MACROECON & INT'L FINANCE WORKSHOP presented by Murray Carlson FRIDAY, April 19, 2013 3:30 pm – 5:00 pm, Room: HOH-706 Heterogeneous Information Diffusion and Horizon Effects in Average Returns Oliver Boguth, Murray Carlson, Adlai Fisher, and Mikhail Simutin∗ April 15, 2013 ABSTRACT We show that when stocks react to fundamentals with heterogeneous delay, the observed short-horizon mean return of a buy-and-hold portfolio is downward biased relative to its average fundamental return. Our theory predicts distinct patterns in portfolio mean returns calculated at different horizons, depending on the degree of information delay for stocks within the portfolio. Consistent with our model, average daily returns of portfolios of small, illiquid, and volatile stocks, rescaled to longer horizons by compounding, show downward bias on the order of 6% annually. Evidence of substantial bias remains in the average monthly returns of some U.S. style portfolios and in the majority of international indices. The direction and magnitude of these findings cannot be explained by standard microstructure frictions such as bid-ask bounce, iid measurement error, or asynchronous trade. Our results contribute to growing evidence of delayed price adjustment as an important friction with broad impacts. The theory and findings also have practical implications for benchmarking and performance evaluation. ∗ We thank Hank Bessembinder, Carole Comerton-Forde, Cam Harvey, Andrew Karolyi, and seminar participants at the University of New South Wales, University of North Carolina at Chapel Hill, the University of Toronto, SUNY-Buffalo, Wilfred Laurier University, the Finance Down Under Conference, the Northern Finance Association Meetings, and the Pacific Northwest Finance Conference for helpful comments. This paper previously circulated under the title “On Horizon Effects and Microstructure Bias in Average Returns and Alphas.” Boguth: oliver.boguth@asu.edu; W. P. Carey School of Business, Arizona State University, PO Box 873906, Tempe, AZ 85287-3906. Carlson and Fisher: murray.carlson@sauder.ubc.ca; adlai.fisher@sauder.ubc.ca; Sauder School of Business, University of British Columbia, 2053 Main Mall, Vancouver, BC, V6T 1Z2. Simutin: mikhail.simutin@rotman.utoronto.ca; Rotman School of Management, University of Toronto, 105 St. George Street, Toronto ON, Canada, M5S 3E6. Support for this project from the Social Sciences and Humanities Research Council of Canada and the UBC Bureau of Asset Management is gratefully acknowledged. Heterogeneous Information Diffusion and Horizon Effects in Average Returns Abstract We show that when stocks react to fundamentals with heterogeneous delay, the observed short-horizon mean return of a buy-and-hold portfolio is downward biased relative to its average fundamental return. Our theory predicts distinct patterns in portfolio mean returns calculated at different horizons, depending on the degree of information delay for stocks within the portfolio. Consistent with our model, average daily returns of portfolios of small, illiquid, and volatile stocks, rescaled to longer horizons by compounding, show downward bias on the order of 6% annually. Evidence of substantial bias remains in the average monthly returns of some U.S. style portfolios and in the majority of international indices. The direction and magnitude of these findings cannot be explained by standard microstructure frictions such as bid-ask bounce, iid measurement error, or asynchronous trade. Our results contribute to growing evidence of delayed price adjustment as an important friction with broad impacts. The theory and findings also have practical implications for benchmarking and performance evaluation. 1. Introduction Over the last two decades, considerable evidence has accumulated that stock prices react to new information with differential delays. In influential early work, Lo and MacKinlay (1990) present evidence of predictability from large stocks to small stocks. An impressive subsequent literature shows that analyst coverage, degree of firm complexity, investor attention depending on day of the week, and other factors can influence the speed with which stocks react to news.1 When stock prices relate to news at different rates, well-known consequences include strong positive autocorrelations and cross-autocorrelations in short-horizon portfolio returns. We demonstrate new implications of heterogeneity in information diffusion for average portfolio returns. We first develop a theoretical environment in which stocks react to fundamentals with different delays. We show that the observed short-horizon mean return of a buy-and-hold portfolio is downward biased relative to its average fundamental return. This result may seem surprising. Prior literature has focused on earning positive profits by active trading strategies that take advantage of slow information diffusion. Our result applies to the average short-horizon returns of a buy-and-hold strategy, and can be understood by considering the portfolio weights implied by a buyand-hold strategy. Consider measuring portfolio returns following a positive shock to fundamentals. Stocks that react slowly to the positive information will be underweighted at the beginning of the measurement interval relative to their fundamental values. At the same time, their observed short-horizon returns are expected to be high. Similarly, follow1 See, for example, Brennan, Jegadeesh, and Swaminathan (1993), Badrinath, Kale, and Noe (1995), Klibanoff, Lamont, and Wizman (1998), Chordia and Swaminathan (2000), Hong, Lim, and Stein (2000), Huberman and Regev (2001), Hirshleifer and Teoh (2003), Hou and Moskowitz (2005), Cohen and Frazzini (2008), Dellavigna and Pollet (2009), Hirshleifer, Lim, and Teoh (2009), Menzly and Ozbas (2010), Chordia, Sarkar, and Subrahmanyam (2011), Hirshleifer, Lim, and Teoh (2011), Tetlock (2011), and Cohen and Lou (2012). 1 ing a negative shock to fundamentals, slow-adjusting stocks are overweighted in the portfolio, while their future returns will tend to be low. The negative cross-sectional correlation between portfolio weights and future observed returns has implications for portfolio returns. In particular, observed average portfolio returns become downward biased relative to their fundamental return at short horizons. These effects are less important when returns are measured over longer intervals, since the effects of slow information diffusion become smaller. Thus, theory predicts specific patterns in average returns measured over different horizons. The setting for our analysis differs fundamentally from prior literature, which has considered bias in average returns arising from iid measurement error in prices. Blume and Stambaugh (1983) and Roll (1983) explain how iid measurement errors in prices cause upward bias in the mean returns of individual stocks and equally-weighted, periodically-rebalanced portfolios, due to Jensen’s inequality. Conrad and Kaul (1993), Canina, Michaely, Thaler, and Womack (1998), and Liu and Strong (2008) provide additional empirical analysis and recommendations to empirical researchers in the presence of iid measurement error. Most recently, Asparouhova, Bessembinder, and Kalcheva (“ABK”, 2010, 2012) show how biases arise in Fama-MacBeth regressions and average returns of equal-weighted style portfolios under iid measurement error, and propose empirical corrections. In contrast to this literature, we focus on delayed price adjustment as a pricing friction. Unlike the standard assumption of iid measurement error, our setting implies price deviations that are correlated with fundamentals. ABK (2012, p. 46) anticipate that the case we investigate, where measurement errors are correlated with fundamentals, should be a profitable direction for research. Our implications for biases in measured mean returns also differ fundamentally from prior literature. Under the iid measurement errors studied previously, equal-weighted portfolio returns are upward biased, and the mean returns of well-diversified value- 2 weighted portfolios remain unbiased. Indeed, the weighting schemes proposed by ABK capture the beneficial aspects of allowing portfolio weights to vary with past returns. In contrast, under the case of heterogeneous information diffusion that we focus on, portfolio returns are downward biased rather than upward biased, and value-weighted portfolios are more severely impacted than equal-weighted or otherwise rebalanced portfolios. We can thus sharply distinguish the empirical predictions for mean returns of iid measurement error versus heterogeneity in information diffusion. An existing literature analyzes differential price adjustment, but focuses on impacts to higher moments of returns, such as individual stock and portfolio variances, autocorrelations, and cross-autocorrelations. Most notably, delayed reaction of some securities to fundamentals causes portfolio returns to be positively autocorrelated and observed portfolio variances to be downward biased (Scholes and Williams, 1977; Lo and MacKinlay, 1990). These same studies conclude, in contrast to our findings, that the implications for observed mean returns are innocuous.2 However, these studies base their conclusions on the properties of logarithmic returns. Empirical researchers are more often concerned with simple returns, as for example when calculating an alpha from a time-series regression or carrying out a standard cross-sectional asset pricing test. Unlike logarithmic returns, which are additive, simple returns involve compounding. In the presence of asynchronous price adjustment, positive autocorrelations, generated by heterogeneity in information diffusion rather than fundamentals, inflates the importance of compounding. The average single-period return must be downward biased to compensate. To provide a more complete framework for empirical analysis, we develop an approximation for the difference between average returns calculated over arbitrary different 2 Scholes and Williams write, “. . .expectations of measured returns [. . .] always equal true mean returns” (p. 113). Lo and MacKinlay state, “. . . nontrading does not affect the mean of observed returns” (p. 187). 3 time scales.3 Our formula shows that when the variance ratio of long-horizon to shorthorizon returns exceeds one, indicating persistence, the average subperiod return scaled by multiplying or compounding is downward biased relative to the buy-and-hold return. We verify that this approximation is empirically accurate for a variety of standard style portfolios. Building on this result, we show that alpha differences across different horizons can be largely explained by applying our formula. Thus, alphas across horizons are not arbitrarily different, but to a close approximation are explained by a few specific moments of returns, most notably variance ratios across horizons. Consistent with our theoretical predictions, we show empirically that short-horizon portfolio return averages, alphas, and Fama-MacBeth coefficients are often not reliable, even in the case of value-weighting or other non-rebalanced strategies. For all portfolios we consider, performance metrics based on daily return measures significantly understate longer-horizon buy-and-hold returns. As theory predicts the biases are particularly strong for illiquid and volatile portfolios such as those containing small stocks and momentum losers. The differences in return measures reach up to 6% annually for U.S. style portfolios. Even monthly returns of U.S. style portfolios show meaningful biases, and only at a quarterly horizon do average returns reliably show small horizon effects. The evidence in international indices is even stronger. Average monthly returns, rescaled by compounding to an annual horizon, consistently understate longer-horizon returns. The magnitude of the effect in emerging and frontier markets, where heterogeneity in information arrivals is most severe, exceeds 10% annually in some cases. Exchange Traded Funds (ETFs) which track a subset of the international indices are available, for which daily closing prices are arguably less succeptible to the microstruc3 Horizon effects in returns have been a fundamental topic in finance since Blume (1974). More recent developments focus on the subtle effects of estimation error (Jacquier, Kane, and Marcus, 2003, 2005) and parameter uncertainty (Pastor and Veronesi, 2003; Pastor and Stambaugh, 2012). 