Do Mutual Funds Time the Market?

Evidence from Portfolio Holdings

George J. Jiang, Tong Yao, and Tong Yu October 2006

forthcoming Journal of Financial Economics

and Yao are from the Department of Finance, Eller College of Management, University of Arizona, Tucson, Arizona, 85721-0108. Emails: [email protected] and [email protected]. Yu is from College of Business Administration, University of Rhode Island. Email: [email protected]. We would like to thank Tyler Brough, Jeff Busse (AFA discussant), Yong Chen, Wayne Ferson, Kathy Kahle, Chris Lamoureux, Bill Schwert (the editor), an anonymous referee, and seminar participants at University of Arizona and the 2005 AFA meeting for helpful comments and suggestions.

Jiang

Do Mutual Funds Time the Market? Evidence from Portfolio Holdings

Abstract
Previous research nds insignicant or even negative market timing ability for mutual funds using tests based on realized fund returns. The return-based tests, however, are subject to articial timing bias due to the dynamic trading effect and option-like features of stock returns. In this paper, we propose new measures of market timing based on mutual fund holdings. Our holdings-based measures do not suffer from the articial timing bias and are statistically more powerful. Applying the holdingsbased tests, we nd that, on average, actively managed US domestic equity funds have positive timing ability. Market timing funds tend to have high industry concentration, large fund size, and a tilt toward small-cap stocks. We also provide evidence that mutual funds use more than public information to predict market returns and industry rotation is key to mutual fund market timing strategy. JEL classication: G10; G11 Keywords: Mutual funds; Market timing; Portfolio holdings

Introduction

Two strands of nance literature stand in curious contrast to each other. On the one hand, many studies have argued that aggregate stock market returns are predictable, and such predictability should have a signicant impact on investors optimal asset allocation. 1 On the other hand, there is scant evidence that investors actually take advantage of such predictability in their portfolio decisions. In particular, earlier studies dating back to Treynor and Mazuy (1966) and Henriksson and Merton (1981), as well as recent studies such as Becker, Ferson, Myers, and Schill (1999) and Jiang (2003), all nd that the average market timing performance of mutual funds is insignicant and sometimes even negative. Fund managers are sophisticated and informed investors. If they cannot exploit market return predictability, it is unlikely that anyone else could.2 Existing market-timing measures, such as those of Henriksson and Merton (1981) and Treynor and Mazuy (1966), are based on nonlinear regressions of realized fund returns against contemporaneous market returns (hereafter return-based measures). It is known, however, that a nonlinear relation between fund and market returns can also be induced by factors other than active market timing. As Jagannathan and Korajczyk (1986) point out, certain dynamic trading strategies by mutual funds may give rise to option-like features in fund returns. Thus, nonlinear relations between fund and market returns may be due to the dynamic trading effect. Jagannathan and Korajczyk (1986) further show that because returns of certain stocks have option-like features, the returns of a passive portfolio investing in these stocks may also have a convex or concave relation with market returns. This is known as the passive timing effect. Both effects are often referred to as articial timing in the literature. Furthermore, Goetzmann, Ingersoll, and Ivkovich (2000) show that the return-based measures are
For evidence of market return predictability, see, for example, Campbell (1987), Campbell and Shiller (1988a, 1988b), Cochrane (1991), Fama and French (1987, 1988, 1989), Fama and Schwert (1977), Ferson (1989), Keim and Stambaugh (1986), Lamont (1998), Lewellen (1999), and Pontiff and Schall (1998). For how such predictability should affect investors asset allocation decisions, see, for example, Balduzzi and Lynch (1999), Barberis (2000), Campbell, Chan, and Viceira (2003), Campbell and Viceira (1999), and Kandel and Stambaugh (1996). 2 In addition, Blake, Lehmann, and Timmermann (1999), Coggin, Fabozzi, and Rahman (1993), Fung, Xu, and Yau (2002), and Graham and Harvey (1996) nd no evidence of positive market-timing ability by pension funds, hedge funds, or investment newsletters.
1

biased downward when funds engage in active timing and trade between the observation dates of fund returns. For instance, when funds engage in daily market timing, the return-based timing measures using monthly fund returns tend to underestimate the timing ability. To overcome these problems, we propose alternative market-timing measures based on observed mutual fund portfolio holdings (hereafter holdings-based measures). To be specic, we estimate the beta of a fund as the weighted average of the betas of individual stocks held in the portfolio, and directly test whether there is any signicant covariance between fund betas at the beginning of a holding period and the holding period market returns. Different from the return-based measures that rely on ex-post realized returns to estimate beta shifting, the holdings-based measures use only ex-ante information on portfolio holdings. Therefore, the holdings-based measures do not suffer from any bias induced by subsequent trading activities during a holding period, or the dynamic trading effect. In addition, we show that the holdings-based measures are also robust to the passive timing effect. When performing market timing tests on passive portfolios formed on stock characteristics, i.e., size, bookto-market ratio, and momentum, we nd that the holdings-based timing measures are overall small and insignicant. However, the return-based timing measures, on average, are signicantly negative. The holdings-based measure has the further benet of improved statistical power. Fund returns are volatile with observations often available only at low frequencies (e.g. monthly) and over a relatively short time period. This limits the statistical power of the return-based tests. For the holdings-based measures, fund betas are calculated from stock betas and portfolio weights. When estimating stock betas, we can take advantage of stock return observations at a higher frequency (e.g. daily) and over extended time periods. As a result, even though individual stock betas are estimated with error, fund betas, as weighted averages of a large number of stock betas, are of much higher accuracy. Our simulations conrm that the holdings-based tests have better statistical power than the return-based tests, even when fund holdings are observed less frequently than fund returns. The use of higher frequency data to enhance statistical power is similar to Bollen and Busse (2001), who perform market timing tests using daily fund returns and nd positive timing ability for a sample

of 230 domestic equity funds. However, the timing performance documented by Bollen and Busse (2001) is at the daily horizon, whereas in this study we focus on market timing activities at relatively long time horizons that are comparable to those in the market return predictability literature. In addition, the mutual fund portfolio holdings data we examine cover a large sample of actively managed equity funds. Reliable large sample data on daily fund returns are not yet available (see discussions in Busse and Irvine (2006)). Mutual fund portfolio holdings have been used in a number of existing studies to evaluate fund performance, notably Grinblatt and Titman (1989, 1993), Grinblatt, Titman, and Wermers (1995), Daniel, Grinblatt, Titman, and Wermers (1997), Wermers (1999, 2000, 2004), and Ferson and Khang (2002). These studies show that performance measures based on portfolio holdings are more powerful in detecting mutual fund stock selection ability. In addition, Daniel, Grinblatt, Titman, and Wermers (1997) and Wermers (2000) examine whether mutual funds exploit time-varying expected returns in characteristic benchmarks based on size, book-to-market ratio, and momentum. Several studies have also used portfolio allocation between cash and equity components to measure market timing and found no signicant timing performance, see, e.g., Kryzanowski, Lalancette, and To (1996), and Becker, Ferson, Myers, and Schill (1999). As pointed out by Warther (1995) and Edelen (1999), the cash position of a fund is often affected exogenously by factors such as fund ow shocks, which tend to confound the effect of market timing. We apply both return-based tests and holdings-based tests to a sample of 2,294 actively managed US equity mutual funds over the period from 1980 to 2002. The fund returns data is from CRSP, and the fund portfolio holdings data is from Thomson Financial. Statistical inference is based on a bootstrapping procedure similar to that proposed by Kosowski, Timmermann, White, and Wermers (2005), which explicitly takes into account the cross-fund correlation and the nite-sample properties of timing measures. Consistent with the existing literature, the return-based timing measures indicate that mutual funds overall have slightly negative but statistically insignicant timing performance. In contrast, the holdings-based measures suggest that mutual funds, on average, have signicantly posi-

tive timing ability. Specically, at one-month and 12-month forecasting horizons, the average timing measures of the holdings-based tests are positive but not signicant. At three-month and six-month forecasting horizons, the average timing measures are signicantly positive. In addition, the proportion of funds with strongly positive timing measures is also higher than what one would expect from a sample of funds without timing ability. These results are robust to different methods used for the estimation of stock and fund betas, as well as variations in the testing procedure. Time variations in fund betas can be driven by both active fund trading activities and passive portfolio weight changes. To examine the active market timing of mutual funds, we further perform the holdings-based tests using active changes of fund beta. The tests provide similar results as those using the level of fund betas, suggesting that mutual funds achieve market timing through active trading. The positive timing ability of mutual funds is also of economic signicance. Using a contingent claim approach, we quantify that for a median fund, the economic value of active market timing can be as much as 0.60% per year. This result is economically signicant in comparison with the documented contribution to fund performance by stock selection. For example, Wermers (2000) reports that mutual fund stock selection activities add to fund performance, on average, by 1.01% per year before transaction costs. This suggests that market timing can be an important investment strategy for mutual funds. We identify links between several fund characteristics and market timing performance. Market timing funds tend to be those with high industry concentration in their portfolios and, to a lesser extent, those with large fund size and a tilt toward small-cap stocks. Further, fund managers adjust fund betas in response to macroeconomic variables such as aggregate dividend yield and earnings-to-price ratio, in a way consistent with how these variables predict market returns. However, the average market timing performance remains signicantly positive even after controlling for these macroeconomic variables. This is evidence that fund managers use not only publicly available macroeconomic information but also private or ner information to time the market. Finally, we document that industry allocation is an important investment strategy employed by

mutual funds to achieve market timing. We nd a stronger market timing effect from industry shifting than from intra-industry portfolio allocations. Furthermore, mutual funds shift portfolio allocations to certain industries in response to macroeconomic conditions that are predictive of future market movement. The remainder of the paper is organized as follows. Section 2 species both the return-based and holdings-based market timing tests and compares the two approaches in terms of articial timing bias and statistical power. Section 3 describes the data and methodology used in our analysis. In Section 4, we examine mutual fund market timing ability by performing both the return-based and holdingsbased tests. Section 5 conduct further analysis of mutual fund market timing activities. Section 6 concludes.

Measures of Market Timing

2.1 Return-based Measures

Let rt and rmt denote the excess fund and market returns over holding period t. The Treynor-Mazuy market-timing measure (see Treynor and Mazuy (1966)) is the estimated coefcient from the following regression:
2 rt = + 0 rmt + rmt + et

(1)

And the Henriksson-Merton measure (see Henriksson and Merton (1981)) is the estimated coefcient from: rt = + 0 rmt + max(rmt , 0) + et (2)

where max(rmt , 0) is the positive part of the market excess return. Various forms of the Treynor-Mazuy and Henriksson-Merton measures have been used in the literature to test mutual fund timing ability, see for example Treynor and Mazuy (1966), Kon (1983), Chang and Lewellen (1984), Henriksson (1984), Cumby and Glen (1990), Ferson and Schadt (1996), Kryzanowski, Lalancette, and To (1997), 5

Becker, Ferson, Myers, and Schill (1999) and Jiang (2003). Most of the studies document negative but insignicant timing performance by mutual funds. Note that both measures in (1) and (2) are based on the covariance between the time-varying fund beta t and the market return rmt :3 t = 0 + rmt + t t = 0 + Irmt >0 + t for the Treynor-Mazuy measure for the Henriksson-Merton measure (3) (4)

where Irmt >0 is an indicator with value 1 when rmt > 0 and 0 otherwise. A signicantly positive indicates positive timing ability by mutual funds.

2.2 Holdings-based Measures

The return-based measures are useful when fund betas are not directly observed. However, when the portfolio holdings of a fund are observed, we can estimate the fund beta and perform tests based directly on (3) and (4). In this study, we estimate the fund beta as the weighted average of the beta estimates for stocks (and other securities) held by the fund:
N

t =
i=1

it bit

(5)

where it is the portfolio weight for stock i at the beginning of holding period t+1, and bit is the beta for stock i estimated using data prior to period t+1. Various methods of estimating stock betas are detailed later in the paper. In the spirit of (3) and (4), we measure market timing by estimating the coefcient directly from the following regressions: t = + rm,t+1 + t+1 t = + Ir m,t+1>0 + t+1
To see this, assume that the fund return follows the market model: rt = + t rmt + (4), we obtain (1) and (2), respectively, with et = t rmt + t .
3 t.

(6) (7)
Combine this with (3) and

t is the fund beta estimated at the beginning of period t + 1. The coefcients estimated where from (6) and (7) are referred to as the holdings-based Treynor-Mazuy measure and the holdings-based Henriksson-Merton measure, respectively.

