in including ln_tna.

23 Table 2. Multivariate Regression Model Results Dependent Variable: Mutual Fund Annual Return Period: January 1970 December 2009 Variables 4A 4B 4C 5A annreturn *** *** *** (0.0031) (0.0044) (0.0044) annreturn3y *** *** *** (0.0070) (0.0086) (0.0086) expratio *** *** *** (0.1043) (0.1125) (0.1071) (0.1125) turnratio ** *** *** *** (0.0004) (0.0007) (0.0007) (0.0007) ln_tna *** (0.0004) (0.0004) (0.0004) Constant *** *** *** *** (0.0025) (0.0030) (0.0020) (0.0030) Observations 104,895 78,629 79,221 78,629 R squared Note: All independent variables are one year lagged values (t 1) Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 C. Decade Specific Regression Models In Section 4, we discuss the possibility that heterogeneity in factors exogenous to our models might explain differences in coefficient estimates among previous studies that used varied datasets. Accordingly, we estimate decade specific regressions for model 5A, the results of which can be found in Table 3. It is readily apparent that significant variation occurs in the estimated coefficients of certain characteristics between decades. These differences manifest themselves most meaningfully in coefficients for the two characteristics measuring performance persistence, annreturn and annreturn3y. In the four decades in which we estimate the regression,

24 the coefficient on annreturn is twice positive and twice negative and three of these four estimates are statistically significant at the 1% level. The results reaffirm the conflicting conclusions regarding positive performance persistence found in previous studies. The coefficients for annreturn3y similarly change signs, though the three statistically significant negative coefficients point toward negative performance persistence as measured using three year compound annual returns. For the remaining three characteristics, the statistical significance of the coefficients varies by both variable and decade. Though the results are not uniform in statistical significance or direction between decades, we conclude that ln_tna and expratio exhibit negative relationships with returns while the relationship between returns and turnratio is indeterminate. Table 3. Decade Specific Regression Results Dependent Variable: Mutual Fund Annual Return Period: January 1970 December 2009 Variables 1970s 1980s 1990s 2000s annreturn *** *** *** (0.0312) (0.0231) (0.0163) (0.0045) annreturn3y *** ** *** *** (0.0536) (0.0456) (0.0318) (0.0089) expratio ** * *** (1.8642) (0.9662) (0.3154) (0.1165) turnratio *** *** (0.0121) (0.0049) (0.0028) (0.0007) ln_tna *** *** *** (0.0066) (0.0031) (0.0013) (0.0004) Constant *** *** *** *** (0.0397) (0.0223) (0.0098) (0.0031) Observations 1,124 2,218 11,835 63,452 R squared Note: All independent variables are one year lagged values (t 1) Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1

25 D. Fixed Effects Regression Models Table 4. Fixed Effects Multivariate Regression Results Dependent Variable: Mutual Fund Annual Return Period: January 1970 December 2009 Variables 4A 4B 4C 5A annreturn *** *** *** (0.0030) (0.0042) (0.0042) annreturn3y *** (0.0059) (0.0070) (0.0070) expratio *** *** *** *** (0.0580) (0.0617) (0.0585) (0.0617) turnratio (0.0002) (0.0004) (0.0004) (0.0004) ln_tna *** *** *** (0.0002) (0.0002) (0.0002) Constant *** *** *** *** (0.0014) (0.0017) (0.0011) (0.0017) Observations 104,895 78,629 79,221 78,629 R squared Note: All independent variables are one year lagged values (t 1) Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 Given the results observed in the decade specific regressions, we estimate a series of fixed effects models in an attempt to remove the effects of heterogeneity in exogenous factors, such as general macroeconomic conditions and market dynamics, on future returns. The regression results for the fixed effects versions of the four multivariate models are displayed in Table 4 above. Table 5 below contains the results of the decade specific fixed effects regressions of model 5A. The inclusion of dummy variables for each year in the model (omitted in the regression output) leads to several significant changes in the estimated coefficients of the four models. The largest effect occurs for estimates of the relationship between past returns and future

26 returns (performance persistence). Whereas the coefficients estimated in the single factor, multivariate, and decade specific models all suggest a negative or indeterminate relationship between past returns and future returns, the coefficients on annreturn and annreturn3y estimated in the fixed effects versions of models 4A, 4B, 4C, and 5A all have positive signs. Statistical significance is stronger for annreturn, with all coefficients significant at the 1% level, compared to annreturn3y for which only one of the three estimated coefficients is statistically significant even at the 10% level. These results provide evidence in favor of the existence of performance persistence once we account for year-to-year variability in exogenous factors and support the conclusions reached by Greenblatt and Titman (1992) and Gruber (1996) in previous studies. Table 5. Fixed Effects Decade Specific Regression Results Dependent Variable: Mutual Fund Annual Return Period: January 1970 December 2009 Variables 1970s 1980s 1990s 2000s annreturn *** *** *** (0.0303) (0.0260) (0.0168) (0.0040) annreturn3y *** *** *** *** (0.0460) (0.0400) (0.0312) (0.0066) expratio *** *** *** (0.7955) (0.6534) (0.2774) (0.0570) turnratio ** ** *** *** (0.0052) (0.0032) (0.0024) (0.0004) ln_tna *** *** (0.0028) (0.0022) (0.0011) (0.0002) Constant *** *** *** *** (0.0170) (0.0151) (0.0087) (0.0015) Observations 1,124 2,218 11,835 63,452 R squared Note: All independent variables are one year lagged values (t 1) Standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1