4 ture biases that affect the individual stocks prices used to compute index returns. Although horizon effects are somewhat smaller for ETFs than for the associated indices, rescaled average monthly returns of ETFs remain consistently below the annual average returns. The differences between short and long-horizon mean returns in international markets thus remain economically significant even in data that is likely to reflect the price impact of trading activity at the individual stock level. Section 2 provides theoretical analysis of the link between heterogeneity in information diffusion and average portfolio returns. Section 3 demonstrates implications for horizon effects in standard empirical methods. Section 4 provide empirical evidence in U.S. style portfolios. Section 5 gives international evidence. All proofs are in the Appendix. 2. A Model of Heterogeneity in Information Diffusion In this section we develop a model in which stocks react to fundamental information with heterogeneous delays. We provide considerable generality in the model. The delay type of an individual firm follows a stationary Markov-switching process, permitting that the speed at which an individual firm adjusts to systematic information may change over time. The model also allows heterogeneous adjustment to multiple sources of systematic news. We first develop an analytical result showing that in the presence of heterogeneity in information diffusion, individual stock and portfolio mean returns are downward biased relative to the mean returns of fundamentals. We then provide a parsimonious calibration showing parameters that permit the model to quantitatively match in a portfolio of small stocks the following empirically observed moments: 1) a slowly decaying (hyperbolic) pattern of loadings on current and lagged daily market returns, 2) a slowly decaying pattern of return autocorrelations, and 3) observed mean returns 5 at different horizons, including apparent downward bias at short-horizons. We also show the parameters under which the model matches empirically observed moments of a value-weighted large-stock portfolio. These parameters imply negligible amounts of slow price adjustment. The setting we analyze complements a set of models developed in earlier literature. First, iid measurement error (e.g., Blume and Stambaugh (1983)) produces effects in value-weighted portfolio returns that differ qualitatively from what we observe in the data. Second, models based on asynchronous trade (e.g., Lo and MacKinlay (1990)) produce qualitatively similar patterns in portfolio autocorrelations to the data. However, prior literature has already shown that even aggressive calibrations of asynchronous trade models cannot match the level and persistence of observed portfolio autocorrelations (Boudoukh, Richardson, and Whitelaw (1994)). We add to prior literature by showing theoretically the implications of asynchronous trade for downward bias and horizon effects in average returns. We also show, complementing Boudoukh, Richardson, and Whitelaw (1994), that even aggressive calibrations of asynchronous trade models cannot capture quantitatively observed horizon effects in mean returns. For the convenience of the reader, the iid measurement error and asynchronous trade models, as well as our new theoretical results for the asynchronous trade model and calibrations, are summarized in the Internet Appendix. We now focus attention on our model of heterogeneity in information diffusion. This is the only model among the broad set we have considered that is able to jointly match empirical moments of value-weighted portfolio betas and lagged betas, autocorrelations, and horizon effects in mean returns. 6 2.1. Model Setup We assume fundamental asset returns over one period have a K-dimensional factor structure, and denote the vector of factor realizations at time t by ft = (f1,t ...fK,t )0 , for t ∈ N. For simplicity, we assume the factors are independent and identically normally distributed with mean µf and diagonal covariance matrix Σf .4 The fundamental value of an individual stock has instantaneous logarithmic returns given by rt∗ = rf t + β 0 ft + εt , (1) where we use an asterisk to denote the fundamental return, rf t is the riskless rate, β is a K × 1 vector of constant factor exposures, and εt is an iid normal variable with zero mean and standard deviation σ. To simplify notation, we suppress subscripts to distinguish individual assets throughout our exposition. To capture the idea of slow information diffusion, we permit that observed stock returns may reflect innovations in lagged as well as contemporaneous fundamentals. We stochastically assign stocks into one of Θ groups. Each group θ ∈ 1, ..., Θ differs by the speed δθ at which fundamental information incorporates into observed asset prices. Observed logarithmic stock returns follow " Θ # X rt = rf t + β δθ (Dθt−1 + ρθt ft ) + εt , (2) θ=1 where, for each θ, δθ is a scalar, Dθt−1 is a K × 1 vector, and ρθt is a diagonal K × K matrix. The state variables Dθt track “information deficits” with respect to all factors in each decay group θ, and the elements of ρθt randomly match information arriving from the K factors to a group θ. The random variables ρθt are required to be stationary, non-negative, and sum to P one: Θ θ=1 ρθt = 1. These restrictions ensure that new information arrival cannot make 4 The analysis can easily be modified to accommodate features such as non-normal returns or stochastic volatility. 7 prices “catch up” on old information faster, and that all information is eventually incorporated into prices. In the implementation, we will further assume that combinations of period−factor information are be assigned to only one decay group, i.e., that ρθt ∈ {0, 1} for all θ, t. Information deficits Dθt accumulate underreaction to past and current factor realizations according to the state equations D1t = (1 − δ1 )(D1t−1 + ρ1t ft ) D2t = (1 − δ2 )(D2t−1 + ρ2t ft ) .. . DΘt = (1 − δΘ )(DΘt−1 + ρΘt ft ), (3) where 1 = δ1 > δ2 > ... > δΘ > 0 are non-stochastic, stock specific information delays. Both the states Dθt and the random matrices ρθt are assumed to be unobservable. The impact of a current factor realization on stock returns is determined by the terms PΘ θ=1 δθ ρθt ft in equation (2). The date−t information delay is therefore determined by the evolution of the family of random matrices ρθt . To model these processes, we define the associated Markov chains skt ∈ {1, 2, ..., Θ} on the “delay groups” skt . The delay group skt = 1 is associated with the parameter δ1 = 1 and therefore produces a price process that reacts instantaneously to factor k information, while the delay state skt = Θ, with δΘ < δθ ∀θ, corresponds to the slowest possible reaction stock prices to current factor−k realizations. The transition matrices for the delay states for each factor k are assumed to be independent and are given by p p · · · pkΘ k1 k2 pk1 pk2 · · · pkΘ Pk = e−λ∆ IΘ + (1 − e−λ∆ ) .. . pk1 pk2 · · · pkΘ 8 (4) where λ is the arrival intensity of a change in state and IΘ is the identity matrix of dimension Θ. This form of the transition matrix for skt conveniently distinguishes between changes in the state variable, which occur according to the intensity parameter λ, and the steady-state distribution of skt as defined by the parameters pkθ . By further defining the diagonal elements of the matrices ρθt (k, k) = 1{skt =θ} , we assume that in delay state skt = θ a constant fraction δθ of the current factor realization immediately impacts on returns, and that the remaining information is released into the market at a geometrically declining rate δθ (1 − δθ )s−t . 2.2. The Impact of Slow Information Diffusion on Return Measurement We now consider a special case of our model of slow information diffusion where returns are driven by a single factor. We consider two classes of assets, one which we refer to as the leader assets for which factor realizations are immediately incorporated into prices and one which we refer to as the lagger assets for which there is a single delay state. Logarithmic returns of the leader assets are assumed to be generated by equation (1) with rf t = 0 and β = 1, rt∗ = p∗t − p∗t−1 = ft , (5) where p∗t is the log price of the fundemental asset. The factor realizations ft are independently normally distributed with mean µf and standard deviation σf . Lagger logarithmic returns rt = pt − pt−1 given by the system of equations rt = δ(Dt−1 + ft ) Dt = (1 − δ)(Dt−1 + ft ). (6) (7) Straightforward manipulation of the dynamic equation for the information deficit Dt shows that it is a stationary random variable. 9 Lemma 1 The information deficit follows an AR(1) process Dt − Dt−1 1−δ = −δ Dt − µf + (1 − δ)σf ξt+1 , δ (8) where ξt are independent standard normal random variables. The unconditional distribution for Dt is normal with E(Dt ) = 1−δ µf δ and Var(Dt ) = (1−δ)2 σ2 . 1−(1−δ)2 f If we further assume that there is a known date t = 0 when p∗0 = p0 = D0 = 0 we can conveniently relate the leader prices to the lagger prices. Lemma 2 Leader and lagger logarithmic prices are cointegrated: p∗t = pt + Dt . (9) The fact that leader and lagger prices are cointegrated leads to the intuitive conclusion that short-run return distributions of the two series will be different but that their long-run return distributions will converge. It is possible to obtain closed-form expressions for characteristics of lagger returns within this simplified setting. The following proposition characterizes the unconditional distribution of short-run returns. Proposition 1 Lagger short-run log returns rt are unconditionally normal with E(rt ) = µf and Var(rt ) = 1 δ δ σ2 2−δ f 1 eµf + 2 [ 2−δ ]σf < eµf + 2 σf 2 2 < σ 2 . Mean short-run simple returns are given by E (ert ) = ∗ = E ert . This proposition makes several notable points that are relevant to applied empirical research. First, the proposition shows that the unconditional mean of logarithmic returns of leader and lagger stocks are equal. Second, although slow diffusion of information has no implications for mean returns, the short-run lagger log returns have lower unconditional variance than the leader log returns. This is a direct reflection of the fact that short run lagger returns are exposed to only a fraction 0 < δ < 1 of the current 10 factor realization. Third, a consequence of these two facts is that mean simple returns of lagger stocks are a downward biased estimate of the fundamental mean simple return that can be measured from leader stocks. Although the first two points have been made in prior work (e.g., Lo and MacKinlay (1990)), the downward bias in expected simple returns has not been emphasized. Conventional wisdom is that value-weighted portfolios produce returns that are less susceptible to bias than reweighted portfolios such as equal-weighted portfolios. We now show that when some stocks respond slowly to factor information, value-weighted portfolios produce returns with an additional source of downward bias for the fundamental mean return. Assume that a fraction π of stocks are leaders and that the remaining fraction 1 − π are laggers. The following proposition characterizes the simple returns for a value weighted portfolio in this setting. Proposition 2 The simple return on a value-weighted portfolio of leader and lagger stocks is given by Rt = wt−1 eft + (1 − wt−1 )eδ(Dt−1 +ft ) , (10) where wt−1 = πeDt−1 . πeDt−1 + 1 − π (11) The mean portfolio return is given by E(Rt ) = E(wt−1 )e where Cov µf + 12 σf2 δ µf + 21 [ 2−δ σf ] +(1−E(wt−1 ))e 1−π , eδ(Dt−1 +ft ) πeDt−1 +1−π 2 +Cov 1−π δ(Dt−1 +ft ) ,e , πeDt −1 + 1 − π (12) < 0. This proposition makes several important points. Equation (11) shows that the value weights are a function of the lagged information deficit Dt−1 and, therefore, stationary. This result follows from the fact that leader and lagger stock prices are cointegrated. Equation (12) shows that there are three potential sources of bias when attempting to 11 measure fundamental mean returns using a value weighted portfolio. First, the mean weight does not equal the proportion of leaders E(wt−1 ) 6= π. Although there is no closed-form expression for this bias, it can be shown to be small. Second, the lagger mean stock return is biased downwards relative to the fundamental mean return, as was demonstrated in Proposition 1. The third source of bias arises because of the negative correlation between lagger value weights and lagger stock returns as given by Cov(1 − wt−1 , eδ(Dt−1 +ft ) ). This additional downward bias in the portfolio simple return, which is caused by downweighting past high factor returns in the lagger stocks, can be highly significant. Interestingly, the use of exogenous portfolio weights, such as equal weights, will eliminate this source of bias and will produce higher mean portfolio returns than does value weighting when slow information diffusion governs a subset of the returns. The next subsection quantifies the magnitude of these biases within the context of a calibrated version of the full model. 