2.3 Comparison of Market Timing Measures

In this section, we compare potential bias and statistical power of the return-based and holding-based market timing measures. We show that compared to the return-based measures, the holdings-based measures do not suffer from articial timing bias and have better statistical power. 2.3.1 Articial Timing Bias

As mentioned earlier, in addition to active market timing there are at least two other factors that can induce a convex or concave relation between fund and market returns. One is the dynamic trading effect, due to certain dynamic trading strategies implemented by mutual funds. The other is the passive timing effect, referring to the fact that returns of a passive portfolio can also be nonlinearly related to market returns. Both effects, as illustrated below, can introduce articial timing bias in the return-based tests. To illustrate the effect of dynamic trading, we consider a simple case where a fund trades in each period but returns are observed every two periods. Suppose a fund manager has no active timing ability but adjusts fund beta during the second period conditional on the realized market return in the rst period. This leads to a correlation between the fund beta in the second period and the market return over the rst period, which further induces a seemingly contemporaneous nonlinear relation between the realized two-period fund and market returns. For example, a positive-feedback manager who increases market exposure after a market run-up would exhibit positive articial timing, and a contrarian manager who reduces market exposure after a market run-up would exhibit negative articial timing. That is, some commonly used trading strategies, seemingly innocuous, can generate nonlinear relations between fund and market returns. The dynamic trading effect described above is also known in the literature as interim trading, 7

because it is caused by fund trading activities between the return observation dates. Note that interim trading can also cause return-based tests to underestimate the true market timing ability. This is the case analyzed by Goetzmann, Ingersoll, and Ivkovich (2000). They perform simulations to show that when funds engage in daily market timing, the Henriksson-Merton measure based on monthly fund returns is biased downward and has low power. The holdings-based tests are robust to the dynamic trading effect. Different from the return-based tests, which measure the contemporaneous relation between realized fund returns and market returns, the holdings-based tests use only ex ante information on portfolio holdings. Namely, fund betas in (5) are computed using portfolio weights observed at the beginning of a holding period. The holdingsbased measures are thus not affected by subsequent fund trading activities during the holding period. In addition, because of the use of beginning-of-period fund betas, the holdings-based tests remain unbiased in the presence of high-frequency market timing activities described in Goetzmann, Ingersoll, and Ivkovich (2000). This, of course, is less a concern for our empirical analysis as we focus on mutual fund timing activities at relatively long horizons, from one month to one year. The passive timing effect has been documented in existing studies for the return-based tests. For instance, Jagannathan and Korajczyk (1986) show that returns of certain stocks exhibit option-like features relative to market returns. As a result, returns of a passive portfolio investing in these stocks may have a convex or concave relation with market returns even when funds are not market timers. In contrast, in the holdings-based tests stock betas are estimated using data prior to the holding period, and thus not affected by any contemporaneous nonlinear relation between stock and market returns. To quantify the passive timing effect, we perform both return-based and holdings-based tests on passive portfolios. These portfolios are formed on stock characteristics, namely size, book-to-market ratio, and momentum, and are similar to those in Daniel, Grinblatt, Titman, and Wermers (1997). The details of portfolio construction are given in Appendix A. The results reported in Appendix A show that the mean and median return-based timing measures of the passive portfolios are signicantly negative. The results are consistent with Ferson and Schadt (1996), who nd negative timing for a

simple buy and hold strategy using return-based measures. In contrast, there is no signicant articial timing bias for the holdings-based tests. The mean and median of the holdings-based measures across passive portfolios are insignicantly different from zero. To further illustrate the pattern and magnitude of the passive timing effect with respect to portfolio characteristics, we plot the return-based and holdings-based timing measures of the passive portfolios in Figures A.1 and A.2. The plots show that negative biases in the return-based measures are highly related to size and most pronounced for small-cap stocks. The pattern is consistent with the results in Jagannathan and Korajczyk (1986), who nd that equal-weighted market returns exhibit a concave relation with value-weighted market returns. Stocks with higher book-to-market ratio (value stocks) and higher momentum (past winners) also tend to have more negative passive timing effect. In contrast, the holdings-based timing measures on passive portfolios are in general small in magnitude, with no systematic biases. Nevertheless, in our empirical analysis we explicitly control for the passive effect in the holdings-based tests. All individual stock beta estimates are adjusted by subtracting the average beta of the matching passive characteristics portfolio. This procedure has an added benet of reducing cross-fund heteroscedasticity in market timing measures. 2.3.2 Statistical Power

The difference in statistical power between the return-based and holdings-based tests results from different information used in each of the tests. In practice, data on fund returns are often available at a low frequency (such as monthly or quarterly) over a limited period of fund life, and thus not sufcient for the precise estimation of market timing measures. The advantage for the holdings-based measure is that with portfolio holdings information, fund betas can be estimated directly from stock betas. When estimating stock betas, we are no longer constrained by the length of fund life and can also take advantage of the stock return data available at a higher sampling frequency (e.g. daily). It is well documented in the literature that beta, as a variance-covariance measure, can be better estimated from data sampled at higher frequency, see, for example, Merton (1980), and French, Schwert, and Stambaugh (1987). 9

To verify the advantage of the holdings-based measures, we simulate returns and portfolio holdings for mutual funds with various levels of market timing ability ( ), and perform both return-based and holdings-based Treynor-Mazuy tests on the simulated data. In the simulation, we take into account two realistic features in the data. First, the fund return data are available at monthly frequency, while the portfolio holdings data are available at quarterly frequency. Second, the portfolio holdings data are available for a shorter time period than the fund return data. Details of the simulation and results are provided in Appendix B. The results show that the standard errors of the holdings-based measures are several times smaller than those of the return-based measures. The simulations conrm that, even when fund holdings are observed less frequently than fund returns, the holdings-based tests have better statistical power than their return-based counterparts.

Data and Methodology

3.1 Data
We combine two mutual fund datasets in our analysis. The rst is the CRSP mutual fund dataset (hereafter the CRSP dataset), which has information on monthly returns and fund characteristics such as total net assets, expense ratio, loads, and turnover ratio, for all US-based mutual funds. The second is the fund portfolios holdings dataset from Thomson Financial (hereafter the Thomson dataset). Its predecessor is the CDA/Spectrum data used in a number of empirical studies, such as Grinblatt and Titman (1989, 1993), Daniel, Grinblatt, Titman, and Wermers (1997), and Wermers (2000, 2004). The Thomson dataset contains quarterly or semiannual information on portfolio holdings for equity mutual funds investing in the US market.4 Most mutual funds in the Thomson dataset are US-based. There are also a number of foreign-based, mostly Canadian, funds. As noted by Daniel, Grinblatt, Titman, and Wermers (1997), CDA/Spectrum, the predecessor of Thomson dataset, does not have survivorship
The Thomson dataset is based on mandatory and voluntary disclosure of fund holdings. Prior to 2004, mutual funds are required to disclose their holdings semiannually. Many funds voluntarily disclose their holdings quarterly. SEC increased the mandatory disclosure frequency from semiannual to quarterly, effective May 2004. Such a change could further improve the power of holdings-based mutual fund performance measures with more observations on fund holdings.
4

10

problem. However, funds are added to the Thomson dataset at a slower pace than to the CRSP dataset. Therefore, a recently started fund may be included in the CRSP dataset but not yet appear in the Thomson dataset. We manually match the funds in the two datasets by fund names and ticker symbols. The matching procedure is similar to that in Wermers (2000). Since our focus is on actively managed US domestic equity funds, we only include US funds with one of the following four investment objectives dened in the Thomson dataset: a) aggressive growth, b) growth, c) growth and income, and d) balanced. 5 When a fund has multiple share classes, returns of these share classes are separately reported in the CRSP dataset but holdings in the Thomson dataset are for the entire fund. We calculate the returns of a multi-class fund as the weighted average returns across share classes, using TNA as the weight. Fund expense ratios, turnover ratios, and loads are similarly calculated. To ensure robust statistical inference, we require a fund to have a minimum of 8 quarters of holdings data and 24 monthly return observations to be included in our sample. The nal matched dataset has 2,294 unique funds over the period from 1980 to 2002 with information on both portfolio holdings and returns.6 Table 1 reports summary statistics of our mutual fund sample during the period from 1980 to 2002. Out of a total of 2,294 funds, more than half (1,390) are growth funds. There are 255 aggressive growth funds, 411 growth and income funds, and 208 balanced funds. By averaging rst over the time series for each fund and then averaging across funds, we obtain the following characteristics: the average total net assets (TNA) of the funds in our sample is $482.60 million, with an annual return of 9.96%, an annual turnover ratio of 0.94, an annualized
The investment objective codes in the Thomson data sometimes contain errors. We manually remove from our sample the index funds, foreign-based funds, US-based international funds, xed-income funds, real estate funds, precious metal funds, and variable annuities that are misclassied as active US domestic equity funds. 6 Various existing studies have also matched funds from CRSP and Thomson/CDA. Wermers (2000) combines the CDA data with the CRSP dataset over the period between 1975 and 1994. His sample contains 1,788 equity funds. Kacperczyk, Sialm, and Zheng (2005) match the CRSP dataset with the Thomson dataset over the period between 1984 and 1999 and come up with 1,971 unique equity funds. Cohen, Coval, and Pastor (2005) also match the Thomson dataset with the CRSP dataset over the period between 1980 and 2002Q2. They report 235 matched funds at the end of 1980 and 1,526 matched funds in 2002Q2. It appears that our mutual fund sample is at least as inclusive as those in the existing literature.
5

11

load of 0.30%,7 and an annual expense ratio of 1.31%. The funds invest 85.89% of assets in common stocks. In addition, the average fund age, calculated as the time between the rst and last monthly return observations in the CRSP dataset, is 13.46 years. The mean number of stocks held in a fund is 102.15, while the median number is 61.90. For a typical fund, we have 131.96 months of return data and 27.24 quarters of portfolio holdings data during the sample period. In our empirical test, the CRSP value-weighted index return is used as a proxy for the market return. Individual stock returns and the one-month T-bill yields (our proxy of the risk-free rate) are also from CRSP. We also obtain observations of several economic variables that are documented in the literature as predictive of market returns, namely the term spread, credit spread, aggregate dividend yield on S&P 500 index, and the earnings-to-price ratio of the S&P 500 index. These variables are constructed using data from Global Insight.

3.2 Statistical Inference: The Bootstrapping Approach

There are a number of issues involved in statistical inference for mutual fund timing tests. When there is only one fund, the test for timing ability can be based on the t-statistic of the timing measure ( ). However, when there are a large number of funds, even though none of them has timing ability, by random chance there will be some funds with signicant timing measures based on the t-statistics. This is analogous to the data-snooping problem discussed in Lo and MacKinlay (1990), and Sullivan, Timmermann, and White (1999). One solution is to test whether all the timing statistics are jointly zero. However, the usual joint test (e.g., the Wald test) requires inversion of the covariance matrix of the timing statistics. When the number of funds is large and the length of the time series is short, the covariance matrix is likely to be poorly estimated or even singular. In this paper, we base the statistical inference on the cross-sectional distribution of timing measures. We focus on the following cross-sectional statistics: the mean, median, standard deviation, skewness, kurtosis, the quartiles at 25% and 75%, as well as the extreme percentiles at 5%, 10%, 90%, and 95%. The statistical signicance at these extreme percentiles can tell us whether there are funds
7

Following Sirri and Tufano (1998), the annualized load is the total load divided by 7.

12

with strong positive timing ability or perversely negative timing ability. There are two additional issues associated with the statistical inference based on the cross-sectional distribution of the market-timing measures. One is the violation of the i.i.d. assumption across funds. The other is the nite sample property of the test statistics. Consider the t-statistics of the timing measures, which are asymptotically normal under quite general conditions. However, since funds hold similar stocks, their betas are correlated. As a result, the timing measures are correlated and the tstatistics are not i.i.d. across funds. This issue is further complicated by the fact that funds may exist for only a short period of time and some funds may not overlap with each other at all. The nite sample distributions for the cross-sectional statistics, particularly those extreme percentiles, may differ from their asymptotic counterparts. To address the above issues, we resort to a bootstrapping approach similar to that proposed in Kosowski, Timmermann, White, and Wermers (2005). In the bootstrapping procedure, we randomly resample the data under the null hypothesis of no market timing while maintaining the covariance structure across fund betas or across fund returns. Statistical inference is then based on whether the empirical distribution of market timing statistics is signicantly different from what one would expect under the null hypothesis that no fund has timing ability.