27 Though the changes in coefficient estimates using a fixed effects model are most pronounced in the two variables related to performance persistence, the remaining characteristic variables estimated coefficients are also altered under a fixed effects model. The coefficients for expratio remain uniformly negative in the four models but slightly decrease in magnitude and are now all statistically significant at the 1% level. Turnratio loses essentially all of its predictive power as the coefficients estimated in the four models are all roughly zero and statistically insignificant at the 10% level. The estimates for ln_tna indicate a negative relationship between size and future performance that is statistically significant at the 1% level, though the effect is muted: a one percent increase in Total Net Assets (TNA) predicts a decrease in future returns by roughly 1 basis point in all three relevant models. In general, our prior conclusions on the relationships between expratio, turnratio, and ln_tna and future returns are largely corroborated by the results of the fixed effects regressions. The notable exception is performance persistence as measured through annreturn and annreturn3y, for which the fixed effects results contradict our previous conclusions of the non-existence of persistence performance. We hypothesize that year-to-year variation in fund returns is strongly correlated with contemporaneous variation in macroeconomic conditions and broader market returns and that once we account for the heterogeneity in the latter through a fixed effects model, we are able to separate out and identify the existence of performance persistence. This is somewhat supported by the R 2 values of the fixed effects models, ranging from to 0.727, which increase dramatically when compared to the general multivariate models. This indicates that a very substantial amount of the variance in current fund returns is explained by factors outside of our models, which seems reasonable given our focus on fund specific characteristics and omission of broader variables such as market returns and macroeconomic indicators.

28 For the sake of comparison, we also estimate decade specific regressions for the fixed effects version of model 5A. As with the fixed effects estimates for the four models above, the decade specific fixed effects regressions produce results largely in line with previous results for expratio, turnratio, and ln_tna. Some small changes do occur such as the coefficients for turnratio becoming more statistically significant (significant at the 5% level in all four decades compared to just two decades in the standard regression). In another similarity with the fixed effects regressions of the four models, the decade specific fixed effects regressions also provide evidence in support of performance persistence in opposition to their non-fixed effects counterparts, with six out of eight coefficients associated with performance persistence exhibiting a positive sign. However, the result is not as robust with only six of eight coefficients statistically significant at the 1% level and two of the statistically significant coefficients carrying a negative sign. This variation in coefficient signs lends further support to the conflicting results found in the normal decade specific regressions and suggests that variability in positive performance persistence exists even after accounting for variability in exogenous factors in the model. We also note the dramatic differences in R 2 values for the same model across decades, with a low value of in the 1990s to a high value of in the 1970s. 6. Conclusion This paper set out to examine the predictive power of various fund characteristics on future returns using an expanded and updated dataset. Our findings support and reaffirm previous conclusions reached for some characteristics and contradict them for others. Specifically, our study finds that expense ratios have a negative relationship with future returns, both in each decade and the overall dataset, and even once we account for variability in exogenous factors through a fixed effects model. These results lend further credibility to the conclusions reached

29 previously by Gruber (1996), Carhart (1997) and Shukla (2000) and suggest that the heightened fees charged by the management are overwhelmingly unjustified in terms of excess performance. Though we acknowledge that the issue of expenses contains far more complexity due to fee structures and strategies of funds than our study allows, we believe that a general policy of preferring low expense funds is still sound in achieving better returns. Fund size as defined by assets under management or total net assets (TNA) is another characteristic for which our results fall largely in line with previous analyses. Like Chen, Hong, Huang, and Kubik (2004), we find increased total net assets to generally have a negative effect on future returns in our multivariate, decade specific, and fixed effects regressions. However, though the results are largely statistically significant at the 1% level, the magnitude of the relationship we find is much smaller than that observed by Chen, Hong, Huang, and Kubik (2004) never exceeding 1 basis point. Thus, while we do observe statistically significant evidence pointing to a negative relationship between size and returns, we believe that the magnitude of the relationship does not merit any actionable advice in mutual fund selection. With regards to turnover ratio, we are unable to reach any conclusions from the results due to a lack of statistical significance and consistency in the signs of the coefficients. Though we do see some evidence of a negative relationship that would indicate that higher turnover cannot overcome higher transaction fees leading to lower after fee performance as described by Carhart (1997) and Wermers (2000), this evidence is limited in its statistical significance across multivariate and decade specific regressions and disappears entirely under a fixed effect model. In examining positive performance persistence, we observe mixed results in line with the conflicting conclusions reached in prior studies. In our standard multivariate regressions, we find strong evidence against positive performance persistence, suggesting that higher prior year

30 returns predict lower future returns. However, these conclusions exhibit uncertainty when subject to decade specific regressions, where the coefficients vary between positive and negative. Once we account for heterogeneity in exogenous factors through a fixed effects model, the multivariate regressions result in strong evidence in favor of positive performance persistence. Again, this relationship is muddled once we take decade specific regressions of the fixed effects model. We are able to conclude that positive performance persistence likely exists as observed by Greenblatt and Titman (1992) and Gruber (1996) after accounting for variability in market dynamics and other exogenous factors, providing support for the picking prior winners fund selection strategy. However, we note that even under the fixed effects model, our decade specific regressions still produce performance persistence coefficients of varying signs that erode some of the credibility of our conclusion. Of the characteristics examined, we believe that performance persistence is a natural candidate for further research. Our study has identified a possible explanation for the conflicting evidence in favor of and against positive performance persistence, namely the potential masking effects of heterogeneity in exogenous factors such as macroeconomic conditions and market sentiment when testing for a positive relationship between past and future returns. However, even the fixed effects model fails to completely eliminate the changes in sign for the performance persistence coefficients when we separate the data into decades. This oddity is further compounded by the large swings in R 2 values for the decade specific fixed effects regression on an identical model. Additional examination into the changes in relationships between returns and fund specific characteristics, performance persistence in particular, would likely yield results of immense interest to academics and market practitioners alike.