2.3. Parameterization and Calibration We provide a calibrated version of our model to quantify the magnitude of heterogenous information diffusion in small stocks. Heterogeneity across stocks can arise from differences in decay rates δθ as well as from the probabilities that any particular information is assigned to a given decay rate. In order to parameterize our model in a parsimonious way, we focus on only two factors, a market and non-market factor, and restrict the set of allowable delay states and transition probabilities. Small stock returns display long memory with significant autocorrelations that persist beyond one trading month (21 days). To capture the slowly-decaying autocorrelations that we find in the data, we choose the information delay parameters δθ , θ ∈ {1, 2, ...7}, to produce a geometric progression {0, 20 , 21 , ..., 25 } of shock half lives. To limit parameters in the transition matrix for delay states, shocks to factor k ∈ {1, 2} are assumed to be entirely incorpo- 12 rated in contemporaneous returns with an unconditional probability pk1 = ak and to affect returns with a lag chosen from the six possible values δθ , θ ∈ {2, 3, ..., 7}, with equal unconditional probabilities (1 − ak )/6. These restrictions on the parameter space allow us to characterize slow reaction to market and non-market factors using only two parameters: a1 is low when returns react slowly to market return innovations, and a2 is low when returns react slowly to non-market news. We obtain daily returns produced by the smallest decile of US stocks during the period 1926-2012 from the data library of Ken French. Two sets of moments are used to calibrate the model: 1) Betas from univariate regressions of daily excess returns on the excess market return with lags from 0 to 42 days, and 2) Autocorrelations with lags from one to 42 days. These moments allows us to determine the slow adjustment of small stock prices to both market news and undiversifiable non-market news. Panel A in Figure 1 shows that positive exposure to lagged market shocks persists for up to two trading months (dashed line). Our model can produce a similar pattern in lagged market betas (solid line) when a1 = 0.35. Thus, our calibration implies that market information is instantaneously incorporated into small stock returns only 35% of the time and that 65% of the time the shock is delayed, potentially so severely that the half-life of the news is 32 days. Panel B in Figure 1 shows that positive autocorrelations in small stock returns persist for up to two trading months (dashed line). To calibrate our model to this pattern, a second, undiversifiable, non-market factor with the same variance as the market factor is required. The small value of the parameter a2 = 0.1 indicates that, unconditionally, small stocks respond slowly to non-market news: Small stocks react instantaneously to non-market news only 5% of the time. In the vast majority of cases (95% of the time) non-market factor information diffuses only very slowly into small stock prices. 13 For comparison, Panels C and D of Figure 1 show lagged betas and autocorrelations for the largest decile of CRSP stocks. Neither the betas nor the autocorrelations display evidence of slow information diffusion, and a calibration of the model with a1 = a2 = 1 provides the best fit to this data. The magnitude of information delay required to explain lagged beta and autocorrelation for small stocks is much higher than can be produced by models of microstructure frictions. Boudoukh, Richardson, and Whitelaw (1994) consider the impact of nonsynchronous trading and show that such models produce geometrically declining autocorrelations, and that even extreme heterogeneity in daily non-trading probabilities ranging from 0 − 85% produces autocorrelations of only 18%. Our model, calibrated to small stock returns, shows that heterogeneity in information diffusion is necessary to explain both the relatively low short-horizon dependencies in returns and the high long-horizon dependencies. This pattern in returns is most plausibly generated by slow information diffusion into observed small-stock prices. 3. Horizon Effects in Empirical Applications We provide an analytical formula that shows how average returns scale across horizons. We then extend our analysis to the scaling of abnormal performance as measured by Jensen’s (1968) alpha and Fama and MacBeth (1973) coefficients, and develop a simple diagnostic tool that shows the impact of microstructure effects at different horizons. 3.1. Horizon Effects in Average Returns Consider the problem of an empiricist evaluating the performance of an investment strategy. A standard return decomposition identifies abnormal return as measured by alpha, the systematic market component, and a residual: R̄in − R̄f n = αin + βin R̄M n − R̄f n + εin , 14 (13) where R̄jn , j ∈ {i, M, f } , denotes the unconditional average n-period gross return R̄jn = E(Rj,t+1 . . . Rj,t+n ) for all t, (14) the index i denotes an arbitrary portfolio, M the market, f the risk-free rate, βin is the market exposure calculated from n-period returns, and εin is the component of returns that is uncorrelated with the market. Setting the base period to one day, R̄i1 gives the average return using daily returns, while R̄i,21 would reflect an average return calculated from n = 21 day periods, or approximately monthly returns. A standard convention for reporting mean returns is to rescale linearly to a different horizon, typically either a month or a year. We correspondingly define the rescaled mean returns RS R̄in ≡ 1 + n R̄i1 − 1 , (15) and the associated ratio of the linearly rescaled to buy-and-hold mean return RS νjn ≡ R̄jn /R̄jn . Rescaling a short-horizon mean to approximate a longer-horizon mean introduces a bias whenever the average net returns of the portfolio do not rescale one-for-one with time. For iid returns, the ratio νjn will be less than one due to compounding, but we generally expect this effect to be small. To see this, define the short-horizon return rescaled by compounding as n RSg R̄in ≡ R̄i1 (16) and the related ratio g RSg νjn ≡ R̄jn /R̄jn . g Under independence of returns, νjn = 1. Consider the rescaling of daily alphas to monthly, quarterly, or annual frequencies, or the rescaling of monthly alphas to quarterly or annual frequencies. In these cases, the effects of compounding alone are small, and 15 RSg g RS R̄in ≈ R̄in or equivalently νjn ≈ νjn .5 Hence, when the returns of the portfolio i are iid, linearly rescaling a daily to a monthly or annual alpha or a monthly to an annual alpha is innocuous. To identify and quantify the primary source of horizon effects in mean returns, we consider normally distributed, single period logarithmic returns: ri1,t = ln (Ri1,t ) ∼ N (µi , σi2 ). Assume that log returns aggregated over n periods have a normal distribu2 tion with variance σin , where n is the relevant horizon. These assumptions are exactly satisfied if ri1,t is a stationary ARMA(p, q) process with Gaussian innovations, and approximately hold in more general cases. We show: Proposition 3 The ratio of rescaled to buy-and-hold mean returns satisfies g νin ≈ νin ≡ RSg R̄in 2 = enσi (1−V Rin )/2 , R̄in 2 / (nσi2 ) is the variance ratio. The net return ratios are where V Rin ≡ σin net νin g,net ≈ νin RS R̄in −1 νin − 1 ≡ = νin + R̄in − 1 R̄in − 1 g RSg −1 νin R̄in − 1 g = νin + . ≡ R̄in − 1 R̄in − 1 The bias in rescaled returns relative to the buy-and-hold return is thus determined by the short-horizon variance σi2 and the variance ratio V Rin . When the variance ratio is net one, for example if returns are iid, then νin = νin = 1 and rescaled short-horizon return means approximate well longer-horizon buy-and-hold averages. Empirically, individual asset returns are typically negatively autocorrelated, implying V Rin < 1, consistent with the upward bias in Blume and Stambaugh (1983). For portfolios, the tendency for net positive spurious autocorrelations suggests νin < νin < 1, implying a downward bias in short-horizon average returns. RSg RS For example, assuming a one percent average monthly return, R̄i12 = 1.1268 and R̄i12 = 1.12. The magnitudes of the biases in (18) will then be tiny. Since νin = 1.12/1.1268 = 0.9939, the bias in alpha due to the first term of (18) will be (νin − 1)R̄in = −0.0061 ∗ .01, less than a basis point per month in absolute terms. 5 16 3.2. Horizon Effects in Alphas Horizon effects in mean returns give rise to horizon effects in performance measures. Consider the problem of an empiricist evaluating the performance of an investment strategy using Jensen’s (1968) alpha αin . A standard convention for reporting alphas is to rescale linearly to a different horizon, typically either a month or a year. For example, Ang and Kristensen (2011), Barber (2007), Lewellen and Nagel (2006), and Li and Yang (2011), compute alphas from daily returns, and rescale linearly to longer horizons. We correspondingly define the rescaled alpha RS αin ≡ nαi1 , (17) and show Proposition 4 The difference between the linearly rescaled and buy-and-hold alphas is RS αin − αin ≈ (νin − 1)R̄in − β1 (νM n − 1)R̄M n − (β1 − βn ) R̄M n − R̄f n . (18) Rescaling a short-horizon alpha to approximate a longer-horizon alpha introduces a bias whenever the average net returns of the portfolio or the market index do not rescale one-for-one with time, or the betas calculated using different return frequencies are not identical. A substantial literature investigates the measurement of betas across different horizons (e.g., Dimson, 1979; Scholes and Williams, 1977), and our paper does not address this issue. Rather, we focus on the scaling of mean returns across horizons. In practice, however, horizon effects in alphas can be substantial. To preview our empirical results, for the portfolio of small stocks, the daily alpha rescaled to a monthly RS frequency is αsmall,21 = 43 basis points, while the alpha obtained from using buy- and-hold returns of 21-day “months” is almost 50% larger, αsmall,21 = 61 basis points. Rescaling the alpha from a monthly to a quarterly frequency also substantially increases the alpha. 17 A common view in the empirical finance literature is that alphas calculated at a daily frequency represent the profits available to an investor with a daily investment horizon, that monthly alphas represent performance from the perspective of an investor with a monthly horizon, and so forth. This view implicitly assumes that investors can transact at observed daily prices. Following arguments in ABK, even if a subset of investors is able to implement such daily rebalancing, gains and losses from the trades sum to zero across all investors. Consequently, rescaled alphas can provide an unbiased measure of profitability for at best only a subset of short-horizon investors who successfully transact at observed prices. Importantly, such performance measures will not properly capture profitability of investors in aggregate. As is clear from (13), any bias in measured average returns of the portfolio or the index due to microstructure effects will bias the calculation of alpha. 