Empirical Analysis of Market Timing

4.1 Return-based Timing Measures

For comparison purposes, we rst perform the return-based tests in (1) and (2) to obtain an estimate of for each fund, using monthly fund and market returns. The t-statistic for is computed using the Newey-West heteroscedasticity and autocorrelation-consistent variance estimator (see Newey and West (1987)), with a six-month lag. We then calculate the cross-sectional statistics for and the tvalues. Statistical inference is based on the bootstrapped p-values following the procedure in Appendix C, with 2,000 replications. Figure 1 plots the density of the cross-sectional distribution of t-statistics for all funds, together with that of the bootstrapped t-statistics. Both density functions are estimated using a Gaussian kernel 13

and the bandwidth choice of Silverman (1986). The comparison of these two density functions provides informal but intuitive inference about the timing ability of mutual funds. For the Treynor-Mazuy t-statistics, the sample density plot is visibly left-shifted relative to the bootstrapped one, indicating overall negative timing ability. For the Henriksson-Merton t-statistics, the density plot is also leftshifted, but to a lesser extent. Table 2 reports cross-sectional statistics of the return-based measures, the t-statistics, and their bootstrapped p-values. Consistent with the previous literature, the mean and median of the timing statistics are slightly negative, suggesting overall negative timing performance. However, both the mean and median timing measures are statistically insignicant based on the bootstrapped p-values. In addition, none of the low percentiles (5 and 10th) and high percentiles (90 and 95th) is statistically signicant. In summary, the return-based measures are on average negative. The pattern is consistent with those documented in previous studies. For example, using various return-based measures, Henriksson (1984), Ferson and Schadt (1996), and Jiang (2003), among others, nd that mutual funds on average exhibit negative market timing. As discussed in Section 2.3, the return-based results are likely subject to the articial timing bias.

4.2 Holdings-based Timing Measures

For the holdings-based tests, we focus on the Treynor-Mazuy measure in (6), with monthly, quarterly, semi-annual, and annual forecasting horizons. In other words, market returns are measured over 1, 3, 6, and 12 months after the date of portfolio holdings. The betas of individual stocks are estimated using one-year daily stock returns prior to the date of portfolio holdings. To account for the effect of nonsynchronous trading, we estimate the market model for each stock using market returns up to ve daily leads and ve daily lags, in addition to the contemporaneous term. That is,
5

ri = ai +
q =5

biq rm, q + ei

(8)
5 q =5 biq .

Following Dimson (1979), the stock beta is the sum of the estimated coefcients, i.e., bi =

We require a stock to have at least 60 daily observations during the estimation period, otherwise we 14

assume a value of one for the stock beta.8 Non-stock securities are assumed to have a beta of zero. As discussed in Section 2.3, to eliminate potential passive timing effects and reduce cross-sectional heteroscedasticity, we further adjust the stock beta estimates by subtracting the average beta of its characteristics-matched portfolio ( bc i ), i.e., bi = bi bc i (9)

The construction of passive characteristics portfolios is detailed in Appendix A. It should be noted that our nding of positive timing ability by mutual funds is not driven by this adjustment. Using unadjusted beta bi to perform the timing tests, there are no material changes to the conclusions. With the estimates of fund beta based on (5), the timing coefcient is estimated from (6) for each fund. The t-statistic of is computed following the Newey-West method with a six-month (2-quarter) lag. 9 Finally, the bootstrapped p-values for the cross-sectional statistics are obtained with 2,000 replications following the procedure in Appendix C. Figure 2 plots the density of the cross-sectional distribution of the t-statistics estimated for all funds, together with that of the bootstrapped t-statistics. The density functions are estimated the same way as in the case of return-based measures. For all four forecasting horizons (1, 3, 6, and 12 months), the sample density functions are right-shifted relative to the bootstrapped ones, indicating overall positive timing ability by mutual funds. Table 3 reports the cross-sectional statistics of the holdings-based measures and their t-statistics, as well as the bootstrapped p-values. First, at the one-month forecasting horizon, the mean and median t-statistics are positive, but not statistically signicant according to the bootstrapped p-values. Second, at the three-month and six-month forecasting horizons, however, the bootstrapped p-values
Stocks with insufcient data for beta estimation account for a very small portion of the portfolio values. We have further experimented with the following variations by (i) setting the stock beta to the matching portfolio beta, and (ii) estimating the stock beta using data available during other time periods. The results are robust to such variations. 9 When we examine long forecasting horizons, such as 12 months, a six-month lag may not be sufcient to capture serial correlations in the overlapping data. However, since the bootstrapping procedure is also based on overlapping data and the same lag order, the bootstrapped p-values should be robust to the choice of lag order. As an extra caution, we also set the lag order to a higher value up to 12-month as well as a data-dependent value of 0.5T 1/3 , and there are no material changes in the results.
8

15

for the mean, median, the 25, 75, and 90th percentiles are all much smaller, below 10%. The results suggest that over these intermediate horizons, there is on average signicantly positive timing ability by mutual funds. There are some funds exhibiting strong positive timing ability, but no evidence of perversely strong negative timing. Third, the evidence is less signicant at the 12-month forecasting horizon. Considering a nearly 100% average annual turnover ratio for mutual funds in our sample, the insignicant timing ability over 12-month horizon should come as no surprise. Finally, compared to the timing measure , the t-statistics have much lower kurtosis. This is consistent with statistical theory that as a pivotal statistic, the t-statistics offer more robust inference. 10 Therefore, while we report the bootstrapped p-values for both the cross-sectional and t-statistics, we focus on the latter for more reliable inference. The positive timing ability documented in Table 3 is in clear contrast with the negative evidence from Table 2. To reconcile the difference, we perform further analysis by constructing matching passive portfolios for mutual funds. Specically, we replace individual stocks in the fund portfolio with their corresponding characteristics-benchmark portfolios (see details in Appendix A). Note that, by construction, the hypothetical buy-and-hold returns of these passive portfolios mainly capture the passive timing effect in fund returns. We then perform return-based market timing tests based on the passive portfolio returns. Consistent with the evidence in Section 2.1 based on characteristicsbenchmark portfolios, the results show that the average passive timing effect is negative. That is, the passive timing effect in fund returns tends to offset the effect of positive timing. As such, even in the presence of active market timing, fund returns may still exhibit a concave or insignicantly convex relation with market returns. We have also performed the holdings-based Henriksson-Merton timing tests in (7). The results, not tabulated in the paper for brevity, have similar patterns as the Treynor-Mazuy tests reported in Table 3.
The timing measures ( ) are non-pivotal statistics and heteroscedastic across funds as their distributions depend on nuisance parameters such as the variance of the error terms. The t-statistics are pivotal and have asymptotic standard normal distributions. Statistical theory (e.g. Hall (1992)) suggests that bootstrapped pivotal statistics have better convergence properties and are more homogeneous.
10

16

4.3 Holdings-based Tests using Alternative Stock Beta Estimates

To ensure that the above results are robust to how stock betas are estimated, we further perform the holdings-based tests using three alternative estimates of individual stock betas, namely, (i) stock beta estimates using daily returns over the past 3 months, (ii) stock beta estimates using monthly returns over the past 5 years, and (iii) stock beta estimates conditioning on rm characteristics and macroeconomic variables and using monthly returns over the past 5 years. When using the past three-month daily returns to estimate stock betas, we perform the same regression as specied in (8). A minimum of 22 daily return observations are required, otherwise we assume a value of one for the stock beta. When using the past 5-year monthly returns to estimate stock betas, we perform the following regression: ri = ai + bi1 rm + bi2 rm 1 + ei (10)

The stock beta estimate is then bi = bi1 + bi2 . We require a stock to have at least 12 monthly return observations, otherwise we assume a value of one for the stock beta. For both of the above stock beta estimates, we further use (9) to obtain the characteristics-adjusted beta bi , which is then used to compute the fund beta following (5). To obtain stock betas conditional on rm characteristics and macroeconomic variables, we follow Avramov and Chordia (2005) and estimate the following model using monthly observations: ri = ai + bi1 rm + bi2 z 1 rm + bi3 Sizei 1 rm + bi4 Sizei 1 z 1 rm +bi5 BMi 1 rm + bi6 BMi 1 z 1 rm + ei (11)

where z 1 is the credit premium, a macroeconomic variable. Sizei 1 is the log of market capitalization, and BMi 1 is the book-to-market ratio. The credit premium is the average yield spread between Moodys Baa-rated corporate bonds and Aaa-rated corporate bonds, where the data is from Global Insight. Data on market capitalization is from CRSP. Book value of equity, used in computing the book-to-market ratio, is based on the end of the most recently reported scal year from Compustat. All three conditioning variables are lagged by one month. The model is estimated on a rolling basis, 17

using monthly data over the past 60 months. The stock beta estimate is based on the estimated coefcients, i.e., bit = bi1 + bi2 zt1 + bi3 Sizeit1 + bi4 Sizeit1 zt1 + bi5 BMit1 + bi6 BMit1 zt1 . Since there are more independent variables involved in model (11), we require a stock to have at least 36 monthly observations. Otherwise, we replace the conditional beta with the unconditional beta estimated using the past 5-year monthly returns. For the conditional beta estimates, we do not make characteristicsbased adjustments. These three alternative estimators may capture different time-varying components of the stock beta. The estimator based on 5 years of monthly returns may capture the long-run component, the estimator based on 3 months of daily returns is more likely to capture the short-term dynamic component, and the third estimator explicitly takes into account of beta variation associated with rm characteristics and macroeconomic conditions. Table 4 reports the cross-sectional distribution of the t-statistics for the holdings-based measure with alternative stock beta estimates. The results are generally consistent with those reported in Table 3 where stock betas are estimated using 12 months of daily returns. At the one-, and 12-month horizons, there is positive but insignicant evidence of timing. At the three- and six-month horizons, the mean and median of the t-statistics are signicantly positive. The only noticeable difference with Table 3 is that when stock betas are estimated using 5 years of monthly returns, the results are slightly weaker at the six-month horizon where the bootstrapped p-values for the mean and median t-statistics are slightly above the 10% level. Overall, with different methods of estimating stock betas, the holdingsbased tests provide robust evidence of positive market timing performance by mutual funds.

4.4 Holdings-based Tests using Active Changes of Fund Beta

The holdings-based tests in essence measure the covariance between the level of fund betas at the beginning of a holding period and the holding period market returns. Note that time-variations in fund betas can be driven by both active fund trading activities and passive portfolio weight changes due to nonproportional changes in stock prices. Changes in fund beta due to active trading are more relevant

18

when measuring active market timing. To examine the role of fund active trading in market timing, we further decompose fund beta into two components: t = th + t t (12)

tth is what the portfolio beta of a fund would be at month t if it passively holds all the positions where t measures the change in beta due to fund in its portfolio from h-month ago (at month t h), and t ), instead of the trading activities in the past h months. In this section, we use fund beta changes ( t ), to perform market timing tests. level of fund betas ( Specically, fund beta changes due to active trading from month t h to t are constructed as follows:
Nt
P Nt

t = t

th t

=
i=1

it bit

th it bit i=1

(13)

where it is the fund portfolio weight of asset i at month t, bit is the beta estimate of asset i at month
th t, and Nt is the number of assets held in the fund portfolio at month t. In the second term, it is the

passive portfolio weight of asset i at month t inferred from fund portfolio holdings at month t h, and
th NtP is the number of assets held by the fund at month t h. The general expression of it is given

by
th = it h nt it pit
P Nt i=1

h nt it pit

(14)

h where pit is the price of asset i at month t, and nt is the number of shares of asset i held by the it

passive portfolio at month t, calculated based on the number of shares of asset i (n ith ) held by the
h fund at month t h. Namely, nt it is obtained by adjusting nith with the share adjustment factor from h CRSP, which takes into account stock splits and stock dividends. When calculating n t it , we also take

into account fund portfolio changes that are not resulted from active fund trading. For instance, even by passively holding all positions in a fund portfolio, a new stock can be added to the portfolio due to spinoffs, and an existing stock can be dropped out of the portfolio due to delisting. In the case of spinoffs, we maintain the assumption that new shares received by a fund are passively held by the fund. 19

In the case of mergers and acquisitions, shares of the target rms are exchanged into the shares of the acquiring rms. CRSP does not provide the exact number of new shares distributed to each existing share in spinoffs or share exchange ratios for mergers and acquisitions. For spinoffs, we calculate the number of new shares received per share of the parent rm using the cash values of spinoffs and the ex-date prices of the new issues in CRSP. For mergers and acquisitions, share exchange ratios are calculated from the delisting values of target rms and the post-merger prices of the acquiring rms in CRSP. In addition, stocks delisted for performance-related reasons during the period between month t h and t are assumed to have zero values in passive fund holdings at month t. Finally, all cash dividends received by a fund are kept in cash positions. For cash and other non-equity holdings, we assume a monthly rate of return equal to the one-month T-bill yield. We perform the Treynor-Mazuy market timing tests in (6) using the active changes of fund beta t . Again, we focus on monthly, quarterly, semi-annual, and annual forecasting horizons. The betas of individual stocks ( bit ) are estimated using one-year daily returns prior to the date of portfolio holdings, without adjusting for stock characteristics. The time lag h is set to be 12 months. That is, we measure active fund beta changes over the past 12 months. The results based on fund beta changes over the past 6 months (i.e., h=6 months) are overall consistent, but slightly weaker with the estimates of timing coefcient in general smaller in magnitude. Similar to the tests using the level of fund betas, t to ensure the robustness of inference for we require a fund to have at least 8 observations of trading-based measures. Table 5 reports the cross-sectional statistics of the holdings-based measures using active changes in fund beta, as well as the t-statistics and bootstrapped p-values. The t-statistic of is computed following the Newey-West method with a six-month (2-quarter) lag. Similar to Table 3, the mean and median estimates of timing coefcient ( ) are positive at all horizons. At both one-month and 12-month forecasting horizons, the mean and median timing measures are statistically insignicant, as indicated by the bootstrapped p-values. At three-month and six-month horizons, however, the mean and median timing measures are all signicantly positive. The 75th, 90th, and 95th percentiles of

20

the t-statistics are also signicantly higher than bootstrapped ones, indicating strong timing ability by some funds. These results suggest that active changes of fund beta are signicantly correlated with future market returns, providing further evidence of positive market timing ability by mutual funds.