3.3. Horizon Effects in Fama-MacBeth Regressions The choice of return horizon also impacts Fama and MacBeth (1973, FM) coefficients and their interpretation. Consider cross-sectional regressions: Rin,t − 1 = ant + bnt Xit + int (19) where for simplicity Xit is a univariate characteristic for stock i at time t, Rin,t is an nperiod gross return on stock i starting at time t, and int is uncorrelated and mean zero. Fama (1976, p. 328) shows that the time-series of FM coefficients bnt can be interpreted as payoffs on a zero-cost investment strategy with portfolio weights proportional to the characteristic.6 To see how the estimated FM coefficients change with time scale, consider estimating (19) using n = 1 and n = 2 period returns. The conditional expectation of a 1-period 6 For example, suppose Xit is a small stock indicator. Then, ant measures the realized return of a large stock portfolio, and bnt represents the realized size premium. 18 return is E (Rin,t | Xit ) = 1 + a1 + b1 Xit . (20) The conditional expectation of the two-period return is E (Rin,t Rin,t+1 | Xit ) = E (1 + a1t + a1t+1 + b1t Xit + b1t+1 Xit+1 | Xit ) +E (a1t a1t+1 | Xit ) +E (a1t b1t+1 Xit+1 + a1t+1 b1t Xit | Xit ) +E (b1t b1t+1 Xit Xit+1 ) | Xit ) . (21) If the time-series of FM coefficients are serially independent and uncorrelated with the characteristic Xit , only the first term of Equation (21) is substantially different from zero. In this case, provided the characteristic Xit is approximately constant across periods, the coefficient means scale approximately linearly, i.e., a2 ≈ 2a1 and b2 ≈ 2b1 , where an and bn respectively denote the time-series averages of the regression coefficients ant and bnt . Serial correlations in the time-series of FM coefficients can be driven by microstructure noise. For example, in the case where positive autocorrelations and cross-serial correlations are driven by asynchronous price adjustment, we expect the short-horizon coefficients to be downward biased, 2a1 < a2 , 2b1 < b2 .7 Generally, different types of microstructure noise could lead to different patterns in the autocorrelations of the FM coefficients. In recent work, ABK focus on the specific case of noise in prices that is uncorrelated with fundamentals. They show that weighting observations i by their lagged returns can correct the bias caused by iid noise in prices. The ABK weighting scheme would not work, however, to correct bias in FM coefficients caused by asynchronous price adjustment, as such a weighting scheme does not address the autocorrelation terms in (21). To account for both iid measurement error, as in ABK, and asynchronous price 7 For example, Xit inversely related to size ensures cross-serial correlations will be positive since large stocks typically lead small stocks. 19 adjustment, one would need to use the ABK weights as well as choose a return horizon long enough to minimize the importance of autocorrelations and cross-autocorrelations in the FM coefficients. 3.4. Diagnosing Microstructure Frictions and Choosing the Return Horizon We propose a simple procedure to diagnose biases in average returns and choose a horizon for empirical analysis. Modifying and extending the notation developed in Section 3.1 to accommodate greater empirical flexibility, consider geometrically rescaling an n-period return to a horizon of m periods: m/n R̄inm ≡ [E(Ri,t+1 . . . Ri,t+n )]m/n ≡ R̄in . (22) If returns are iid, R̄inm = R̄imm ≡ R̄im for all n, m > 0. (23) This suggests using a plot of R̄inm versus the return-period length n as a diagnostic. When this plot is approximately flat, then rescaling effects have a small impact. Figure 2 demonstrates this diagnostic technique using data simulated from our Section 2 model using both the small and large stock calibrations. Panel A clearly shows that for small stocks, where evidence of slow information diffusion is strong, value weighted portfolio return means are downward biased at short horizons. This bias is highly economically significant, with rescaled daily mean returns a full 40 basis points per month lower than the rescaled annual mean returns. On the other hand, Panel B shows that for the large stocks, where lagged betas and return autocorrelations provided no evidence of slow information diffusion, no horizon effects are present in average holding period returns. 20 4. Empirical Evidence in U.S. Indices and Style Portfolios We show the magnitudes of the rescaling biases in attribute-sorted portfolios and market indices, and implement empirical methods to avoid these biases. 4.1. Data Using data from CRSP and Compustat, we form portfolio returns on the basis of the following characteristics: Market Equity: The market value of the equity of the firm at the end of calendar year τ − 1 is used to form portfolios beginning in July of year τ . Book-to-Market: The ratio of the book value of equity to the market value of equity, using the end of the previous calendar year book and market equity values.8 Similar to Fama and French (1993), the book-to-market ratio at the end of calendar year τ − 1 is used to form portfolios starting in July of year τ . Momentum: The cumulative return of individual stocks in the calendar months t − 12 to t − 2 is used to form portfolios starting in month t. Price: The price per share at the end of month t − 2 is used to construct portfolios beginning in month t. Short-term reversal: The return in month t − 2 is used to form portfolios starting in month t. Volatility: The standard deviation of daily stock returns estimated from months t − 13 to t − 2 is used to form portfolios starting in month t. 8 Book equity used to calculate the book-to-market ratio is defined as stockholders’ book equity plus balance sheet deferred taxes plus investment tax credit less the redemption value of preferred stock. If the redemption value of preferred stock is not available, we use its liquidation value. If the stockholders’ equity value is not on Compustat, we compute it as the sum of the book value of common equity and the value of preferred stock. Finally, if these items are not available, stockholders’ equity is measured as the difference between total assets and total liabilities. 21 Illiquidity: The Amihud (2002) price impact measure defined as average ratio of absolute daily returns to daily dollar volume, estimated in year τ − 1 is used to construct portfolios beginning in year τ . Z-Score: The bankruptcy predictor introduced in Altman (1968) and measured in fiscal year ending in calendar year τ − 1 is used to form portfolios beginning in July of year τ . For each attribute, we construct decile portfolios and study the returns of the top and bottom groups. We calculate daily time series of each decile under three weighting schemes.9 First, we compute initially equal-weighted (IEW) returns by investing $1 in each stock at the beginning of each rebalancing period and holding the resulting portfolio until the next rebalancing period.10 Second, we calculate initially value-weighted (IVW) returns, where, at the beginning of each rebalancing period, we invest in each stock an amount proportional to its most recent market capitalization.11 Under the first two weighting schemes, market equity, book-to-market, illiquidity, and Z-score portfolios are rebalanced annually, and the other portfolios are rebalanced monthly. Finally, we compute equal-weighted (EW) returns by rebalancing stocks in a portfolio to equal weights daily. 4.2. Horizon Effects in Average Returns and Alphas Table 1 compares average daily, monthly (21 day), quarterly (63 day), semi-annual (126 day), and annual (252 day) returns, rescaled geometrically to different horizons. We 9 The sample periods are 1964-2009 for the book-to-market and Z-score portfolios, 1927-2009 for the momentum portfolio, and 1926-2009 for the remaining portfolios. 10 See Asparouhova, Bessembinder, and Kalcheva (2010) for discussion of initially equal-weighted portfolios. Many empirical studies construct IEW portfolios by allocating stocks to portfolios according to some stock characteristic, with equal portfolio weights at the initial formation (e.g., Greenwood (2005); Brennan and Wang (2010); Sadka (2010); Kaniel, Ozoguz, and Starks (2012)). Event studies evaluating BHARs also use this weighing scheme (e.g., Loughran and Ritter (1995)). 11 Note that IVW differs from conventional value-weighting if the market capitalization changes for reason unrelated to stock performance, such as share issuance and repurchases. 22 consider the CRSP market index, and decile 1 and 10 portfolios of stocks sorted by market equity, book-to-market ratio, and momentum. Following our theoretical analysis, we expect rescaled daily returns to be smaller than buy-and-hold returns of longer horizons, with the downward bias being largest for portfolios of illiquid and volatile stocks. The results are consistent with this prediction. For the initially equal-weighted small stock portfolio, the geometrically rescaled daily means are substantially below the buy-and-hold averages for monthly returns (1.41 percent versus 1.63), quarterly returns (4.30 versus 5.73), and annual returns (18.34 versus 25.87). The biases are even larger for initially equal-weighted momentum losers, with geometric rescaled daily versus buy-and-hold returns of 0.67 versus 1.00 at a monthly horizon, and 8.34 versus 14.97 at an annual horizon. The differences between rescaled and buy-and-hold returns are smaller, although still meaningful, for portfolios of large stocks and initially value-weighted portfolios. These results confirm our prediction of systematic biases in mean returns calculated from daily returns. Importantly, these biases do not completely disappear at a monthly return horizon. For example, the annualized monthly return of the initially equal-weighted momentum loser portfolio is 12.72, while the corresponding measures of quarterly, semi-annual, and annual returns are all relatively similar between 14.64 and 14.97. The effects of rescaling are thus strong in daily returns and remain relevant in monthly returns. In the Appendix we show robustness of the results of Table 1 to linear instead of geometric rescaling. As anticipated in Section 3, the results are very similar to those with geometric rescaling. Since linear rescaling easily accommodates statistical tests based on means, is more common empirically, and as shown in Section 3 is implicit in performance regressions that calculate alphas at different horizons, in the remainder of tables we use linear rescaling. In Table 2 we test the significance of differences in linearly rescaled (RS) versus 23 buy-and-hold (BH) returns, and demonstrate the accuracy of our analytical approxnet net is close to the developed in Section 3. The statistic νin imation of the ratio νin estimated RS/BH ratio for all portfolios, validating the accuracy of the approximation given in Proposition 3. The results also show that the rescaled returns are below the buy-and-hold returns for all portfolios and horizons considered. Even apparently small biases, for example for the large stock portfolio (−0.02 monthly, −0.15 quarterly) are statistically significant, indicating that the return difference is stable over time. Many of the portfolios show much larger differences between rescaled short-horizon returns and longer-horizon returns. For example, in Panel B the rescaled monthly return average consistently understates the quarterly averages, with the largest differences for portfolios of low-priced stocks (1.21 quarterly) and small stocks (0.82), and sizeable differences for other portfolios (e.g., 0.24 for value stocks). Interestingly, portfolios with the highest autocorrelation do not necessarily have the largest differences between rescaled and buy-and-hold returns. For example, the low liquidity portfolio has high autocorrelations (0.20 in daily returns and 0.17 in monthly returns), but the low volatility of this portfolio keeps horizon effects contained to about 0.40 quarterly. The portfolio of liquid stocks exhibits similar horizon effects, achieved through a higher return volatility paired with lower autocorrelation, so that the return spread (high illiquidity minus low illiquidity) is nearly unaffected by the choice of horizon. In contrast, the high- and low-volatility portfolios have comparable daily autocorrelations (0.20 vs 0.17), but the standard deviation of the high-volatility portfolio is about three times larger than the low-volatility portfolio. As a result, an apparently small difference in means based on daily returns (3.03 − 2.62 = 0.42 when rescaled to a quarterly frequency) translates to a large difference in actual quarterly returns of 4.52 − 2.79 = 1.73. Table 3 clarifies how the rescaling bias we focus on in this paper differs from the ef- 24 fects of iid measurement error. Focusing on the small stock portfolio, the average daily return rebalanced to equal weights (RB = 0.24) substantially exceeds the average daily return of the buy-and-hold portfolio (BH = 0.07). This is the measurement error bias of Blume and Stambaugh (1983), and is highly significant. At monthly and annual horizons, the difference between the returns of rebalanced portfolios and the buy-and-hold portfolio continues to grow. For example, the annual return of the rebalanced portfolio is 138.8, while the buy-and-hold return is 25.87. The biases caused in multi-period returns by frequent portfolio rebalancing are discussed by Roll (1983) and subsequent authors. The rescaling bias that we focus on can be seen in the difference between the rescaled (RS) and buy-and-hold (BH) returns at monthly and annual horizons. In contrast to the upward biases caused by iid measurement error and portfolio rebalancing, the rescaled returns are downward biased, as indicated by the negative signs and statistically significant t-statistics in all entries of the row RS-BH for long-only portfolios. We show how horizon effects in mean returns translate into horizon effects in alphas. Table 4 presents linearly rescaled daily alphas and monthly alphas for the attributesorted portfolios. We decompose the difference in the two alphas into return bias, factor bias, and beta bias, using Equation (18). The largest estimates for factor and beta bias are 0.05 and 0.06 respectively. The average return biases are larger, in the range of 0.10 to 0.40 monthly. Table 5 similarly decomposes the difference in alphas calculated from monthly and quarterly data. The raw return biases are again the most substantial contributor to the alpha bias. Thus, horizon effects in portfolio mean returns, which are the focus of our paper, are the dominant component driving alpha differences across horizons. 25 4.3. Choosing a Horizon In Figure 3, we follow the method proposed in Section 3 of plotting the rescaled returns R̄inm versus the return-period length n = 1, ..., 252 for fixed m = 21. All of the plots show the characteristic shape associated with asynchronous trade and partial price adjustment demonstrated in Section 3. For low n, the plots slope upward, and for larger n the plots flatten out and stabilize. In a number of portfolios (small stocks, momentum losers, high volatility, high inverse price, low reversal), the difference between R̄inm for n = 1 (daily) and the flat part of the graph is 40-50 basis points monthly, or more. Most of the portfolios show a horizon effect of at least 10 basis points monthly for n = 1. In many of the plots a strong upward slope is still apparent for monthly returns (n = 21). Only at a quarterly horizon do the plots reliably flatten out for all portfolios. Plotting the return-measurement interval versus rescaled returns, as shown in Figure 3, provides a useful diagnostic tool for empiricists. Choosing a measurement interval sufficiently long that it is on the flat portion of the rescaled return plot is a simple way to alleviate concerns about the impact of microstructure frictions on average returns. Conversely, using a return-measurement interval on the sloping portion of the rescaled return graph should suggest consideration of the effects of measurement errors. In future work we anticipate that the effects of microstructure frictions demonstrated here could be directly incorporated into empirical moment conditions when using higher-frequency data. 5. International Evidence International portfolios exhibit considerable heterogeneity in liquidity.12 In this section, we study horizon effects among region and country portfolios, as well as among style 12 Summarizing a large empirical literature, Bekaert, Harvey, and Lundblad (2007) note in their abstract: “Given the cross-sectional and temporal variation in their liquidity, emerging equity markets provide an ideal setting to examine the impact of liquidity on expected returns.” 26 portfolios constructed from international equities. From Datastream, we identify all available country and regional US-dollar-denominated MSCI indices. To be included in our sample, we require indices to have at least 10 years of valid data and eliminate overly redundant regions13 to produce our international return sample of 56 country and 49 regional indices. We supplement this international data with developed market style portfolios from Ken French’s website. We download 6 portfolios formed on size and book-to-market ratio and 6 portfolios formed on size and momentum for five markets: Asia Pacific excluding Japan, European, Global, Japanese, and North American. Following the methodology outlined on the website, we compute returns on small, big, high book-to-market, low book-to-market, winner, and loser portfolios for each market. Tables 6, 7, and 8 compare average monthly returns rescaled to annual frequency (RS) with average buy-and-hold annual returns (BH).14 Several observations from the Tables are particularly noteworthy. First, rescaling biases are significant. For example, the average difference between BH and RS returns is 4.1% per year in country portfolios and 2.4% in regional index portfolios. These magnitudes are economically significant and clearly call for caution in interpreting statistics calculated using average monthly returns. Second, rescaling biases in the international portfolios are systematic. Rescaled returns are lower than buy-and-hold returns for every country, every region, and every style portfolio. Figure 1 plots the histogram of the ratios of RS and BH returns and confirms this observation graphically: The ratio is always below 1 and for several indices approaches 0.5. Third, consistent with our theoretical predictions, horizon effects are stronger in 13 MSCI maintains a number of regional indices (e.g., Europe Excluding Ireland) that closely overlap with other broader indices that we study (e.g., Europe). For brevity, we exclude such regional indices. 14 Our focus on monthly returns is due to their wide use in the international finance. 27 smaller, less liquid markets where information diffusion is likely to be slow. For example, Table 6 shows that the average difference between BH and RS returns in countries classified as Developed by MSCI is 2% annually, whereas for the Emerging and Frontier countries the difference is 5% on average and exceeds 10% for several countries. Regional indices in Table 7 and international style portfolios in Table 8 exhibit similar patterns, with stronger biases in regions with emerging markets. For example, the bias in the small stock portfolio in Asia Pacific (ex. Japan), at 4.84% per year, is nearly six times larger than it is in North America. Finally, Table 8 also shows that the biases in the long and the short sides of the factor portfolios partially offset, and the bias in the net long-short factor portfolios is smaller. This suggests that biases in the average returns of country and regional indices will closely translate into biases in alphas. An empiricist measuring performance of international portfolios, particularly those in smaller, less liquid markets can inadvertently introduce substantial biases in average returns and alphas, even when monthly returns are used. To determine whether the differences between rescaled and buy-and-hold returns reflect an econometric bias or tradable profit opportunities, we compare horizon effects in non-investable market indices and investable exchange-traded funds. Table 9 summarizes the results for MSCI country indices from Datastream and the corresponding iShares country ETFs from CRSP. The average (median) difference between rescaled and buy-and-hold returns amounts to −1.87% (−1.33%) per year for indices. The magnitude is lower by more than a third for ETFs: −1.17% (−0.81%). Smaller horizon effects in ETFs suggest that their prices reflect information more fully than do index prices. The results presented in the table indicate that a considerable portion of the difference between rescaled and buy-and-hold returns of the indices is due to a non-tradable econometric bias. Importantly, the horizon effects are sizeable even in 28 ETFs, suggesting some room for profitable investment strategies as a consequence of slow information diffusion into stock prices. 6. Conclusion We establish theoretically, using standard models of asynchronous trading and partial price adjustment, that short-horizon average returns of value-weighted and other non-rebalanced portfolios are biased. In contrast to the existing literature on iid measurement error, the bias under these sources of microstructure friction is downward rather than upward, and impacts value-weighted portfolios more than equal-weighted portfolios. To explain these biases, we develop an analytical approximation linking the average returns of a portfolio over arbitrary horizons. Average returns over a long horizon are closely approximated using only a few simple moments of returns, namely, the short-horizon average return and variance, and the variance ratio of long- to shorthorizon returns. The formula explains why bid-ask bounce, which generates a variance ratio of long-horizon to short-horizon returns less then one, produces an upward bias in short-run returns. On the other hand, asynchronous price adjustment produces a variance ratio greater than one, and a downward bias in short-horizon portfolio returns. We show that, consistent with theory, buy-and-hold portfolios of small, illiquid, and volatile stocks have average returns that are substantially downward biased at a daily horizon. These biases remain significant in many cases even for monthly returns. We propose a simple solution to account for these biases. Plotting rescaled average returns versus the return horizon permits easy diagnosis of market microstructure frictions by a sharp slope at short-horizons. Choosing a sufficiently long return horizon on the flat portion of this graph ensures that pricing frictions should not cause a problem in measured return means, alphas, or Fama-MacBeth coefficients. Given the increas- 29 ing importance of short-horizon returns in empirical work, awareness of these biases is important, and we anticipate continued advances in empirical methods that explicitly account for microstructure frictions. 