4.5 Economic Value of Market Timing

A few studies, such as Henriksson and Merton (1981), Merton (1981), and Glosten and Jagannathan (1994), have pointed out that market timing is equivalent to a contingent claim on the market index. We follow their intuition to quantify the economic value of market timing ability of mutual funds in our sample. Specically, given the fund return process in (1), market timing generates an additional
2 terminal payoff rmt to a fund with the maturity of the payoff equal to the forecasting horizon, where

rmt is the market return in excess of the constant risk free rate rf . Assume that the market index Pm follows a geometric Brownian motion, dPm /Pm = um dt + m dW , the value of market timing V can
2 under the risk neutral measure (Q): be derived as the expected present value of rmt

V =

1 2 2 E Q [rmt ] = (1 + rf ) (em 1) 1 + rf

(15)

To illustrate, note that at the six-month horizon, the median timing measure is 0.25 percent in Table 3 (based on the level of fund beta) and 0.24 percent in Table 5 (based on active changes in fund beta). We further assume an annual risk-free rate of 3% and monthly market return variance of 0.2% based on the CRSP value-weighted index. Based on the above timing measures, the economic values of market timing from (15) are, respectively, 0.63% and 0.60% per year. These values are economically large, indicating that market timing can be an important investment strategy for mutual funds. We note, however, that the numbers can not be directly interpreted as realized contribution to fund performance since they are derived under the assumption of continuous trading and zero transaction costs. Also, the timing coefcients used in computing the economic values are the holdings-based measures that capture only the aspect of active timing, without adjusting for the negative passive timing effect in fund returns.

21

Further Analysis

Having documented statistical and economic signicance of mutual fund market timing performance, we conduct further analysis to investigate the following issues. First, which types of funds are more likely to be successful market timers? Second, do fund managers use public information, or any additional private information, in predicting and timing market returns? And third, motivated by a recent study by Avramov and Wermers (2006), we examine whether industry shifting is an important element of mutual funds market timing strategy.

5.1 Characteristics of Market Timers

While the results in the previous section indicate positive timing ability for the average fund, they also show a wide variation of timing performance across funds. It is thus tempting to identify funds or fund managers that do posses superior market timing ability. Unfortunately, as noted earlier, out of a large sample of funds, it is difcult to pin point whether the timing ability of a specic fund is due to true ability or pure luck. However, we can investigate whether there is any commonality among market timers. For example, are small funds or large funds better market timers? Do successful market timing funds charge high fees to extract rents from investors? Does active market timing result in frequent trading and thus high turnover? Furthermore, we are interested in whether market timing funds are associated with certain investment styles, and whether they make concentrated sector bets in their portfolios. To examine the above issues, we consider seven fund characteristics. The rst three are the total net assets, expense ratio, and turnover. In each year, we obtain the percentile rank of each variable across funds. The annual percentile ranks are then averaged over time for each fund. 11 The next three are the characteristics of stocks held by a fund, namely the market capitalization or size, book-to-market ratio, and momentum as measured by the past 12-month returns. In each quarter, we rst obtain the percentile ranks of size, book-to-market ratio, and momentum for all stocks in CRSP. We then compute
11

We use the percentile ranks, instead of the underlying variable, to remove any potential time trend in the variable.

22

the weighted average of the percentile ranks for each fund, where the weights are proportional to the values of stock positions in the portfolio. Finally, these quarterly measures are averaged over time for each fund. The last fund characteristic we include is the industry concentration index (ICI). This measure is computed for each fund following Kacperczyk, Sialm, and Zheng (2005) and using FamaFrench 12-industry SIC classication. ICI is also computed quarterly and then averaged over time. We perform the following panel-data regression to gauge the relation between various fund characteristics and the timing measure: it = ai + (b0 + bCi )rm,t+1 + eit it = ai + (b0 + bCi )rm,t+1 + eit (16) (17)

it is the fund beta calculated from stock beta where Ci denotes the vector of fund characteristics, and estimates using the past one-year daily returns. The models are estimated with both xed fund and time effects. Since the xed effects may not completely capture the covariance of the error terms, we further rely on the bootstrapped p-values for statistical inference. The results are reported in Table 6. A noted pattern is that active market timing funds tend to have high industry concentration in their portfolios, as the coefcient of industry concentration index (ICI) is signicantly positive in both regressions (16) and (17). This suggests a possible link between concentrated sector bets and market timing. To a lesser extent, active market timers also tend to have large fund size and a tilt toward small-cap stocks. Coefcients of these fund characteristics are signicant in only one of the above regressions. Note that by the denition of p-values in Equation C.1 of Appendix C, the high p-value (close to 1) for the coefcient of stock size (SIZE) suggests that the coefcient estimate is signicantly below the corresponding bootstrapped values. As noted in Chen, Hong, Huang, and Kubik (2004), fund size is detrimental to stock selection performance due to liquidity issues. As a result, large funds often resort to strategies involving sector bets or industry shifts. We perform further analysis on industry allocation in Section 5.3 to shed more lights on this issue. Finally, small-cap stocks have a wide dispersion of betas, thus making it easier for funds to shift portfolio beta by changing the weights of such stocks. Kacperczyk, Sialm, and Zheng (2005) also 23

document that funds with high industry concentration tend to overweigh small-cap stocks.

5.2 Public Information and Market Timing

Academic studies have documented that macroeconomic variables such as the interest rate and aggregate dividend yield can predict stock market returns. Note that the horizons at which these variables predict market returns are very similar to the horizons at which we nd active market timing for mutual funds. It is thus natural to ask whether mutual funds have relied on these macroeconomic variables to time the market. Here, we focus on ve return-predictive economic variables well documented in the nance and economics literature, namely, the short-term interest rate, term premium, credit premium, aggregate dividend yield, and aggregate earnings-to-price ratio. We use the one-month T-bill yield as the short-term interest rate. The term premium is the yield spread between the ten-year T-bond and one-month T-bill. The credit premium is the average yield spread between Moodys Baa-rated and Aaa-rated corporate bonds. The aggregate dividend yield and earnings-to-price ratio are based on the S&P 500 index. To see if these variables are part of fund managers information set, we perform the following regressions for each fund: t = a + bMt1 + et t = a + bMt1 + et (18) (19)

where Mt1 is the vector of ve economic variables (lagged by one month) and b is a vector of coefcients. Since the above models involve more regressors than the holdings-based tests in (6), we require t (or t ) for a fund to be included in the analysis. We report at least 10 quarterly observations on the mean and median t-statistics of the estimated coefcients, as well as their bootstrapped p-values, in Panel A of Table 7. Overall, the results based on the level of fund betas and the active change of fund betas are consistent. The mean and median coefcient estimates of short-term interest rate, term premium, and credit premium are all statistically insignicant. But the mean and median coefcient estimates of the aggregate dividend yield and the aggregate earnings-to-price ratio are signicantly 24

positive in both regressions. That is, fund managers do take advantage of some of the publicly available macroeconomic information when they adjust the market exposure of fund portfolios. A further interesting question is whether fund managers possess any private information about market returns that is not captured by these macroeconomic variables. For this purpose, we perform the following holdings-based tests for each fund, in the spirit of Ferson and Schadt (1996): t = a + bMt1 + rm,t+1 + et t = a + bMt1 + rm,t+1 + et (20) (21)

We are interested in whether the timing measure remains signicant after controlling for the returnpredictive economic variables. The results are reported in Panel B of Table 7, where the bootstrapped p-values are computed following the procedure detailed in Appendix C. For both models (20) and (21), adding macroeconomic variables does not substantially change the test statistics of the timing measures. At three- and six-month forecasting horizons, the bootstrapped p-values for the mean and median t-statistics are all below 10%. In summary, fund managers adjust the market exposure of fund portfolios in response to public information, especially to variables such as aggregate dividend yield and aggregate earnings-to-price ratio. But they do not solely rely on those variables to time the market. In other words, fund managers may possess information about market returns beyond what is contained in the ve macroeconomic variables. Identifying such information remains an interesting task.

5.3 Industry Allocation and Market Timing

The above results suggest that macroeconomic information related to business cycle is used by fund managers in market timing. In a recent study, Avramov and Wermers (2006) document that optimal portfolios of mutual funds formed on macroeconomic information exhibit large variations in industry tilts over the business cycle. They nd that shift of industry weights is an important element to achieve superior portfolio returns. The results in Table 6 also suggest that funds with high industry concentration in their portfolios are more likely to be active market timers. In the following we examine a 25

related issue, i.e., whether mutual timing funds actively shift industry weights to time the market. To perform the analysis, we construct industry beta and industry-adjusted beta for funds in our sample. The industry beta of a fund is the weighted average of industry portfolio betas:
NI

tI

=
k=1

I I bkt kt

(22)

I where kt is fund portfolio weight in industry k at the beginning of period t + 1, bI kt is the beta of

industry k , calculated as the value-weighted average of individual stock betas (estimated using the past one-year daily returns), and NI is the number of industries. We classify stocks into 12 industries following the Fama-French industry SIC classication. The industry-adjusted beta of a fund is the t is the fund t I , where difference between the beta of the fund portfolio and the industry beta: t beta computed in (5), based on individual stock betas estimated using the past one-year daily returns. We perform holdings-based market timing tests using both the industry beta and the industryadjusted beta, with results reported in Table 8. The measures based on industry beta represent market timing achieved by industry rotation, whereas the measures based on industry-adjusted beta represent market timing achieved through intra-industry allocation. The results are consistent when the tests are based on the level of fund beta (Panel A) and when they are based on the active change of fund beta (Panel B). According to the bootstrapped p-values, at three-month and six-month horizons, the mean and median timing measures based on both industry betas and industry-adjusted betas are signicantly positive. However, the magnitude of the mean and median timing measures based on industry betas is substantially higher than those based on industry-adjusted betas. The above results suggest that, on average, mutual funds shift portfolio weights to high beta industries in anticipation of market upswings, and to low beta industries when expecting market downturns. Industry rotation is an important element of the mutual fund market timing strategy. To further link these ndings with those in the previous section, we investigate whether mutual funds shift industry weights in response to business-cycle related macroeconomic information. We perform the analysis by regressing quarterly changes of fund industry beta tI against the ve macroeconomic variables considered in the previous section, namely short-term interest rate, term premium, 26

credit premium, aggregate dividend yield, and aggregate earnings-to-price ratio. The regression results show that changes in fund industry beta are negatively correlated with the level of short rate, positively correlated with both aggregate dividend yield and aggregate earnings-to-price ratio. The median coefcients of these variables are, respectively, -0.06 (0.97), 0.45 (0.00), and 0.02 (0.02), where the numbers in the brackets are the bootstrapped p-values. The results suggest that funds tend to shift to high (low) beta industries when the short rate is low (high), the aggregate dividend yield is high (low), or the aggregate earnings-to-price ratio is high (low). In other words, fund managers respond to macroeconomic information when shifting industry weights of fund portfolios. Finally, we note that our nding is consistent with some anecdotal evidence regarding how fund managers achieve market timing. For example, in a letter to fund shareholders the manager of Fidelity Contrafund provided the following discussions on the outlook of the market and his investment strategy:12 I believe the economy will begin to slow. Its hard to predict when, but with higher oil prices, the Federal Reserve Board raising interest rates, high consumer debt and a possible slowing in the appreciation of real estate, economic growth should slip ... [A] sluggish economic backdrop often favors less-cyclical, steadier blue-chip growth stocks ... I have condence in our ability to nd those companies. Less-cyclical, steadier blue-chip stocks typically have lower betas, and tilting toward these stocks due to pessimistic market views is a textbook example of market timing. Interestingly, many fund managers seem to be more explicit when discussing stock selection strategies, but less so when discussing market timing strategies. The above example shows that certain strategies pursued by mutual funds, such as industry rotation, may have the actual effect of market timing, while not explicitly referred to as such.
Fund Talk: The Managers Overview, Shareholder Update and Semiannual Report, Fidelity Contrafund, June 30, 2005.
12