30 7. Appendix: Proofs Proof of Lemma 1 Let ft = µf + σf ξt where the random variables ξt are independent standard normals. Substituting this expression for ft in equation (7) produces equation (8). Since 0 < δ < 1, the process for Dt is AR(1) and the formulas for the unconditional mean and variance can be obtained by applying the expectation and variance operators to both sides of equation (8). Proof of Lemma 2 By assumption, p∗0 = p0 + D0 . To apply the inductive proof of the result, assume that p∗t−1 = pt−1 + Dt−1 . Equations (5), (6), and (7) yield pt + Dt = pt−1 + δ(Dt−1 + ft ) + (1 − δ)(Dt−1 + ft ) = pt−1 + Dt−1 + ft = p∗t−1 + ft = p∗t . Proof of Proposition 1 Formulas for the unconditional mean and variance of lagger returns follow by applying the expectation and variance operator to both sides of equation (6) and then substituting the expressions for E(Dt ) and Var(Dt ) from Lemma 1. The upper bound on lagger variance follows from the fact that 0 < δ < 1 < 2 − δ implies that Proof of Proposition 2 31 δ 2−δ < 1. Formula (11) is derived as follows: ∗ wt−1 πept−1 = ∗ πept−1 + (1 − π)ept−1 πept−1 +Dt−1 = πept−1 +Dt−1 + (1 − π)ept−1 πeDt−1 . = πeDt−1 + (1 − π) (24) (25) (26) Expression (12) follows from applying the expectation operator to equation (10) and applying the definition of covariance. Proof of Proposition 3 RSg We first observe that R̄in = en(µi +σi /2) . The RS measure depends only on the first 2 and second moments of daily returns. By contrast, buy-and-hold returns over a monthly horizon depend on how daily returns aggregate. In particular, by assumption of joint normality of daily log returns, the monthly log returns are also normally distributed, i.e., 2 2 BH ri,1 + · · · + ri,n ∼ N (nµi , σin ). As a consequence, the BH statistic is R̄in = enµi +σin /2 . The ratio of the two statistics is νin ≡ RSg R̄in 2 2 2 = e(nσi −σin )/2 = enσi (1−V Rin )/2 . BH R̄in (27) The ratio in the net returns is net νin ≡ RSg R̄in −1 νin − 1 = νin + BH . BH R̄in − 1 R̄in − 1 (28) Proof of Proposition 4 Scaling the short-horizon Jensen’s alpha, given by equation (13) in the case where n = 1, to its n-period long-horizon value yields the expression RS nαi1 ≡ αin = nR̄i1 − nR̄f 1 + β1 nR̄M 1 − nR̄f 1 = (nR̄i1 − 1) − (nR̄f 1 − 1) + β1 (nR̄M 1 − 1) − (nR̄f 1 − 1) = (1 + (nR̄i1 − 1)) − (1 + (nR̄f 1 − 1)) + β1 (1 + (nR̄M 1 − 1)) − (1 + (nR̄f 1 − 1)) = νin R̄in − νf n R̄f n + β1 νM n R̄M n − νf n R̄f n , 32 RS where νjn ≡ (1 + n(R̄j1 − 1))/R̄jn = R̄jn /R̄jn . Subtracting the long-horizon buy-and- hold alpha αin produces RS − αin = αin νin R̄in − νf n R̄f n + β1 νM n R̄M n − νf n R̄f n − R̄in − R̄f n + βn R̄M n − R̄f n = (νin − 1)R̄in − (νf n − 1)R̄f n − β1 νM n R̄M n − νf n R̄f n + βn R̄M n − R̄f n . This expression can be equivalently written RS αin − αin = (νin − 1)R̄in − β1 (νM n − 1)R̄M n + (βn − β1 ) R̄M n − R̄f n −(1 − β1 )(νf n − 1)R̄nf . (29) The approximation (18) follows if rescaling has an insignificant impact on the average return of the risk-free asset, νf n ≈ 1. 33 References Altman, E. I., 1968, “Financial Ratios, Discriminant Analysis and the Prediction of Corporate Bankruptcy,” Journal of Finance, 23, 589–609. Amihud, Y., 2002, “Illiquidity and Stock Returns: Cross-Section and Time-Series Effects,” Journal of Financial Markets, 5, 31–56. Ang, A., and D. Kristensen, 2011, “Testing Conditional Factor Models,” Journal of Financial Economics,, Forthcoming. Asparouhova, E., H. Bessembinder, and I. Kalcheva, 2010, “Liquidity Biases in Asset Pricing Tests,” Journal of Financial Economics, 96, 215–237. , 2012, “Noisy Prices and Inference Regarding Returns,” Journal of Finance,, Forthcoming. Badrinath, S. G., J. R. Kale, and T. H. Noe, 1995, “Of Shepherds, Sheep, and the Crossautocorrelations in Equity Returns,” The Review of Financial Studies, 8, 401–430. Barber, B. M., 2007, “Monitoring the Monitor: Evaluating CalPERS’ Activism,” Journal of Investing, pp. 66–80. Bekaert, G., C. R. Harvey, and C. T. Lundblad, 2007, “Liquidity and expected returns: Lessons from emerging markets,” Review of Financial Studies, 20(6), 1783–1831. Blume, M. E., 1974, “Unbiased Estimators of Long-Run Expected Rates of Return,” Journal of the American Statistical Association, 69, 634–638. Blume, M. E., and R. F. Stambaugh, 1983, “Biases in Computed Returns - an Application to the Size Effect,” Journal of Financial Economics, 12, 387–404. Boudoukh, J., M. P. Richardson, and R. F. Whitelaw, 1994, “A Tale of Three Schools: Insights on Autocorrelations of Short-Horizon Stock Returns,” Review of Financial Studies, 7, 539–573. Brennan, M. J., N. Jegadeesh, and B. Swaminathan, 1993, “Investment Analysis and the Adjustment of Stock Prices to Common Information,” Review of Financial Studies, 6, 799–824. Brennan, M. J., and A. W. Wang, 2010, “The Mispricing Return Premium,” Review of Financial Studies, 23, 3437–3468. Canina, L., R. Michaely, R. H. Thaler, and K. L. Womack, 1998, “Caveat Compounder: A Warning about Using the Daily CRSP Equal-Weighted Index to Compute Long-Run Excess Returns,” Journal of Finance, 53, 403–416. Chordia, T., A. Sarkar, and A. Subrahmanyam, 2011, “Liquidity Dynamics and CrossAutocorrelations,” Journal of Financial and Quantitative Analysis, 46, 709–736. Chordia, T., and B. Swaminathan, 2000, “Trading Volume and Cross-Autocorrelations in Stock Returns,” Journal of Finance, 55, 913–935. Cohen, L., and A. Frazzini, 2008, “Economic links and predictable returns,” Journal of Finance, 63(4), 1977–2011. Cohen, L., and D. Lou, 2012, “Complicated firms,” Journal of Financial Economics, 104(2), 383–400. Conrad, J. S., and G. Kaul, 1993, “Long-Term Market Overreaction or Biases in Computed Returns?,” Journal of Finance, 48, 39–63. 34 Dellavigna, S., and J. M. Pollet, 2009, “Investor Inattention and Friday Earnings Announcements,” Journal of Finance, 64(2), 709–749. Dimson, E., 1979, “Risk Measurement When Shares Are Subject to Infrequent Trading,” Journal of Financial Economics, 7, 197–226. Fama, E. F., 1976, Foundations of finance : portfolio decisions and securities prices. Basic Books, New York. Fama, E. F., and K. R. French, 1993, “Common Risk-Factors in the Returns on Stocks and Bonds,” Journal of Financial Economics, 33, 3–56. Fama, E. F., and J. D. MacBeth, 1973, “Risk, Return, and Equilibrium - Empirical Tests,” Journal of Political Economy, 81, 607–636. Greenwood, R., 2005, “Short- and long-term demand curves for stocks: theory and evidence on the dynamics of arbitrage,” Journal of Financial Economics, 75(3), 607 – 649. Hirshleifer, D., S. S. Lim, and S. H. Teoh, 2009, “Driven to Distraction: Extraneous Events and Underreaction to Earnings News,” Journal of Finance, 64(5), 2289–2325. , 2011, “Limited Investor Attention and Stock Market Misreactions to Accounting Information,” Review of Asset Pricing Studies, 1(1), 35–73. Hirshleifer, D., and S. H. Teoh, 2003, “Limited attention, information disclosure, and financial reporting,” Journal of Accounting and Economics, 36(1-3), 337–386. Hong, H., T. Lim, and J. C. Stein, 2000, “Bad news travels slowly: Size, analyst coverage, and the profitability of momentum strategies,” Journal of Finance, 55, 265–295. Hou, K., and T. J. Moskowitz, 2005, “Market frictions, price delay, and the cross-section of expected returns,” Review of Financial Studies, 18(3), 981–1020. Huberman, G., and T. Regev, 2001, “Contagious speculation and a cure for cancer: A nonevent that made stock prices soar,” Journal of Finance, 56(1), 387–396. Jacquier, r., A. Kane, and A. J. Marcus, 2003, “Geometric or Arithmetic Mean: A Reconsideration,” Financial Analysts Journal, 59, 46–53. , 2005, “Optimal Estimation of the Risk Premium for the Long Run and Asset Allocation: A Case of Compounded Estimation Risk,” Journal of Financial Econometrics, 3(1), 37 – 55. Jensen, M. C., 1968, “Performance of Mutual Funds in Period 1945-1964,” Journal of Finance, 23, 389–416. Kaniel, R., A. Ozoguz, and L. Starks, 2012, “The high volume return premium: Cross-country evidence,” Journal of Financial Economics, 103(2), 255–279. Klibanoff, P., O. Lamont, and T. A. Wizman, 1998, “Investor Reaction to Salient News in Closed-End Country Funds,” Journal of Finance, 53, 673–699. Lewellen, J., and S. Nagel, 2006, “The conditional CAPM does not explain asset-pricing anomalies,” Journal of Financial Economics, 82, 289–314. Li, Y., and L. Yang, 2011, “Testing Conditional Factor Models: A Nonparametric Approach,” Journal of Empirical Finance, 18(5), 972 – 992. 35 Liu, W., and N. Strong, 2008, “Biases in Decomposing Holding-Period Portfolio Returns,” Review of Financial Studies, 21, 2243–2274. Lo, A. W., and A. C. MacKinlay, 1990, “An Econometric Analysis of Nonsynchronous Trading,” Journal of Econometrics, 45, 181–211. Loughran, T., and J. R. Ritter, 1995, “The New Issue Puzzle,” Journal of Finance, 50, 23–51. Menzly, L., and O. Ozbas, 2010, “Market Segmentation and Cross-predictability of Returns,” Journal of Finance, 65(4), 1555–1580. Pastor, L., and R. F. Stambaugh, 2012, “Are Stocks Really Less Volatile in the Long Run?,” Journal of Finance, 67, 431 – 478. Pastor, L., and P. Veronesi, 2003, “Stock Valuation and Learning about Profitability,” Journal of Finance, 58, 1749–1789. Roll, R., 1983, “On Computing Mean Returns and the Small Firm Premium,” Journal of Financial Economics, 12, 371–386. Sadka, R., 2010, “Liquidity risk and the cross-section of hedge-fund returns,” Journal of Financial Economics, 98(1), 54 – 71. Scholes, M. S., and J. Williams, 1977, “Estimating Betas from Nonsynchronous Data,” Journal of Financial Economics, 5, 309–327. Tetlock, P. C., 2011, “All the News That’s Fit to Reprint: Do Investors React to Stale Information?,” Review of Financial Studies, 24, 1481 – 1512. 36 Table 1. Horizon Effects in Style Portfolios Performance Metric Daily Monthly Quarterly Semi-Annual Annual Daily Monthly Quarterly Semi-Annual Annual Daily Monthly Quarterly Semi-Annual Annual Daily Monthly Quarterly Semi-Annual Annual Daily Monthly Quarterly Semi-Annual Annual Daily Monthly Quarterly Semi-Annual Annual Daily Monthly Quarterly Semi-Annual Annual 1 day 0.04 0.04 0.04 0.05 0.06 0.07 0.08 Holding Horizon 1 mo 3 mo 6 mo 1 year 1 day A. CRSP Value-Weighted Index 0.84 2.55 5.17 10.61 0.86 2.62 5.30 10.89 2.62 5.31 10.90 5.33 10.94 10.97 Holding Horizon 1 mo 3 mo 6 mo B. Market Capitalization, Initially Value-Weighted Big 0.80 2.43 4.92 10.09 0.06 1.26 0.81 2.45 4.95 10.14 1.46 2.49 5.04 10.32 5.06 10.37 10.49 C. Market Capitalization, Initially Equal-Weighted Big 0.81 2.46 4.97 10.20 0.07 1.41 0.83 2.52 5.10 10.46 1.63 2.58 5.22 10.71 5.19 10.64 10.69 D. Book-to-Market, Value 1.16 3.52 7.17 1.20 3.63 7.39 3.64 7.42 7.31 Initially Value-Weighted E. Book-to-Market, Value 1.36 4.13 8.42 1.49 4.52 9.25 4.70 9.62 9.65 Initially Equal-Weighted 1.51 1.51 1.71 1.80 14.85 15.33 15.38 15.16 15.37 17.55 19.36 20.16 20.24 19.91 0.04 0.02 0.77 0.78 0.48 0.59 F. Momentum, Initially Value-Weighted Winners 4.60 9.41 19.71 -0.01 -0.15 4.61 9.44 19.77 0.03 4.72 9.67 20.27 9.76 20.47 20.26 G. Momentum, Initially Equal-Weighted Winners 5.23 10.74 22.62 0.03 0.67 5.51 11.32 23.93 1.00 5.74 11.82 25.04 12.05 25.56 24.69 Small 3.83 4.44 5.12 Small 4.30 4.98 5.73 Growth 2.33 2.37 2.42 Growth 1.44 1.79 1.97 Losers -0.46 0.09 0.37 Losers 2.02 3.04 3.47 1 year 7.80 9.07 10.49 10.43 16.22 18.97 22.09 21.95 22.65 8.78 10.21 11.78 11.78 18.34 21.46 24.94 24.94 25.87 4.72 4.79 4.90 4.96 9.65 9.80 10.04 10.16 10.52 2.90 3.62 3.98 3.82 5.89 7.37 8.11 7.79 7.82 -0.91 0.18 0.75 0.73 -1.81 0.37 1.51 1.47 1.82 4.09 6.17 7.07 7.21 8.34 12.72 14.64 14.93 14.97 Notes: This table reports average buy-and-hold returns (BH, on the diagonal) as well as geometrically rescaled short horizon returns (RSg, off-diagonal). Returns are calculated using periods of n = 1, 21, 63, 126, and 252 days corresponding to Daily, Monthly, Quarterly, Semi-Annual, and Annual frequencies, and are scaled to the corresponding Holding Horizon. For all calculations, Panel A uses the value-weighted CRSP index, and the remaining use either initially equally-weighed (IEW) or initially value-weighted (IVW) daily returns for size, value, and momentum portfolios. A complete description of the portfolios is provided in Section 4 of the text. Table 2. Decomposing Horizon Effects A. Daily Performance Metric, Monthly Holding Horizon RS BH RS-BH σRS ρRS VR RS/BH net νin RS BH RS-BH σRS ρRS VR RS/BH net νin Market Capitalization H L HL 0.81 1.40 -0.59 0.83 1.63 -0.80 -0.02 -0.23 0.21 [-2.02] [-4.53] [4.47] 1.09 1.45 0.08 0.13 1.17 1.91 0.97 0.86 0.97 0.88 Book-to-Market H L HL 1.35 0.48 0.87 1.49 0.59 0.89 -0.14 -0.12 -0.02 [-9.24] [-6.62] [-1.25] 0.88 1.22 0.25 0.19 2.59 1.77 0.91 0.80 0.91 0.80 Short-Term Reversal H L HL 1.01 1.01 0.00 1.16 1.25 -0.09 -0.15 -0.25 0.10 [-6.34] [-7.12] [4.42] 1.30 1.41 0.18 0.21 1.83 2.12 0.87 0.80 0.87 0.81 H 1.01 1.35 -0.34 [-7.75] 1.73 0.20 2.04 0.75 0.76 Volatility L HL 0.87 0.14 0.91 0.44 -0.04 -0.30 [-6.03] [-7.52] 0.66 0.17 1.75 0.96 0.96 Momentum H L 1.70 0.67 1.80 1.00 -0.10 -0.34 [-5.83] [-7.25] 1.35 1.47 0.15 0.25 1.49 2.40 0.94 0.67 0.95 0.68 H 1.04 1.12 -0.08 [-6.44] 0.92 0.20 1.83 0.93 0.93 HL 1.03 0.80 0.23 [5.82] Illiquidity L HL 0.97 0.08 1.04 0.09 -0.07 -0.01 [-3.68] [-0.48] 1.48 0.08 1.30 0.93 0.93 Inverse Price H L HL 1.80 0.96 0.84 2.21 1.00 1.21 -0.41 -0.04 -0.37 [-7.14] [-5.06] [-6.71] 1.47 0.96 0.23 0.13 2.58 1.37 0.81 0.96 0.83 0.96 H 1.16 1.23 -0.07 [-6.54] 0.87 0.17 1.83 0.94 0.95 Z-Score L 0.43 0.66 -0.23 [-7.65] 1.16 0.27 2.64 0.65 0.65 HL 0.73 0.57 0.16 [6.05] B. Monthly Performance Metric, Quarterly Holding Horizon RS BH RS-BH σRS ρRS VR RS/BH net νin RS BH RS-BH σRS ρRS VR RS/BH net νin Market Capitalization H L