27

Conclusion

Using traditional return-based timing measures, existing studies have so far found insignicant and sometimes even negative market timing performance by mutual funds. These results seem to suggest that mutual fund managers do not exploit the predictability of market returns documented in the economics and nance literature. In this paper, we propose new measures of market timing performance for mutual funds using information on fund portfolio holdings. Compared to traditional timing measures based on realized fund returns, our proposed market timing measures are not subject to articial timing bias, and yet have better statistical power. We perform the holdings-based tests and nd that, on average, actively managed US domestic equity funds possess positive timing ability. There are also some funds with strong timing skills, and the proportion of these funds in our sample exceeds what one would expect if no fund has any timing ability. Furthermore, the tests based on fund beta changes show that mutual funds time the market through active trading. We provide further evidence that fund managers not only adjust the market exposure of fund portfolios in response to macroeconomic conditions, but also use private information to time the market. Active market timers tend to have high industry concentration in their portfolios and, to a lesser extent, large fund size and an investment style tilting toward small-cap stocks. We also document that industry allocation plays an important role in mutual fund market timing activities, and that mutual funds shift industry weights in response to macroeconomic information. The ndings of positive market ability by mutual funds have potentially important implications. We show that the positive market ability by mutual funds is of both statistical and economic signicance. Market timing as an investment strategy may deserve more attention when evaluating the active ability of mutual funds.

28

References
Avramov, D. and T. Chordia, 2005, Asset pricing models and nancial market anomalies, Review of Financial Studies, forthcoming. Avramov, D. and R. Wermers, 2006, Investing in mutual funds when returns are predictable, Journal of Financial Economics, forthcoming. Barberis, N. C., 2000, Investing for the long run when returns are predictable, Journal of Finance 55, 225-264. Balduzzi, P., and A. W. Lynch, 1999, Transaction costs and predictability: Some utility cost calculations, Journal of Financial Economics 52, 47-78. Becker, C., W. Ferson, D. H. Myers, and M.J. Schill, 1999, Conditional market timing with benchmark investors, Journal of Financial Economics 52, 119-148. Blake, D., B. Lehmann, and A. Timmermann, 1999, Asset allocation dynamics and pension fund performance, Journal of Business 72, 429-461. Bollen, N., and J. A. Busse, 2001, On the timing ability of mutual fund managers, Journal of Finance 56, 1075-1094. Busse, J., and P. Irvine, 2006, Bayesian alphas and mutual fund performance, Journal of Finance, forthcoming. Campbell, J. Y., 1987, Stock returns and the term structure, Journal of Financial Economics 18, 373399. Campbell, J. Y., Y. L. Chan, and L. M. Viceira, 2003, A multivariate model of strategic asset allocation, Journal of Financial Economics 67, 41-80. Campbell, J. Y., and R. J. Shiller, 1988a, The dividend-price ratio and expectations of future dividends and discount factors, Review of Financial Studies 1, 195-228. Campbell, J. Y., and R. J. Shiller, 1988b, Stock prices, earnings, and expected dividends, Journal of Finance 43, 661-676. Campbell, J. Y. and L. M. Viceira, 1999, Consumption and portfolio decisions when expected returns are time varying, Quarterly Journal of Economics 114, 433-495. Chang, E. C. and W. G. Lewellen, 1984, Market timing and mutual fund investment performance, Journal of Business 57, 57-72. Chen, J., H. Hong, M. Huang, and J. Kubik, 2004, Does Fund Size Erode Mutual Fund Performance? The Role of Liquidity and Organization, American Economic Review 94, 1276-1302. Cochrane, J., 1991, Production-based asset pricing and the link between stock returns and economic uctuations, Journal of Finance 46, 207-234. Coggin, T. D., F. J. Fabozzi, and S. Rahman, 1993, The investment performance of US equity pension fund managers, Journal of Finance 48, 1039-1056. Cohen, R., J. Coval, and L. Pastor, 2005, Judging fund managers by the company they keep, Journal of Finance 60, 1057-1096. Cumby, R. and J. Glen, 1990, Evaluating the performance of international mutual funds, Journal of Finance 45, 497-521. 29

Daniel, K., M. Grinblatt, S. Titman, and R. Wermers, 1997, Measuring mutual fund performance with characteristic-based benchmarks, Journal of Finance 52, 1035-1058. Dimson, E., 1979, Risk measurement when shares are subject to infrequent trading, Journal of Financial Economics 7, 197-226. Edelen, R.M., 1999, Investor Flows and the Assessed Performance of Open-end Mutual Funds, Journal of Financial Economics 53, 439-466. Fama, E. F., and K. R. French, 1987, Permanent and temporary components of stock prices, Journal of Political Economy 96, 246-273. Fama, E. F., and K. R. French, 1988, Dividend yields and expected stock returns, Journal of Financial Economics 22, 3-25. Fama, E. F., and K. R. French, 1989, Business conditions and expected returns on stocks and bonds, Journal of Financial Economics 25, 23-49. Fama, E. F., and W. Schwert, 1977, Asset returns and ination, Journal of Financial Economics 5, 115-146. Ferson, W. E., 1989, Changes in expected security returns, risk and the level of interest rates, Journal of Finance 44, 1191-1217. Ferson, W. E., and K. Khang, 2002, Conditional performance measurement using portfolio weights: evidence for pension funds, Journal of Financial Economics 65, 249-282. Ferson, W. E., and R. Schadt, 1996, Measuring fund strategy and performance in changing economic conditions, Journal of Finance 51, 425-462. French, K. R., G. W. Schwert, and R.F. Stambaugh, 1987, Expected stock returns and volatility, Journal of Financial Economics 19, 3-30. Fung, H., E. Xu, and J. Yau, 2002, Global hedge funds: risk, return, and market timing, Financial Analyst Journal 58, 19-30. Graham, J., and C. Harvey, 1996, Market timing ability and volatility implied in investment newsletters asset allocation recommendations, Journal of Financial Economics 42, 397-421. Grinblatt, M., and S. Titman, 1989, Mutual fund performance: An analysis of quarterly portfolio holdings, Journal of Business 62, 393-416. Grinblatt, M., and S. Titman, 1993, Performance measurement without benchmarks: An examination of mutual fund returns, Journal of Business 66, 47-68. Grinblatt, M., S. Titman, and R. Wermers, 1995, Momentum investment strategies, portfolio performance, and herding: A study of mutual fund behavior, American Economic Review 85, 1088-1105. Goetzmann, W. N., J. Ingersoll Jr., and Z. Ivkovich, 2000, Monthly measurement of daily timers, Journal of Financial and Quantitative Analysis 35, 257-290. Hall, P., 1992, The Bootstrap and Edgeworth Expansion, New York: Springer-Verlag. Henriksson, R. D., 1984, Market timing and mutual fund performance: An empirical investigation, Journal of Business 57, 73-96. Henriksson, R. D., and R. C. Merton, 1981, On market timing and investment performance II: Statistical procedures for evaluating forecasting skills, Journal of Business 54, 513-534. Jagannathan, R., and R. Korajczyk, 1986, Assessing the market timing performance of managed portfolios, Journal of Business 59, 217-236. 30

Jiang, W., 2003, A nonparametric test of market timing, Journal of Empirical Finance 10, 399-425. Kacperczyk, M., C. Sialm, and L. Zheng, 2005, On the industry concentration of actively managed equity mutual funds, Journal of Finance 60, 1983-2011. Kryzanowski, L., Lalancette, S., To, M., 1996, Performance attribution using an APT with prespecied factors, Unpublished working paper, Concordia University and University of British Columbia. Keim, D., and R. Stambaugh, 1986, Predicting returns in stock and bond markets, Journal of Financial Economics 17, 357-390. Kandel, S., and R. Stambaugh, 1996, On the predictability of stock returns: An asset-allocation perspective, Journal of Finance 51, 385-424. Kon, S. J., 1983, The market timing performance of mutual fund managers, Journal of Business 56, 323-47. Kosowski, R., A. Timmermann, H. White, and R. Wermers, 2005, Can mutual fund stars really pick stocks? New evidence from a bootstrap analysis, Journal of Finance, forthcoming. Lamont, O., 1998, Earnings and expected returns, Journal of Finance 53, 1563-1587. Lewellen, J., 1999, The time-series relations among expected return, risk, and book-to-market, Journal of Financial Economics 54, 5-43. Lo, A., and C. MacKinlay, 1990, Data snooping biases in tests of nancial asset pricing models, Review of Financial Studies 3, 431-468. Merton, R. C., 1980, On estimating the expected return on the market; an exploratory investigation, Journal of Financial Economics 8, 323361. Merton, R.C., 1981, On market timing and investment performance part I: An equilibrium theory of value for market forecasts, Journal of Business 54, 363-406. Newey, W., and K. West, 1987, A simple positive semi-denite, heteroscedasticity and autocorrelation consistent covariance matrix, Econometrica, 55, 703-708. Pontiff, J., and L. Schall, 1998, Book-to-market ratios as predictors of market returns, Journal of Financial Economics 49, 141-160. Silverman, B. W., 1986, Density Estimation For Statistics And Data Analysis, London: Chapman and Hall. Sirri, E., and P. Tufano, 1998, Costly search and mutual fund ows, Journal of Finance 53, 1589-1622. Sullivan, R., A. Timmermann, and H. White, 1999, Data-snooping, technical trading rules and the bootstrap, Journal of Finance 54, 1647-1692. Treynor, J., and K. Mazuy, 1966, Can mutual funds outguess the market?, Harvard Business Review 44, 131-136. Warther, V., 1995, Aggregate mutual fund ows and security returns, Journal of Financial Economics 39, 209-235. Wermers, R., 1999, Mutual fund herding and the impact on stock prices, Journal of Finance 54, 581622. Wermers, R., 2000, Mutual fund performance: An empirical decomposition into stock-picking talent, style, transaction costs, and expenses, Journal of Finance 55, 1655-1703. Wermers, R., 2004, Is money really smart? New evidence on the relation between mutual fund ows, manager behavior, and performance persistence, Working Paper, University of Maryland. 31

Appendix A. Market Timing Tests on Passive Portfolios

The passive characteristics portfolios in our analysis are constructed following Daniel, Grinblatt, Titman, and Wermers (1997). First, at the end of every month during our sample period, we sort all stocks into ve groups according to market capitalization using the NYSE breakpoints. Second, within each of the ve size-sorted groups, we sort stocks into ve subgroups according to book-to-market ratio (BM). The market value of a rm is measured at the end of each month, and the book value of equity is based on the most recently reported scal year. Finally, within each of the 25 size-and-BM-sorted subgroups, we further sort stocks into ve subgroups based on momentum, which is measured by a stocks total return during the previous 6 months. Within each subgroup, we form equally-weighted portfolios that are re-balanced in the end of each month. This results in a total of 125 passive characteristics portfolios. We perform both the return-based and holdings-based market timing tests on these 125 passive portfolios, using the value-weighted CRSP index return as market return. The return-based tests are based on monthly returns, and the holdings-based tests are based on portfolio betas measured at the beginning of each quarter and market returns over the quarter. Stock betas used for calculating portfolio betas are estimated using the past 1-year daily returns (see Section 4.2). Statistical inference is based on the bootstrapped p-values of the timing coefcient and its Newey-West t-statistics, following the procedure described in Appendix C with 2,000 replications. The results are reported in Table A. The high p-values (close or equal to 1) for the return-based timing coefcients, according to the denition in Equation C.1 of Appendix C, suggest that the empirical estimate of market timing is consistently smaller than the bootstrapped timing coefcients, thus evidence of signicant negative timing. We further plot the return-based and holdings-based Treynor-Mazuy timing coefcients of the 125 passive portfolios against the portfolio characteristics in Figures A.1 and A.2. The 25 panels represent size-and-BM-sorted portfolios indexed by (i, j ), where i denotes the size ranking from top (small) to bottom (large), and j the BM ranking from left (low) to right (high). Within each panel or the sizeand-BM-sorted portfolio, we plot the timing coefcients of ve momentum portfolios (sorted from low to high on momentum). The x axis in each panel denotes the momentum ranking. The results are summarized in Section 2.3.1.