HL 2.50 4.90 -2.40 2.58 5.73 -3.15 -0.08 -0.82 0.75 [-1.80] [-2.33] [2.31] 5.39 9.18 0.08 0.22 1.14 1.44 0.97 0.86 0.98 0.90 Book-to-Market H L HL 4.46 1.78 2.67 4.70 1.97 2.73 -0.24 -0.19 -0.06 [-3.10] [-1.89] [-0.66] 6.49 7.42 0.29 0.14 1.27 1.21 0.95 0.91 0.96 0.91 Short-Term Reversal H L HL 3.49 3.76 -0.28 3.80 4.31 -0.50 -0.32 -0.54 0.22 [-2.97] [-2.42] [1.56] 8.03 9.41 0.23 0.19 1.26 1.30 0.92 0.87 0.93 0.90 H 4.06 4.52 -0.46 [-1.69] 11.31 0.19 1.17 0.90 0.92 Notes: Table continues on the next page. Volatility L HL 2.73 1.33 2.79 1.73 -0.06 -0.40 [-2.34] [-1.56] 4.01 0.08 1.18 0.98 0.98 Momentum H L 5.41 3.01 5.74 3.47 -0.33 -0.47 [-3.37] [-1.92] 7.53 10.41 0.15 0.20 1.26 1.21 0.94 0.87 0.96 0.90 H 3.37 3.54 -0.17 [-3.14] 5.72 0.17 1.25 0.95 0.96 HL 2.40 2.27 0.13 [0.62] Illiquidity L HL 3.11 0.26 3.31 0.23 -0.19 0.02 [-2.49] [0.48] 7.73 0.12 1.17 0.94 0.95 Inverse Price H L HL 6.62 2.99 3.63 7.83 3.07 4.76 -1.21 -0.08 -1.13 [-2.68] [-2.70] [-2.59] 10.84 5.12 0.23 0.10 1.41 1.15 0.85 0.97 0.90 0.98 H 3.70 3.82 -0.13 [-2.82] 5.37 0.13 1.19 0.97 0.98 Z-Score L 1.98 2.25 -0.27 [-1.52] 8.64 0.23 1.21 0.88 0.89 HL 1.72 1.57 0.15 [0.93] Table 2. Decomposing Horizon Effects, Continued C. Daily Performance Metric, Quarterly Holding Horizon RS BH RS-BH σRS ρRS VR RS/BH net νin RS BH RS-BH σRS ρRS VR RS/BH net νin Market Capitalization H L HL 2.43 4.21 -1.78 2.58 5.73 -3.15 -0.15 -1.51 1.37 [-2.67] [-3.16] [3.12] 1.09 1.45 0.08 0.13 1.33 2.76 0.94 0.74 0.95 0.79 Book-to-Market H L HL 4.04 1.43 2.61 4.70 1.97 2.73 -0.65 -0.54 -0.12 [-6.87] [-4.67] [-1.17] 0.88 1.22 0.25 0.19 3.30 2.14 0.86 0.73 0.88 0.73 Short-Term Reversal H L HL 3.03 3.02 0.01 3.80 4.31 -0.50 -0.78 -1.29 0.51 [-5.14] [-4.09] [2.66] 1.30 1.41 0.18 0.21 2.30 2.77 0.80 0.70 0.81 0.73 H 3.03 4.52 -1.49 [-4.61] 1.73 0.20 2.40 0.67 0.70 Volatility L HL 2.62 0.42 2.79 1.73 -0.18 -1.31 [-5.43] [-4.36] 0.66 0.17 2.06 0.94 0.95 Momentum H L HL 5.10 2.00 3.10 5.74 3.47 2.27 -0.64 -1.47 0.83 [-4.89] [-4.09] [2.62] 1.35 1.47 0.15 0.25 1.88 2.90 0.89 0.58 0.91 0.62 H 3.13 3.54 -0.40 [-5.00] 0.92 0.20 2.29 0.89 0.90 Illiquidity L HL 2.90 0.24 3.31 0.23 -0.41 0.01 [-3.46] [0.09] 1.48 0.08 1.52 0.88 0.89 Inverse Price H L HL 5.39 2.87 2.52 7.83 3.07 4.76 -2.44 -0.20 -2.24 [-3.91] [-5.26] [-3.73] 1.47 0.96 0.23 0.13 3.63 1.57 0.69 0.94 0.75 0.94 H 3.48 3.82 -0.34 [-5.50] 0.87 0.17 2.17 0.91 0.92 Z-Score L 1.29 2.25 -0.96 [-5.02] 1.16 0.27 3.20 0.57 0.58 HL 2.20 1.57 0.63 [3.84] Notes: This table reports and decomposes the horizon effect bias in the initially equally weighted buy-and-hold style portfolios. The rescaled (RS) performance measures are obtained from daily (in Panels A and C) or monthly (in Panel B) returns, and are rescaled and compared to monthly (in Panel A) or quarterly (in Panels B and C) 1 PNn Rnt − 1, where Rnt is the portfolio gross buy-and-hold (BH) returns. BH returns are computed as Nn t=1 n-day return, t indices n-day periods, Nn is the number of n-day intervals in the sample, and n = 1, 21, 63 correspond to daily, monthly, and quarterly periods, respectively. RS returns are computed by rescaling the daily (n1 = 1 in Panels A and C) or monthly (n1 = 21 in Panel B) portfolio returns to a corresponding monthly (n2 = 21 in Panel i 1 PNn2 hPτ n2 /n1 A) or quarterly (n2 = 63 in Panels B and C) frequency: τ =1 t=(τ −1)n2 /n1 +1 (Rn1 t − 1) . t-statistics for N n2 the difference between RS and BH returns are in square brackets. Also reported are the standard deviation σRS and autocorrelation ρRS of the short horizon (daily in Panels A and C, monthly in Panel B) returns, the ratio V R of variance of performance-measurement-horizon returns variance to variance of holding-horizon returns, as well net as the empirically measured bias (RS/BH) and its analytical approximation νin from Proposition 3. A complete description of the portfolios is provided in Section 4 of the text. Table 3. Blume and Stambaugh (1983), Roll (1983), and Horizon Effects Portfolio Rule D High M A BH 0.04 [5.25] 0.83 [5.00] 10.69 [4.87] RB 0.04 [5.69] 0.91 [5.27] 11.66 [5.11] 0.24 [25.7] 5.53 [13.3] RS 0.04 [5.25] 0.81 [4.90] 9.71 [4.54] 0.07 [6.80] RB-BH 0.00 [3.51] 0.08 [3.23] 0.97 [2.81] 0.17 [30.7] -0.02 [-2.02] -0.98 [-2.68] RS-BH Low D M A A. Market Capitalization 0.07 1.63 25.87 [6.80] [5.22] [4.00] D High – Low M A -0.03 [-3.22] -0.80 [-3.41] -15.18 [-2.85] 138.80 [3.70] -0.19 [-23.0] -4.62 [-13.1] -127.14 [-3.48] 1.40 [4.90] 16.84 [3.78] -0.03 [-3.22] -0.59 [-2.84] -7.13 [-2.27] 3.90 [17.6] 112.93 [3.21] -0.17 [-30.3] -3.82 [-17.6] -111.96 [-3.21] -0.23 [-4.53] -9.03 [-3.37] 0.21 [4.47] 8.05 [3.14] 0.04 [5.53] 0.89 [4.62] 12.08 [4.04] B. Book-to-Market 0.02 0.59 7.82 [2.00] [1.89] [1.81] BH 0.06 [7.79] 1.49 [5.30] 19.91 [4.55] RB 0.15 [18.0] 3.37 [10.8] 51.74 [7.20] 0.05 [4.58] 1.22 [3.84] 16.08 [3.56] 0.10 [12.9] 2.15 [10.5] 35.66 [6.56] RS 0.06 [7.79] 1.35 [4.86] 16.18 [4.47] 0.02 [2.00] 0.48 [1.52] 5.72 [1.40] 0.04 [5.53] 0.87 [4.54] 10.46 [3.60] RB-BH 0.09 [34.2] 1.88 [16.8] 31.83 [5.88] 0.03 [14.6] 0.63 [9.68] 8.26 [5.23] 0.06 [21.3] 1.26 [14.8] 23.58 [5.68] -0.14 [-9.24] -3.73 [-3.54] -0.12 [-6.62] -2.10 [-2.62] -0.02 [-1.25] -1.62 [-2.21] RS-BH BH 0.08 [8.93] 1.80 [7.78] 24.69 [6.81] C. Momentum 0.03 1.00 [3.19] [2.92] 14.97 [2.60] 0.05 [6.45] 0.80 [3.40] 9.72 [2.35] RB 0.10 [11.4] 2.30 [9.74] 32.79 [7.29] 0.15 [14.9] 3.59 [9.39] 65.37 [5.33] -0.05 [-5.94] -1.29 [-4.73] -32.58 [-3.14] RS 0.08 [8.93] 1.70 [7.42] 20.40 [6.60] 0.03 [3.19] 0.67 [2.06] 8.01 [1.83] 0.05 [6.45] 1.03 [4.77] 12.39 [4.73] RB-BH 0.02 [22.4] 0.49 [13.6] 8.10 [4.15] 0.12 [58.9] 2.59 [23.1] 50.39 [5.82] -0.09 [-45.6] -2.10 [-21.1] -42.29 [-5.46] -0.10 [-5.83] -4.29 [-5.26] -0.34 [-7.25] -6.96 [-3.23] 0.23 [5.82] 2.68 [1.39] RS-BH Notes: This table reports average returns (in percent) computed over daily (D), monthly (M), and annual (A) holding horizons for portfolios formed on the basis of market capitalization, book-to-market ratio, and momentum. BH are 1 PNn average returns of the annually-rebalanced initially equal-weighted buy-and-hold portfolio, Rnt − 1, Nn t=1 where Rnt is the portfolio gross n-day return, t indices n-day periods, Nn is the number of n-day intervals in the sample, and n = 1, 21, 252 correspond to daily, monthly, and annual periods, respectively. RB returns are computed similarly for portfolios whose components are rebalanced daily to equal weights. RS are average daily returns of the annually-rebalanced initially equal-weighed portfolio rescaled to monthly (n = 21) or annual (n = 252) horizons, i 1 PNn hPτ n τ =1 t=(τ −1)n+1 (R1t − 1) . Also shown are the average differences between RB and BH, and between RS Nn and BH performance measures. Corresponding t-statistics are in square brackets. Results are reported for the portfolios of high and low deciles as well as for their difference. A complete description of the portfolios is provided in Section 4 of the text. Table 4. Horizon Effects in Alphas: Daily vs. Monthly Horizon Alpha Daily t(Alpha) Beta Small Big S-B 0.43 -0.01 0.44 [2.54] [-0.38] [2.53] 1.31 1.01 0.30 Value Growth V-G 0.45 -0.49 0.94 [3.96] [-4.27] [7.11] 1.18 1.36 -0.18 0.64 -0.38 1.02 Winners Losers W-L 0.74 -0.52 1.25 [7.23] [-3.67] [8.00] 1.21 1.61 -0.40 0.69 0.15 0.54 [4.14] [3.37] [3.12] 1.48 0.95 0.54 Winners Losers W-L -0.01 -0.13 0.12 [-0.12] [-1.02] [0.87] 1.31 1.52 -0.21 High Low H-L -0.23 0.20 -0.43 [-1.36] [5.34] [-2.54] 1.70 0.69 1.01 0.07 0.23 -0.16 F. Volatility [0.29] 1.70 [4.57] 0.67 [-0.66] 1.03 High Low H-L 0.19 -0.06 0.25 [2.77] [-0.67] [2.59] 1.02 1.32 -0.30 0.26 -0.02 0.28 High Low H-L 0.30 -0.55 0.85 [4.18] [-3.58] [6.23] 1.08 1.41 -0.33 0.38 -0.30 0.68 High Low H-L Monthly t(Alpha) Beta A. Size 0.61 [2.69] 1.36 -0.01 [-0.39] 1.02 0.62 [2.69] 0.34 Alpha Comp. 1 Decomposition Comp. 2 Comp. 3 Alpha Bias -0.23 -0.02 -0.21 0.03 0.02 0.01 0.02 0.00 0.02 -0.18 0.00 -0.18 B. Book-to-Market [3.29] 1.02 [-2.34] 1.35 [5.59] -0.33 -0.14 -0.12 -0.02 0.01 0.01 0.00 -0.06 0.00 -0.06 -0.19 -0.11 -0.08 0.83 -0.22 1.05 C. Momentum [6.47] 1.19 [-0.97] 1.61 [4.61] -0.42 -0.10 -0.34 0.23 0.03 0.04 -0.01 -0.01 0.00 -0.01 -0.09 -0.30 0.21 1.04 0.18 0.86 D. Inverse Price [3.74] 1.54 [3.84] 0.92 [3.00] 0.62 -0.41 -0.04 -0.37 0.03 0.02 0.01 0.03 -0.02 0.05 -0.35 -0.03 -0.31 E. Short-Term Reversal 0.11 [0.80] 1.31 -0.15 0.09 [0.50] 1.50 -0.25 0.02 [0.12] -0.19 0.10 0.03 0.03 0.00 0.00 -0.01 0.01 -0.12 -0.22 0.10 -0.34 -0.04 -0.30 0.05 0.02 0.03 0.00 -0.01 0.01 -0.30 -0.03 -0.27 G. Illiquidity [3.55] 0.99 [-0.15] 1.32 [2.45] -0.33 -0.08 -0.07 -0.01 0.02 0.03 -0.01 -0.02 0.00 -0.02 -0.07 -0.04 -0.03 H. Z-score [3.74] 1.04 [-1.16] 1.33 [3.12] -0.29 -0.07 -0.23 0.16 0.01 0.01 0.00 -0.01 -0.03 0.02 -0.08 -0.25 0.17 Notes: This table reports the rescaled daily alpha, the one-month buy-and-hold alpha, as well as the decomposition of net BH the difference following Equation (18). Component 1 is the bias in portfolio returns, Comp. 1 ≡ (νin − 1)(R̄in − 1), net BH component 2 is the bias in factor returns, Comp. 2 ≡ −βi (νM n − 1)(R̄M n − 1), and component 3 is the beta bias, BH Comp. 3 ≡ −(βi − βin )(R̄M n − 1). Daily betas are estimated with 10 Dimson (1979) lags. A complete description of the test portfolios is provided in Section 4 of the text. Table 5. Horizon Effects in Alphas: Monthly vs. Quarterly Horizon Alpha Monthly t(Alpha) Beta Small Big S-B 0.93 0.02 -0.91 [1.42] [0.25] [-1.36] 1.90 0.99 -0.91 Quarterly t(Alpha) Beta A. Size 1.58 [1.72] 1.91 -0.02 [-0.28] 1.01 -1.61 [-1.71] -0.90 Value Growth V-G 1.42 -1.40 2.82 [2.59] [-2.90] [5.17] 1.47 1.59 -0.12 1.87 -1.17 3.04 Winners Losers W-L 2.18 -1.02 3.20 [5.65] [-1.53] [4.67] 1.36 1.82 -0.46 2.25 0.49 1.76 [2.74] [3.43] [2.07] 2.03 0.95 1.08 Winners Losers W-L -0.14 -0.08 -0.06 [-0.34] [-0.14] [-0.14] 1.58 1.72 -0.13 High Low H-L -0.34 0.69 -1.03 [-0.49] [4.55] [-1.43] 2.01 0.67 1.34 0.00 0.68 -0.69 F. Volatility [-0.01] 2.04 [3.76] 0.69 [-0.80] 1.35 High Low H-L 0.66 -0.09 0.75 [2.98] [-0.27] [2.16] 1.06 1.35 -0.29 0.73 -0.07 0.80 High Low H-L 1.03 -1.37 2.39 [3.46] [-1.76] [3.66] 1.12 1.73 -0.61 1.12 -1.00 2.12 High Low H-L Alpha Comp. 1 Decomposition Comp. 2 Comp. 3 Alpha Bias -0.82 -0.08 0.75 0.15 0.08 -0.07 0.02 0.04 0.02 -0.66 0.04 0.70 B. Book-to-Market [3.06] 1.22 [-2.31] 1.49 [4.87] -0.26 -0.24 -0.19 -0.06 0.05 0.06 0.00 -0.29 -0.12 -0.17 -0.47 -0.25 -0.23 2.46 -0.96 3.41 C. Momentum [5.69] 1.33 [-1.29] 1.96 [4.40] -0.63 -0.33 -0.47 0.13 0.10 0.14 -0.03 -0.06 0.25 -0.31 -0.29 -0.08 -0.21 2.91 0.58 2.33 D. Inverse Price [2.44] 2.25 [3.52] 0.90 [1.90] 1.35 -1.21 -0.08 -1.13 0.15 0.07 0.08 0.39 -0.09 0.48 -0.67 -0.11 -0.57 E. Short-Term Reversal 0.18 [0.43] 1.51 -0.32 0.05 [0.08] 1.86 -0.54 0.13 [0.27] -0.35 0.22 0.12 0.13 -0.01 -0.13 0.26 -0.39 -0.33 -0.16 -0.18 -0.46 -0.06 -0.40 0.06 0.02 0.04 0.05 0.04 0.00 -0.35 0.00 -0.35 G. Illiquidity [2.98] 1.07 [-0.20] 1.38 [2.04] -0.31 -0.17 -0.19 0.02 0.08 0.10 -0.02 0.02 0.05 -0.04 -0.07 -0.04 -0.03 H. Z-score [3.60] 1.12 [-1.21] 1.58 [2.89] -0.46 -0.13 -0.27 0.15 0.04 0.06 -0.02 -0.01 -0.18 0.17 -0.09 -0.38 0.30 Notes: This table reports the rescaled monthly alpha, the quarterly buy-and-hold alpha, as well as the decomposition net BH of the difference following Equation (18). Component 1 is the bias in portfolio returns, Comp. 1 ≡ (νin −1)(R̄in −1), net BH component 2 is the bias in factor returns, Comp. 2 ≡ −βi (νM n − 1)(R̄M n − 1), and component 3 is the beta bias, BH Comp. 3 ≡ −(βi − βin )(R̄M n − 1). Monthly betas are estimated with 3 Dimson (1979) lags. A complete description of the test portfolios is provided in Section 4 of the text. Table 6. Horizon Effects in Country Index Portfolios Country Argentina Australia Austria Belgium Brazil Canada Chile China Colombia Croatia Czech Republic Denmark Egypt Estonia Finland France Germany Greece Hong Kong Hungary India Indonesia Ireland Israel Italy Japan Jordan Kenya Malaysia Mauritius Mexico Lebanon Morocco Netherlands New Zealand Nigeria Norway Pakistan Peru Philippines Poland Portugal Russia Singapore Slovenia South Africa South Korea Spain Sri Lanka Sweden Switzerland Taiwan Thailand Turkey United Kingdom United States MSCI Category Frontier Developed Developed Developed Emerging Developed Emerging Emerging Emerging Frontier Emerging Developed Emerging Frontier Developed