Appendix B. Statistical Power of Market Timing Tests

To compare the statistical power of the return-based and holdings-based timing measures, we perform the following simulations. Daily returns of N stocks are simulated using the following process: rit rf = i (rmt rf ) +
it

(B.1)

2 ), and it is i.i.d. over where i U (0, 2), eit N (0, 2 ), market excess return rmt rf N (um , m time and across stocks. Assuming 22 trading days in a month, we aggregate stock and market returns at both monthly and quarterly frequencies. We assume that a fund manager has perfect knowledge of a stocks beta and divides stocks into a high-beta group (H) and a low-beta group (L), each with N/2 number of stocks. She calculates the

32

Table A. Articial Timing Bias of Return-based versus Holdings-based Timing Tests 5% 10% 25% Mean Median 75% -0.16 (0.98) -0.31 (0.93) 0.32 (0.20) 1.03 (0.22) 90% 0.61 (0.18) 1.16 (0.53) 0.45 (0.31) 1.62 (0.21) 95% 0.72 (0.21) 1.36 (0.62) 0.58 (0.28) 2.15 (0.13) Stdev 1.32 (0.00) 2.55 (0.00) 0.28 (0.57) 0.86 (0.60) Skew 0.14 (0.46) 0.03 (0.42) -0.01 (0.50) 0.42 (0.07) Kurto -0.77 (1.00) -0.38 (0.84) 0.15 (0.48) 0.09 (0.33)

Panel A: Return-based Tests (1-Month Horizon) Treynor -3.26 -3.07 -2.33 -1.23 -1.14 -Mazuy p (1.00) (1.00) (1.00) (1.00) (1.00) t -6.50 -5.39 -3.97 -2.24 -2.15 p (1.00) (1.00) (1.00) (1.00) (1.00) Panel B: Holdings-based Tests (3-Month Horizon) Treynor -0.29 -0.22 -0.06 0.13 0.16 -Mazuy p (0.16) (0.20) (0.17) (0.18) (0.12) t -0.68 -0.50 -0.14 0.48 0.42 p (0.05) (0.08) (0.14) (0.16) (0.18)

average betas of the two groups: H = iH i /(2N ) and L = iL i /(2N ). At the beginning of each quarter, the fund manager receives a signal about market returns in the next quarter: St = rm,t+1 rf + es,t+1 (B.2)

2 where es,t+1 N (0, s ). To take advantage of the timing signal St , the fund manager adjusts portfolio weights it in the following way: the weight for each stock in the high-beta group is set as Ht = 1/N + 2 (St um )/( H L )/N , while the weight for each stock in the low-beta group is set as Lt = 1/N 2 (St um )/( H L )/N . There is no active trading during the quarter. The fund beta + (St um ), where = N i /N is the average beta of at the beginning of the quarter is t = i=1 the N stocks. In the above setup, s measures the accuracy of the fund managers market timing information, and measures the aggressiveness of the fund managers timing strategy. The fund return is r t+1 rf = um )(rm,t+1 rf ) + (rm,t+1 rf )2 + t (rm,t+1 rf ) + iH Ht i,t+1 + iL Lt i,t+1 = ( es,t+1 (rmt rf ) + et+1 , where et+1 = iH Ht i,t+1 + iL Lt i,t+1 . For simplicity, we set the risk-free rate rf = 0, and the expected market return um = 0.5%/22. We set the number of stocks N =60, approximately the median number of stocks held by funds in our sample (from Table 1). The daily standard deviation of the idiosyncratic stock return it is set as e = 0.17/ 22, where 0.17 is approximately the average standard deviation of the residuals of monthly market model for all CRSP stocks. The daily standard deviation of the market return is set as m = 0.045/ 22, close to that of the CRSP value-weighted index returns. The total number of monthly stock return and fund return observation is set as T = 132 + 60, where 132 is the mean number of monthly return observations for a typical fund in our sample, and the additional 60 months of stock returns are for the purpose of estimating stock betas. We vary two parameters in the simulation. First, the noise-signal ratio of St , i.e. s /m , takes values of 0.5, 1, 2, and 4. Second, the timing measure is set to 0.30, 0.70, and 1.60, which are close to the median, 75th and 95th percentiles of the holdingsbased Treynor-Mazuy timing measures at the 3-month forecasting horizon (see Table 3).

33

Table B. Statistical Power of Return-based versus Holdings-based Measures = 0.30

s m

= 0.70 HB3 0.302 (0.16) 0.302 (0.16) 0.303 (0.17) 0.303 (0.22) RB 0.691 (0.73) 0.690 (0.74) 0.695 (0.75) 0.692 (0.83) HB1 0.700 (0.09) 0.700 (0.12) 0.701 (0.18) 0.699 (0.33) HB2 0.700 (0.14) 0.700 (0.16) 0.701 (0.21) 0.700 (0.35) HB3 0.701 (0.17) 0.701 (0.18) 0.702 (0.23) 0.701 (0.36) RB 1.575 (0.83) 1.577 (0.86) 1.578 (0.95) 1.582 (1.25)

= 1.60 HB1 1.600 (0.13) 1.600 (0.21) 1.597 (0.39) 1.608 (0.76) HB2 1.602 (0.22) 1.602 (0.28) 1.600 (0.43) 1.612 (0.79) HB3 1.601 (0.20) 1.601 (0.26) 1.598 (0.42) 1.610 (0.78)

RB 0.300 (0.71) 0.300 (0.71) 0.300 (0.71) 0.302 (0.73)

HB1 0.300 (0.08) 0.301 (0.08) 0.301 (0.10) 0.301 (0.16)

HB2 0.300 (0.12) 0.301 (0.12) 0.301 (0.14) 0.301 (0.19)

0.50 1.00 2.00 4.00

In each simulation, we perform both the return-based and holdings-based Treynor-Mazuy tests. To be consistent with our data structure, the return-based tests (RB) are based on 132 monthly fund returns (from month 61 to month 182), while the holdings-based tests are based on 27 quarterly fund betas (from quarter 21 to quarter 47). Stock betas are estimated using the past 1-year daily returns (HB1), the past 3-month daily returns (HB2), and the past ve-year monthly returns (HB3), respectively (see Sections 4.2 and 4.3). Table B reports the means and standard deviations of both measures with 10,000 replications. The results conrm that the holdings-based tests have more statistical power than their return-based counterparts. For example, when = 0.30 and s /m = 0.25, the standard deviation of the return-based measure is 0.71, while those of the three holdings-based measures are 0.08, 0.12, and 0.16, respectively.

Appendix C. Bootstrapping Procedures

C.1. Bootstrapping Procedure for Holdings-based Tests
This procedure is used for holdings-based market timing tests, including the analysis of fund charac1t , 2t , ..., Kt ) denote the teristics in Section 5.1. and industry allocation in Section 5.3. Let b t = ( beta estimates for K funds at the beginning of period t + 1 and r m,t+1 the market return over the period t + 1. Let (bt , rm,t+1 ) be a cross-sectional statistic (say, the 90th percentile) of the holdings-based measure ( or its t-statistic). The bootstrapping procedure is to obtain the distribution of a particu lar cross-sectional statistic j (bt , rm,t+1 ) under the null hypothesis that there is no timing ability, and then compare it to (bt , rm,t+1 ). To ensure no timing ability, in each bootstrap we keep b t unchanged but randomly sample the market returns rm,t +1 . With bt and rm,t+1 we estimate the holdings-based measure for all funds and calculate the bootstrapped cross-sectional statistic j (bt , rm,t+1 ). This pro cedure is repeated for a large number (J) of times to obtain a series of j (bt , rm,t+1 ), j=1, ... J. Since the procedure preserves the covariance structure of bt with correlated fund betas, the distribution of j (bt , rm,t+1 ) approximates the distribution of (b t , rm,t+1 ) under the null hypothesis of no timing 34

ability. The bootstrapped p-value of (bt , rm,t+1 ) is computed as: 1 p= J

J

I j >
j =1

(C.1)

where I is an indicator function with value 1 if j (bt , rm,t+1 ) > (bt , rm,t+1 ) and 0 otherwise, j > J is set as 2,000 in our empirical analysis. A low value of p (close to 0) implies that the estimated timing measure is consistently higher than its bootstrapped values and thus evidence of positive timing ability, while a high value of p (close to 1) implies that the estimated timing measure is consistently lower than its bootstrapped values and thus evidence of negative timing ability.

C.2. Bootstrapping Procedure for Return-based Tests

For the return-based timing tests, we adopt a parametric bootstrapping procedure. First, we estimate the following market return model for each fund: rt = + 0 rmt + et (C.2)

0 , as well as the residuals e We retain the OLS estimates of parameters and t for all funds. To ensure no timing ability, the bootstrapped fund returns r t are generated from the randomly sampled while keeping e based on the estimated model (C.2), using and t xed. With market returns rmt bootstrapped fund returns, we estimate the return-based timing measures for all funds and calculate the cross-sectional statistic j (rt , rmt ). This procedure preserves the covariance structure of fund returns under the null hypothesis of no timing ability. The bootstrapped p-values of the cross-sectional statistics are computed following equation (C.1).

C.3. Bootstrapping Procedure for Holdings-based Tests Controlling for Public Information
This procedure is used for holdings-based market timing tests in (20) and (21) with control of public information, of which the tests in (18) and (19) are special cases. We rst estimate the following models using OLS based on monthly series of market returns, and macroeconomic variables: Mjt = aj + bj Mjt1 + ejt ,
5

j = 1, ..., 5

(C.3) (C.4)

rmt = ar +
j =1

brj Mjt1 + ert

where Mjt denotes the j th macroeconomic variable. We retain the parameter estimates a j , bj , a r , and brj , as well as the residuals e jt and e rt . In each bootstrap, the economic variables Mjt and market return rmt are generated based on the estimated models (C.3) and (C.4) by jointly resampling from the residuals e rt and e jt1 , with j=1, ..., 5. With the bootstrapped macroeconomic variables and market returns, we perform holdings-based tests for each fund. The bootstrapped p-values are obtained via equation (C.1). 35

Table 1 Summary Statistics of Fund Characteristics This table reports the characteristics of mutual funds in the sample with a breakdown according to investment objectives. Total net assets, annual return, turnover, annual load, expense ratio, percentage investment in stocks, and average number of stocks held are rst averaged over time for each fund, and then averaged across funds. A funds annual load is calculated as the total load divided by 7. Median number of stocks held is the cross-sectional median of the time series averages. Age is the average life of funds in the sample, where fund life is dened as the time between the rst and last reported monthly returns. NOBS return denotes the average number of monthly return observations across funds, and NOBS holdings denotes the cross-sectional average number of quarters with portfolio holdings information. All Funds Number of funds Total net asset ($ millions) Annual return (%) Turnover (per year) Annual load (%/year) Expense ratio (%/year) Investment in stocks (%) Age (years) Average no. of stocks held Median no. of stocks held NOBS return NOBS holdings 2294 482.60 9.96 0.94 0.30 1.31 85.89 13.46 102.15 61.90 131.96 27.24 Aggressive Growth 255 545.95 11.76 1.13 0.39 1.43 91.77 17.21 91.57 66.38 165.65 36.93 Growth 1390 397.52 9.97 0.98 0.27 1.34 91.81 11.86 104.70 61.95 119.90 24.52 Growth & Income 441 773.93 9.71 0.73 0.35 1.19 85.06 16.15 90.94 59.65 150.39 30.75 Balanced 208 360.96 8.18 0.94 0.30 1.16 53.49 13.46 122.39 63.36 132.91 26.30

36

Table 2 Return-based Timing Tests This table reports the cross-sectional distribution of the Treynor-Mazuy timing measure and the Henriksson-Merton timing measure ( ), as well as the Newey-West t-statistics (t). The return horizon is one month. The bootstrapped p-values (p) for the timing measures and the Newey-West t-statistics are reported, respectively, in the parentheses underneath. Stdev, Skew, and Kurto denote the cross-sectional standard deviation, skewness, and excess kurtosis, respectively. 5% 10% 25% Mean Median 75% 90% 0.79 (0.46) 1.53 (0.42) 95% 1.25 (0.45) 2.12 (0.37) 0.29 (0.63) 1.92 (0.28) Stdev 1.09 (0.43) 1.59 (0.06) 0.23 (0.74) 1.24 (0.04) Skew 0.32 (0.46) 0.28 (0.32) -0.25 (0.57) -0.03 (0.54) Kurto 9.10 (0.35) 8.54 (0.04) 6.38 (0.36) 3.96 (0.50)

Panel A: Treynor-Mazuy Measure (1-Month Horizon) -2.10 -1.63 -0.79 -0.26 -0.15 0.30 p (0.79) (0.81) (0.81) (0.70) (0.69) (0.52) t -2.86 -2.14 -1.27 -0.28 -0.28 0.71 p (0.88) (0.84) (0.77) (0.67) (0.66) (0.51)

Panel B: Henriksson-Merton Measure (1-Month Horizon) -0.42 -0.32 -0.15 -0.04 -0.01 0.09 0.20 p (0.58) (0.62) (0.60) (0.59) (0.56) (0.60) (0.62) t -2.09 -1.63 -0.86 -0.08 -0.08 0.72 1.47 p (0.83) (0.78) (0.69) (0.57) (0.58) (0.44) (0.33)

37

Table 3 Holdings-based Tests: Treynor-Mazuy Timing Measure This table reports the cross-sectional distribution of the holdings-based Treynor-Mazuy timing measure ( ) and the Newey-West t-statistics (t) for the 1, 3, 6, and 12 months horizons. The stock betas are estimated using the past 1-year daily returns. The timing measures are pre-multiplied by 100. The bootstrapped p-values (p) for the timing measures and the Newey-West t-statistics are reported, respectively, in the parentheses underneath. Stdev, Skew, and Kurto denote the cross-sectional standard deviation, skewness, and excess kurtosis, respectively. 5% 10% 25% Mean 0.33 (0.17) 0.23 (0.26) 0.35 (0.05) 0.65 (0.05) 0.27 (0.08) 0.71 (0.08) Median 0.19 (0.20) 0.21 (0.29) 0.31 (0.03) 0.67 (0.04) 0.25 (0.03) 0.73 (0.07) 0.18 (0.10) 0.80 (0.11) 75% 1.01 (0.11) 0.89 (0.39) 0.73 (0.03) 1.44 (0.04) 0.61 (0.04) 1.65 (0.05) 0.44 (0.12) 1.88 (0.09) 90% 2.10 (0.10) 1.63 (0.38) 1.24 (0.08) 2.23 (0.05) 1.00 (0.08) 2.45 (0.06) 0.74 (0.19) 3.01 (0.07) 95% 3.04 (0.11) 2.19 (0.33) 1.64 (0.14) 2.69 (0.07) 1.30 (0.13) 2.90 (0.10) 0.97 (0.24) 3.84 (0.07) Stdev 1.71 (0.36) 1.23 (0.76) 0.84 (0.52) 1.33 (0.25) 0.68 (0.32) 1.48 (0.19) 0.52 (0.44) 1.83 (0.14) Skew 0.78 (0.33) 0.38 (0.14) 0.22 (0.45) -0.16 (0.72) 0.13 (0.56) -0.36 (0.89) -0.49 (0.68) 0.00 (0.53) Kurto 12.74 (0.80) 3.94 (0.32) 6.94 (0.97) 1.59 (0.74) 6.43 (0.92) 2.33 (0.51) 6.09 (0.84) 2.42 (0.53)

Panel A: 1-Month Horizon -1.92 -1.19 -0.44 p (0.39) (0.36) (0.35) t -1.63 -1.21 -0.48 p (0.13) (0.18) (0.19) Panel B: 3-Month Horizon -0.81 -0.48 -0.05 p (0.17) (0.14) (0.05) t -1.48 -0.97 -0.14 p (0.11) (0.08) (0.05) Panel C: 6-Month Horizon -0.73 -0.43 -0.04 p (0.38) (0.30) (0.09) t -1.64 -1.07 -0.15 p (0.20) (0.16) (0.09)

Panel D: 12-Month Horizon -0.63 -0.37 -0.08 0.18 p (0.42) (0.34) (0.21) (0.20) t -1.90 -1.24 -0.33 0.81 p (0.22) (0.19) (0.18) (0.12)

38

Table 4 Holdings-based Tests: Alternative Stock Beta Estimates This table reports the cross-sectional distribution of the Newey-West t-statistics (t) for the holdings-based Treynor-Mazuy timing measure with alternative estimates of stock betas. The stock betas are estimated from (i) the past 3-month daily returns (3m), (ii) the past ve-year monthly returns (5y), and (iii) the past ve-year monthly returns conditional on rm characteristics and macroeconomic variables (5yc). The bootstrapped p-values (p) for the Newey-West t-statistics are reported in the parentheses underneath. Stdev, Skew, and Kurto denote the cross-sectional standard deviation, skewness, and excess kurtosis, respectively. 5% 10% 25% -0.40 (0.09) -0.64 (0.32) -0.55 (0.14) -0.39 (0.08) -0.42 (0.14) -0.45 (0.09) -0.48 (0.11) -0.56 (0.28) -0.59 (0.24) Mean 0.32 (0.10) 0.18 (0.30) 0.28 (0.13) 0.44 (0.04) 0.53 (0.07) 0.50 (0.03) 0.43 (0.07) 0.52 (0.12) 0.53 (0.05) 0.30 (0.18) 0.31 (0.32) 0.49 (0.14) Median 0.28 (0.12) 0.15 (0.36) 0.26 (0.17) 0.41 (0.04) 0.57 (0.06) 0.52 (0.03) 0.42 (0.07) 0.55 (0.11) 0.50 (0.06) 0.27 (0.18) 0.27 (0.34) 0.40 (0.18) 75% 1.05 (0.16) 0.91 (0.37) 1.05 (0.18) 1.26 (0.04) 1.38 (0.05) 1.39 (0.01) 1.34 (0.03) 1.43 (0.09) 1.53 (0.02) 1.18 (0.18) 1.29 (0.27) 1.50 (0.10) 90% 1.79 (0.23) 1.60 (0.35) 1.80 (0.17) 2.04 (0.05) 2.16 (0.05) 2.19 (0.02) 2.11 (0.07) 2.37 (0.06) 2.56 (0.01) 2.06 (0.18) 2.33 (0.21) 2.78 (0.03) 95% 2.34 (0.23) 2.18 (0.27) 2.34 (0.33) 2.61 (0.06) 2.72 (0.05) 2.72 (0.06) 2.65 (0.08) 3.01 (0.06) 3.26 (0.03) 2.76 (0.13) 3.01 (0.21) 3.76 (0.02) Stdev 1.23 (0.86) 1.24 (0.68) 1.24 (0.71) 1.37 (0.26) 1.44 (0.06) 1.38 (0.14) 1.41 (0.22) 1.55 (0.09) 1.61 (0.03) 1.52 (0.28) 1.70 (0.25) 1.88 (0.02) Skew 0.13 (0.31) 0.59 (0.09) 0.12 (0.33) 0.12 (0.33) 0.53 (0.07) 0.09 (0.37) 0.23 (0.83) 0.15 (0.32) 0.16 (0.29) 0.13 (0.33) 0.08 (0.45) 0.30 (0.19) Kurto 1.45 (0.84) 3.63 (0.39) 1.47 (0.80) 2.46 (0.44) 4.58 (0.17) 0.71 (0.94) 1.72 (0.57) 1.64 (0.73) 0.72 (0.94) 2.47 (0.34) 3.06 (0.39) 2.15 (0.44)

Panel A: 1-Month Horizon 3m t -1.64 -1.17 p (0.11) (0.09) 5y t -1.62 -1.24 p (0.14) (0.23) 5yc t -1.62 -1.17 p (0.08) (0.06) Panel B: 3-Month Horizon 3m t -1.66 -1.15 p (0.09) (0.07) 5y t -1.56 -1.17 p (0.19) (0.21) 5yc t -1.67 -1.21 p (0.20) (0.16) Panel C: 6-Month Horizon 3m t -1.73 -1.21 p (0.14) (0.11) 5y t -1.76 -1.29 p (0.35) (0.33) 5yc t -1.91 -1.38 p (0.38) (0.32)

Panel D: 12-Month Horizon 3m t -2.06 -1.43 -0.61 p (0.28) (0.20) (0.19) 5y t -2.19 -1.52 -0.77 p (0.44) (0.38) (0.39) 5yc t -2.16 -1.59 -0.73 p (0.41) (0.38) (0.32)

39

Table 5 Holdings-based Tests using Active Changes of Fund Beta This table reports the cross-sectional distribution of the holdings-based Treynor-Mazuy timing measure ( ) and the Newey-West t-statistics (t) using active changes of fund beta. The active change of fund beta is dened in (13). The timing measures are pre-multiplied by 100. The bootstrapped p-values (p) for the timing measures and the Newey-West t-statistics are reported, respectively, in the parentheses underneath. Stdev, Skew, and Kurto denote the cross-sectional standard deviation, skewness, and excess kurtosis, respectively. 5% 10% 25% Mean 0.45 (0.12) 0.30 (0.19) 0.33 (0.08) 0.48 (0.09) 0.28 (0.08) 0.61 (0.09) Median 0.28 (0.11) 0.30 (0.20) 0.23 (0.06) 0.49 (0.07) 0.24 (0.04) 0.60 (0.08) 0.13 (0.15) 0.56 (0.16) 75% 1.22 (0.05) 1.04 (0.25) 0.72 (0.05) 1.27 (0.10) 0.62 (0.04) 1.41 (0.11) 0.37 (0.19) 1.43 (0.20) 90% 2.39 (0.10) 1.79 (0.28) 1.44 (0.05) 2.01 (0.13) 1.08 (0.07) 2.33 (0.09) 0.66 (0.26) 2.37 (0.21) 95% 3.17 (0.18) 2.37 (0.28) 1.95 (0.09) 2.58 (0.13) 1.51 (0.09) 2.92 (0.11) 0.89 (0.32) 2.91 (0.27) Stdev 2.08 (0.17) 1.30 (0.84) 1.02 (0.25) 1.34 (0.55) 0.76 (0.23) 1.48 (0.32) 0.49 (0.60) 1.61 (0.54) Skew 2.64 (0.05) 0.07 (0.42) 0.47 (0.39) 0.02 (0.48) 0.35 (0.43) 0.06 (0.46) -0.64 (0.64) -0.30 (0.79) Kurto 29.88 (0.14) 2.43 (0.76) 8.40 (0.85) 2.19 (0.77) 4.94 (0.97) 3.40 (0.45) 7.35 (0.75) 4.10 (0.36)

Panel A: 1-Month Horizon -2.42 -1.44 -0.40 p (0.53) (0.42) (0.22) t -1.73 -1.22 -0.49 p (0.11) (0.12) (0.16) Panel B: 3-Month Horizon -1.01 -0.58 -0.13 p (0.26) (0.16) (0.09) t -1.60 -1.14 -0.31 p (0.10) (0.13) (0.08) Panel C: 6-Month Horizon -0.83 -0.48 -0.07 p (0.41) (0.29) (0.10) t -1.65 -1.12 -0.23 p (0.14) (0.13) (0.09)

Panel D: 12-Month Horizon -0.60 -0.38 -0.10 0.13 p (0.33) (0.31) (0.22) (0.25) t -2.02 -1.27 -0.41 0.51 p (0.22) (0.16) (0.18) (0.20)

40

Table 6 Characteristics of Market Timers This table reports the coefcient estimates ( b) of fund characteristics in regressions (16) and (17). The regression is performed jointly for all funds with a vector of fund characteristics. Fund betas are computed from stock betas estimated using the past 1-year daily returns. TNA, EXPENSE, and TURNOVER denote the average cross-fund percentile ranks of total net asset (TNA), expense ratio, and turnover for each fund. SIZE, B/M RATIO, and MOMENTUM denote the average cross-sectional percentile ranks of stock characteristics in fund portfolios based on the market capitalization, book-to-market ratio, and momentum. ICI is the average industry concentration index for each fund computed following Kacperczyk, Sialm, and Zheng (2005). The bootstrapped p-values (p) for the coefcient estimates and t-statistics are reported in the parentheses underneath. TNA EXPENSE TURNOVER SIZE B/M RATIO MOMENTUM ICI

Panel A: Regression of the Level of Fund Beta 3-Month Horizon b 1.15 -0.92 p (0.12) (0.82) t 2.06 -1.43 p (0.08) (0.86) 6-Month Horizon b 0.76 -0.74 p (0.14) (0.81) t 1.94 -1.50 p (0.11) (0.83) 3-Month Horizon b 0.98 -0.52 p (0.20) (0.72) t 1.42 -0.51 p (0.17) (0.75) 6-Month Horizon b 0.54 -0.75 p (0.27) (0.82) t 1.09 -1.09 p (0.25) (0.83) 0.37 (0.60) 0.61 (0.60) -0.38 (0.74) -0.84 (0.78) -0.34 (0.65) -2.31 (0.68) -0.27 (0.68) -2.52 (0.69) -0.51 (0.66) -1.84 (0.69) -0.17 (0.49) -1.57 (0.49) 0.95 (0.15) 1.93 (0.14) 0.77 (0.20) 2.08 (0.18) 0.36 (0.07) 3.44 (0.04) 0.25 (0.08) 3.33 (0.06)

Panel B: Regression of the Active Change of Fund Beta 1.13 (0.10) 1.98 (0.12) 0.89 (0.29) 1.74 (0.28) -0.82 (0.91) -4.32 (0.93) -0.71 (0.90) -4.93 (0.91) -0.61 (0.68) -1.66 (0.71) -0.17 (0.48) -1.39 (0.48) 0.29 (0.39) 0.93 (0.37) 0.17 (0.56) 0.82 (0.55) 0.41 (0.06) 3.06 (0.04) 0.22 (0.10) 2.41 (0.08)

41

Table 7 Public Information and Market Timing Panel A reports the Newey-West t-statistics (t) of the mean and median coefcient estimates of the macroeconomic variables in regressions (18) and (19). The regressions are performed for each fund with a vector of ve economic variables short-term interest rate (Short Rate), term premium (Term Premium), credit premium (Credit Premium), aggregate dividend yield (Dividend Yield), and aggregate earnings-to-price ratio (Earnings/Price). Panel B reports the cross-sectional distribution of the Newey-West t-statistics (t) for the estimates of the regressions (20) and (21) at 3-month (3-M) and 6-month (6-M) horizons. Fund betas are computed from stock betas estimated using the past 1-year daily returns. The bootstrapped p-values (p) for t-statistics are reported in the parentheses underneath. Stdev, Skew, and Kurto denote the cross-sectional standard deviation, skewness, and excess kurtosis, respectively. Panel A: Fund Beta and Macroeconomic Variables Short Rate Term Premium Credit Premium -0.13 (0.55) -0.03 (0.50) -0.21 (0.65) -0.12 (0.59) Dividend Yield 1.10 (0.04) 0.98 (0.06) 1.06 (0.00) 0.91 (0.01) Earnings/Price 1.23 (0.02) 1.14 (0.03) 1.15 (0.01) 1.07 (0.01) Regression of the Level of Fund Beta Mean t -0.26 -0.23 p (0.65) (0.61) Median t -0.23 -0.32 p (0.63) (0.66) Regression of the Active Change of Fund Beta Mean t -0.19 0.06 p (0.64) (0.46) Median t -0.21 0.08 p (0.67) (0.45) Panel B: Market Timing Controlling for Public Information 5% 10% 25% Mean Median 75% 90% 95% Stdev Skew Kurto

Holdings-based Tests Using the Level of Fund Beta 3-M 6-M t p t p -1.43 (0.09) -1.64 (0.16) -0.90 (0.05) -1.06 (0.11) -0.06 (0.02) -0.19 (0.06) 0.70 (0.02) 0.66 (0.06) 0.72 (0.02) 0.71 (0.04) 1.47 (0.02) 1.53 (0.05) 2.16 (0.03) 2.31 (0.07) 2.73 (0.03) 2.90 (0.07) 1.29 (0.20) 1.54 (0.10) 0.04 (0.42) -1.31 (0.98) 1.94 (0.46) 19.26 (0.03)

Holdings-based Tests Using the Active Change of Fund Beta 3-M 6-M t p t p -1.71 (0.27) -1.72 (0.18) -1.09 (0.13) -1.09 (0.10) -0.34 (0.09) -0.21 (0.05) 0.46 (0.08) 0.76 (0.03) 0.45 (0.08) 0.69 (0.04) 1.22 (0.08) 1.62 (0.02) 1.96 (0.07) 2.60 (0.02) 2.57 (0.05) 3.40 (0.01) 1.40 (0.09) 1.60 (0.06) 0.47 (0.12) 0.52 (0.15) 5.85 (.16) 2.83 (.51)

42

Table 8 Industry Allocation and Market Timing This table reports the cross-sectional distribution of the holdings-based Treynor-Mazuy timing measure ( ) based on fund industry beta in (22) and industry-adjusted beta, respectively. The tests are also performed using beta changes due to active trading. The estimates of timing coefcient capture the effects of industry rotation and intra-industry allocation in market timing. The bootstrapped p-values (p) for the timing coefcient estimates ( ) are reported in the parentheses underneath. Stdev, Skew, and Kurto denote the cross-sectional standard deviation, skewness, and excess kurtosis, respectively. 5% 10% 25% Mean Median 75% 90% 95% Stdev Skew Kurto

Panel A: Holdings-based Tests Using the Level of Fund Beta Effect of Industry Rotation at 3-Month Horizon -0.39 -0.24 -0.04 0.22 0.19 0.45 p (0.19) (0.16) (0.09) (0.06) (0.04) (0.10) Effect of Intra-Industry Allocation at 3-Month Horizon -0.56 -0.34 -0.07 0.07 0.08 0.26 p (0.19) (0.13) (0.06) (0.08) (0.05) (0.12) Effect of Industry Rotation at 6-Month Horizon -0.25 -0.18 -0.03 0.19 0.16 0.36 p (0.24) (0.25) (0.16) (0.08) (0.06) (0.08) Effect of Intra-Industry Allocation at 6-Month Horizon -0.56 -0.32 -0.08 0.06 0.08 0.24 p (0.40) (0.31) (0.10) (0.10) (0.05) (0.11) 0.79 (0.11) 0.50 (0.14) 0.62 (0.13) 0.44 (0.21) 1.09 (0.14) 0.68 (0.19) 0.81 (0.17) 0.58 (0.26) 0.56 (0.61) 0.42 (0.89) 0.44 (0.42) 0.37 (0.62) 0.56 (0.32) -0.74 (0.68) 0.34 (0.36) -0.38 (0.84) 7.77 (0.91) 8.47 (0.75) 8.00 (0.80) 9.31 (0.51)

Panel B: Holdings-based Tests Using the Active Change of Fund Beta Effect of Industry Rotation at 3-Month Horizon -0.42 -0.24 -0.05 0.25 0.18 0.45 p (0.23) (0.17) (0.11) (0.08) (0.03) (0.09) Effect of Intra-Industry Allocation at 3-Month Horizon -0.67 -0.37 -0.11 0.10 0.07 0.28 p (0.28) (0.22) (0.10) (0.07) (0.04) (0.15) Effect of Industry Rotation at 6-Month Horizon -0.36 -0.21 -0.03 0.17 0.18 0.35 p (0.27) (0.19) (0.12) (0.04) (0.02) (0.07) Effect of Intra-Industry Allocation at 6-Month Horizon -0.54 -0.32 -0.09 0.10 0.06 0.27 p (0.29) (0.21) (0.11) (0.06) (0.03) (0.07) 0.82 (0.11) 0.64 (0.26) 0.58 (0.12) 0.55 (0.15) 1.02 (0.35) 0.98 (0.31) 0.82 (0.18) 0.88 (0.18) 0.61 (0.43) 0.63 (0.55) 0.44 (0.49) 0.50 (0.40) 0.37 (0.58) 0.96 (0.27) 0.27 (0.80) 1.08 (0.25) 2.52 (0.35) 28.58 (0.20) 1.89 (0.48) 12.89 (0.61)

43

0.5 0.45 0.4 0.35 0.3 0.25 0.2 0.15 0.1 0.05 0 6 4

Panel A: The TreynorMazuy Measure

0.5 0.45 0.4 0.35 0.3 0.25 0.2 0.15 0.1 0.05 0 6 4

Panel B: The HenrikssonMerton Measure

Fig 1 The Density of t-statistics for the Return-based Timing Measures This gure plots the density of the cross-sectional distribution of the t-statistics (the solid line) for the return-based timing measures, together with that of the bootstrapped t-statistics (the dotted line).

44

0.5 0.45 0.4 0.35 0.3 0.25 0.2 0.15 0.1 0.05 0 5

Panel A: 1month Horizon

0.5 0.45 0.4 0.35 0.3 0.25 0.2 0.15 0.1 0.05

Panel B: 3month Horizon

0.5 0.45 0.4 0.35 0.3 0.25 0.2 0.15 0.1 0.05 0 5

Panel C: 6month Horizon

0.5 0.45 0.4 0.35 0.3 0.25 0.2 0.15 0.1 0.05

Panel D: 12month Horizon

Fig 2 The Density of t-statistics for the Holdings-based Timing Measures This gure plots the density of the cross-sectional distribution of the t-statistics (the solid line) for the holdings-based timing measures (with 1, 3, 6, and 12 months horizons), together with that of the bootstrapped t-statistics (the dotted line).

45

4 2 0 2 4 4 2 0 2 4 4 2 0 2 4 4 2 0 2 4 4 2 0 2 4 1 1 1 1 1

(1,1 )

4 2 0 2

(1,2 )

4 2 0 2

(1,3 )

4 2 0 2

(1,4 )

4 2 0 2

(1,5 )

2 3 4 (2,1 )

4 4 2 0 2

2 3 4 (2,2 )

4 4 2 0 2

2 3 4 (2,3 )

4 4 2 0 2

2 3 4 (2,4 )

4 4 2 0 2

2 3 4 (2,5 )

2 3 4 (3,1 )

4 4 2 0 2

2 3 4 (3,2 )

4 4 2 0 2

2 3 4 (3,3 )

4 4 2 0 2

2 3 4 (3,4 )

4 4 2 0 2

2 3 4 (3,5 )

2 3 4 (4,1 )

4 4 2 0 2

2 3 4 (4,2 )

4 4 2 0 2

2 3 4 (4,3 )

4 4 2 0 2

2 3 4 (4,4 )

4 4 2 0 2

2 3 4 (4,5 )

2 3 4 (5,1 )

4 4 2 0 2

2 3 4 (5,2 )

4 4 2 0 2

2 3 4 (5,3 )

4 4 2 0 2

2 3 4 (5,4 )

4 4 2 0 2

2 3 4 (5,5 )

Fig A.1 Return-based timing coefcients of passive portfolios This gure plots the return-based timing coefcients of the 125 passive portfolios. The panels represent 25 size-and-BM-sorted portfolios, where the 5 rows are sorted from small to large according to size and the 5 columns are further sorted from low to high according to the B/M ratio. Within each panel, the x axis represents ve portfolios sorted from low to high according to momentum.

46

4 2 0 2 4 4 2 0 2 4 4 2 0 2 4 4 2 0 2 4 4 2 0 2 4 1 1 1 1 1

(1,1 )

4 2 0 2

(1,2 )

4 2 0 2

(1,3 )

4 2 0 2

(1,4 )

4 2 0 2

(1,5 )

2 3 4 (2,1 )

4 4 2 0 2

2 3 4 (2,2 )

4 4 2 0 2

2 3 4 (2,3 )

4 4 2 0 2

2 3 4 (2,4 )

4 4 2 0 2

2 3 4 (2,5 )

2 3 4 (3,1 )

4 4 2 0 2

2 3 4 (3,2 )

4 4 2 0 2

2 3 4 (3,3 )

4 4 2 0 2

2 3 4 (3,4 )

4 4 2 0 2

2 3 4 (3,5 )

2 3 4 (4,1 )

4 4 2 0 2

2 3 4 (4,2 )

4 4 2 0 2

2 3 4 (4,3 )

4 4 2 0 2

2 3 4 (4,4 )

4 4 2 0 2

2 3 4 (4,5 )

2 3 4 (5,1 )

4 4 2 0 2

2 3 4 (5,2 )

4 4 2 0 2

2 3 4 (5,3 )

4 4 2 0 2

2 3 4 (5,4 )

4 4 2 0 2

2 3 4 (5,5 )

Fig A.2 Holdings-based timing coefcients of passive portfolios This gure plots the holdings-based timing coefcients of the 125 passive portfolios. The panels represent 25 size-and-BM-sorted portfolios, where the 5 rows are sorted from small to large according to size and the 5 columns are further sorted from low to high according to the B/M ratio. Within each panel, the x axis represents ve portfolios sorted from low to high according to momentum.

47