Developed Developed Developed Developed Emerging Emerging Emerging Developed Developed Developed Developed Frontier Frontier Emerging Frontier Emerging Frontier Emerging Developed Developed Frontier Developed Frontier Emerging Emerging Emerging Developed Emerging Developed Frontier Emerging Emerging Developed Frontier Developed Developed Emerging Emerging Emerging Developed Developed RS 25.50 12.35 11.07 13.12 30.50 11.66 19.83 6.35 21.77 11.92 16.18 14.37 20.17 18.03 12.99 12.27 12.08 9.41 20.33 18.68 13.71 22.71 5.88 8.14 8.54 11.07 4.70 25.42 12.80 24.64 24.13 16.95 11.56 13.43 8.84 20.80 14.94 13.51 22.34 13.21 24.02 5.62 28.75 15.29 8.45 15.75 14.25 11.47 12.72 16.02 12.75 12.48 16.06 26.66 12.26 10.45 BH 32.71 12.99 13.88 15.28 32.58 12.29 23.57 7.99 26.30 16.94 17.21 16.65 29.95 22.35 16.52 13.43 13.53 12.74 23.66 20.72 18.15 30.89 8.22 9.46 10.19 13.41 6.44 37.39 16.57 30.24 28.37 18.79 13.07 14.30 10.39 25.09 18.96 18.92 24.72 18.49 46.62 6.91 40.29 19.16 15.64 16.90 18.47 13.05 17.19 17.85 13.87 14.64 21.50 48.30 13.85 11.19 RS/BH 0.78 0.95 0.80 0.86 0.94 0.95 0.84 0.79 0.83 0.70 0.94 0.86 0.67 0.81 0.79 0.91 0.89 0.74 0.86 0.90 0.76 0.74 0.72 0.86 0.84 0.83 0.73 0.68 0.77 0.81 0.85 0.90 0.88 0.94 0.85 0.83 0.79 0.71 0.90 0.71 0.52 0.81 0.71 0.80 0.54 0.93 0.77 0.88 0.74 0.90 0.92 0.85 0.75 0.55 0.89 0.93 RS-BH -7.21 [-0.87] -0.64 [-0.86] -2.81 [-1.53] -2.16 [-2.38] -2.08 [-0.55] -0.63 [-1.26] -3.74 [-2.16] -1.64 [-0.82] -4.53 [-1.71] -5.03 [-1.24] -1.02 [-0.56] -2.28 [-2.49] -9.78 [-2.69] -4.32 [-1.71] -3.54 [-1.31] -1.16 [-1.64] -1.46 [-1.29] -3.32 [-1.89] -3.33 [-1.59] -2.04 [-0.78] -4.44 [-2.11] -8.18 [-1.63] -2.33 [-1.34] -1.32 [-1.49] -1.65 [-1.24] -2.34 [-2.00] -1.75 [-2.16] -11.97 [-1.15] -3.78 [-1.65] -5.61 [-1.63] -4.24 [-2.28] -1.85 [-0.50] -1.51 [-1.69] -0.87 [-1.86] -1.55 [-1.46] -4.29 [-0.98] -4.02 [-1.96] -5.41 [-1.72] -2.39 [-1.29] -5.28 [-2.52] -22.59 [-0.94] -1.29 [-1.58] -11.54 [-1.99] -3.87 [-1.71] -7.19 [-1.96] -1.15 [-1.02] -4.21 [-2.46] -1.58 [-1.66] -4.47 [-1.24] -1.83 [-2.41] -1.12 [-1.43] -2.16 [-1.47] -5.43 [-1.61] -21.64 [-1.60] -1.59 [-2.14] -0.74 [-2.57] Years 25 43 43 43 25 43 25 20 20 10 18 43 18 10 31 43 43 25 43 18 20 25 25 20 43 43 25 10 25 10 25 10 18 43 31 10 43 20 20 25 20 25 18 43 10 20 25 43 20 43 43 25 25 25 43 43 Notes: This table reports the horizon effect bias in the MSCI country index portfolios. Rescaled (RS) returns are average monthly returns mutiplied by 12. Buy-and-hold (BH) returns are average annual returns. t-statistic for the difference between RS and BH returns is shown in square brackets. Sample period ends in 2012 and spans the number of years shown in the last column. MSCI category is as of January 2013. Table 7. Horizon Effects in Regional Index Portfolios Regional Index All Country Americas All Country Asia All Country Asia Excluding Japan All Country Asia Pacific All Country Asia Pacific Excluding Japan All Country EAFE + Emerging Markets All Country Europe All Country Europe + Middle East All Country Far East All Country Far East Excluding Japan All Country Far East Excluding Japan and Hong Kong All Country Pacific All Country Pacific Excluding Japan All Country Pacific Excluding Japan and Hong Kong All Country World All Country World Excluding United States Brazil, Russia, India and China (BRIC) EAFE EAFE + Canada EAFE Excluding Japan EAFE Excluding United Kingdom Emerging Markets Emerging Markets Asia Emerging Markets Eastern Europe Emerging Markets Europe Emerging Markets Europe + Middle East Emerging Markets Europe, Middle East and Africa Emerging Markets Excluding Asia Emerging Markets Far East Emerging Markets Latin America EMU Europe European Union Far East Frontier Markets Frontier Markets Africa Frontier Markets Central and Eastern Europe + CIS Frontier Markets Europe, Middle East and Africa Frontier Markets Excluding Gulf Cooperation Council G7 Golden Dragon Nordic North America Pacific Pacific Excluding Japan South East Asia World World Excluding United States Zhong Hua RS 10.65 3.10 12.31 4.80 11.61 7.48 10.14 5.94 3.98 12.33 11.16 4.73 11.65 10.58 8.39 7.61 13.42 10.83 10.78 11.27 10.87 15.00 11.33 13.22 14.33 12.71 13.10 16.41 11.26 22.87 10.16 11.46 9.37 11.36 10.33 19.41 10.27 10.75 11.97 10.63 7.83 14.67 10.66 11.13 12.71 5.81 10.16 10.91 13.19 BH 11.66 4.61 15.44 5.95 14.29 8.37 11.08 7.42 4.98 15.17 14.32 5.80 14.15 12.98 9.22 8.53 18.58 12.10 11.98 12.43 12.22 18.38 15.11 16.56 18.42 16.64 16.05 20.36 14.76 27.46 11.00 12.58 10.38 13.90 14.49 24.91 16.26 14.82 16.51 11.44 9.61 16.54 11.40 13.30 14.62 9.02 11.03 12.06 16.60 RS/BH 0.91 0.67 0.80 0.81 0.81 0.89 0.91 0.80 0.80 0.81 0.78 0.82 0.82 0.82 0.91 0.89 0.72 0.90 0.90 0.91 0.89 0.82 0.75 0.80 0.78 0.76 0.82 0.81 0.76 0.83 0.92 0.91 0.90 0.82 0.71 0.78 0.63 0.73 0.73 0.93 0.82 0.89 0.94 0.84 0.87 0.64 0.92 0.90 0.79 RS-BH -1.02 [-2.68] -1.51 [-1.95] -3.13 [-2.21] -1.15 [-2.00] -2.68 [-2.36] -0.89 [-1.81] -0.94 [-1.73] -1.48 [-1.40] -1.00 [-1.71] -2.85 [-2.00] -3.16 [-2.09] -1.07 [-1.93] -2.50 [-2.25] -2.40 [-2.34] -0.83 [-2.20] -0.92 [-1.87] -5.17 [-2.48] -1.26 [-2.60] -1.20 [-2.65] -1.15 [-2.31] -1.36 [-2.36] -3.38 [-2.61] -3.79 [-2.43] -3.33 [-1.61] -4.09 [-2.05] -3.92 [-2.18] -2.95 [-1.88] -3.95 [-1.95] -3.50 [-2.10] -4.58 [-1.79] -0.84 [-1.49] -1.12 [-2.05] -1.01 [-1.67] -2.54 [-2.09] -4.16 [-1.96] -5.50 [-1.68] -5.99 [-1.81] -4.07 [-1.89] -4.53 [-2.10] -0.81 [-2.54] -1.78 [-1.33] -1.86 [-2.71] -0.74 [-2.64] -2.17 [-2.29] -1.90 [-2.13] -3.20 [-1.54] -0.87 [-2.99] -1.15 [-2.54] -3.41 [-1.78] Years 25 25 25 25 25 25 25 25 25 25 25 25 25 25 25 25 18 43 43 43 43 25 25 18 25 25 16 25 25 25 25 43 25 43 10 10 10 10 10 36 16 43 43 43 43 18 43 43 20 Notes: This table reports the horizon effect bias in the MSCI regional index portfolios. Rescaled (RS) returns are average monthly returns mutiplied by 12. Buy-and-hold (BH) returns are average annual returns. t-statistic for the difference between RS and BH returns is shown in square brackets. Sample period ends in 2012 and spans the number of years shown in the last column. Table 8. Horizon Effects in Developed Market Factor Portfolios RS BH RS-BH RS/BH RS BH RS-BH RS/BH RS BH RS-BH RS/BH RS BH RS-BH RS/BH RS BH RS-BH RS/BH Small Big SMB High Low HML A. Asia Pacific Excluding Japan -2.79 16.35 8.95 7.40 -0.68 20.46 12.82 7.64 -2.11 -4.12 -3.87 -0.24 [-2.42] [-2.40] [-2.58] [-0.32] 0.80 0.70 Winners Losers WML 11.21 16.04 -4.84 [-2.58] 0.70 14.00 16.72 -2.72 [-2.19] 0.84 8.92 12.66 -3.74 [-1.54] 0.70 17.63 22.34 -4.71 [-3.09] 0.79 -8.71 -9.68 0.97 [0.51] 9.04 10.92 -1.89 [-2.29] 0.83 9.86 10.86 -0.99 [-1.75] 0.91 -0.83 0.07 -0.89 [-2.32] B. Europe 11.80 6.82 13.48 8.50 -1.68 -1.69 [-1.94] [-2.74] 0.88 0.80 4.98 4.97 0.01 [0.01] 4.39 6.18 -1.80 [-1.71] 0.71 15.35 17.62 -2.26 [-3.29] 0.87 -10.96 -11.43 0.47 [0.53] 9.42 10.64 -1.22 [-1.88] 0.89 8.47 9.21 -0.74 [-1.81] 0.92 0.95 1.43 -0.48 [-1.30] C. Global 11.18 6.39 12.20 7.73 -1.01 -1.34 [-1.76] [-2.42] 0.92 0.83 4.80 4.47 0.33 [0.70] 5.81 7.00 -1.19 [-1.35] 0.83 13.26 14.85 -1.59 [-2.17] 0.89 -7.45 -7.85 0.40 [0.45] 2.90 4.21 -1.31 [-1.24] 0.69 3.10 3.83 -0.72 [-0.87] 0.81 -0.20 0.39 -0.59 [-0.73] D. Japan 5.96 0.29 6.56 2.51 -0.59 -2.21 [-0.71] [-1.39] 0.91 0.12 5.67 4.05 1.62 [1.12] 2.82 3.44 -0.62 [-0.56] 0.82 4.07 6.20 -2.13 [-1.24] 0.66 -1.25 -2.76 1.51 [0.80] 12.68 13.52 -0.84 [-1.21] 0.94 10.29 11.25 -0.96 [-2.18] 0.91 2.38 2.27 0.11 [0.25] 8.98 10.07 -1.09 [-1.15] 0.89 16.47 18.04 -1.57 [-1.76] 0.91 -7.49 -7.98 0.48 [0.46] E. North America 13.22 9.51 3.71 14.41 10.58 3.83 -1.19 -1.07 -0.13 [-2.20] [-1.61] [-0.25] 0.92 0.90 Notes: This table reports the horizon effect bias in the developed market factor portfolios. Rescaled (RS) returns are average monthly returns mutiplied by 12. Buy-and-hold (BH) returns are average annual returns. t-statistic for the difference between RS and BH returns is shown in square brackets. Sample period is 1991-2012. Table 9. Horizon Effects in Non-Investable and Investable Country Portfolios Country Australia Brazil Canada France Germany Hong Kong Italy Japan Malaysia Singapore South Africa South Korea Spain Sweden Switzerland Taiwan United Kingdom RS 12.34 21.71 12.16 8.66 9.79 9.09 7.31 1.62 8.19 9.96 18.07 20.92 10.92 13.01 9.63 9.95 6.59 MSCI Index BH RS/BH 14.11 28.73 13.85 9.48 10.94 10.60 8.11 2.69 12.06 12.85 18.97 22.81 11.66 15.71 10.29 10.86 7.92 0.87 0.76 0.88 0.91 0.90 0.86 0.90 0.60 0.68 0.78 0.95 0.92 0.94 0.83 0.94 0.92 0.83 RS-BH -1.76 -7.03 -1.68 -0.82 -1.15 -1.52 -0.80 -1.06 -3.87 -2.89 -0.89 -1.89 -0.74 -2.70 -0.66 -0.91 -1.33 [-1.31] [-1.54] [-1.58] [-1.13] [-1.20] [-1.06] [-0.75] [-1.30] [-1.27] [-1.44] [-0.84] [-1.23] [-0.66] [-1.68] [-1.23] [-0.52] [-1.48] RS 11.72 19.98 11.48 8.04 9.47 8.17 6.68 1.15 10.10 8.17 17.48 19.87 10.33 11.47 8.64 8.04 6.01 iShares ETF BH RS/BH 12.79 26.39 12.66 8.51 10.27 9.23 7.15 1.96 10.85 9.74 17.77 21.23 10.61 13.30 8.86 8.40 6.96 0.92 0.76 0.91 0.94 0.92 0.89 0.94 0.59 0.93 0.84 0.98 0.94 0.97 0.86 0.98 0.96 0.86 RS-BH -1.06 -6.41 -1.18 -0.47 -0.80 -1.06 -0.46 -0.81 -0.75 -1.56 -0.28 -1.35 -0.28 -1.83 -0.22 -0.36 -0.95 [-0.92] [-1.51] [-1.27] [-0.67] [-0.80] [-0.74] [-0.46] [-0.92] [-0.26] [-0.99] [-0.34] [-0.83] [-0.26] [-1.43] [-0.44] [-0.24] [-1.18] Index(RS-BH) vs ETF(RS-BH) -0.70 -0.62 -0.51 -0.35 -0.35 -0.46 -0.34 -0.25 -3.12 -1.32 -0.61 -0.54 -0.46 -0.87 -0.44 -0.55 -0.38 [-2.14] [-1.09] [-2.66] [-2.28] [-1.91] [-1.88] [-1.80] [-1.39] [-1.10] [-0.95] [-1.53] [-2.26] [-1.76] [-1.75] [-1.89] [-1.35] [-1.92] Years 16 12 16 16 16 16 16 16 16 16 9 12 16 16 16 12 16 Notes: This table reports the horizon effect bias in the non-investable MSCI country indices and the corresponding exchange-traded funds. It also shows the difference in biases in the indices and ETFs. Rescaled (RS) returns are average monthly returns mutiplied by 12. Buy-and-hold (BH) returns are average annual returns. t-statistics are shown in square brackets. The sample covers the number of years shown in the last column, ending in 2012. A. Beta vs Lag 0.5 B. Autocorrelation vs Lag 0.3 Simulated Small Stock 0.4 Simulated AR(1) matching first order AR Hyperbolic Small Stock 0.25 0.2 0.3 0.15 0.2 0.1 0.05 0.1 0 0 −0.1 −0.05 0 5 10 15 20 25 30 35 40 C. Beta vs Lag −0.1 0 5 10 Simulated Large Stock 0.15 0.2 25 30 35 40 Simulated AR(1) matching first order AR Hyperbolic Large Stock 0.25 0.2 0.3 20 D. Autocorrelation vs Lag 0.3 0.4 15 0.1 0.05 0.1 0 0 −0.1 −0.05 0 5 10 15 20 Lag 25 30 35 40 −0.1 0 5 10 15 20 Lag 25 30 35 40 Figure 1. Panels A and C plot slope coefficients from univariate regressions of daily excess returns on lagged excess market returns versus lag in days. Panels B and D plot autocorrelation coefficients versus lag in days. Red dashed lines are generated by returns from the smallest and largest deciles of US stocks during the period 1926-2012. Solid blue lines are generated by returns simulated from equation (2) with parameters chosen to match the empirical moments: σ1 = σ2 = 0.2, λ = 12, δ = (1, 0.5, 0.29, 0.156, 0.08, 0.04, 0.02), (pik1 , pik2 , ..., pik7 ) = (ak , (1 − ak )/6, (1 − ak )/6, ...(1 − ak )/6). Small Stocks: µ1 = µ2 = 0.06, β = [1, 1.5], a1 = 0.35, a2 = 0.1. Large Stocks: µ1 = µ2 = 0.05, β = [1, 1], a1 = 1, a2 = 1. A. Small Stocks 2 1.9 Mean Return 1.8 1.7 1.6 1.5 1.4 1.3 1.2 0 50 100 150 200 250 100 150 Holding Period (days) 200 250 B. Large Stocks 1.6 1.5 Mean Return 1.4 1.3 1.2 1.1 1 0.9 0.8 0 50 Figure 2. Panels A and B plot average simple returns versus holding period in days. Returns are simulated from equation (2) with parameters chosen to match the empirical moments: σ1 = σ2 = 0.2, λ = 12, δ = (1, 0.5, 0.29, 0.156, 0.08, 0.04, 0.02), (pik1 , pik2 , ..., pik7 ) = (ak , (1 − ak )/6, (1 − ak )/6, ...(1 − ak )/6). Small Stocks: µ1 = µ2 = 0.06, β = [1, 1.5], a1 = 0.35, a2 = 0.1. Large Stocks: µ1 = µ2 = 0.05, β = [1, 1], a1 = 1, a2 = 1. Monthly Return, Percent Big Small 1.20 2.00 1.10 1.90 1.00 1.80 0.90 1.70 0.80 1.60 0.70 1.50 0.60 1.40 0.50 1.30 0.40 0 21 42 63 84 105 126 147 168 189 210 231 252 1.20 0 21 42 63 84 Monthly Return, Percent Value Growth 1.90 1.00 1.80 0.90 1.70 0.80 1.60 0.70 1.50 0.60 1.40 0.50 1.30 0.40 1.20 0.30 1.10 0 21 42 63 84 105 126 147 168 189 210 231 252 0.20 0 21 42 63 84 Monthly Return, Percent Winner 1.40 2.10 1.30 2.00 1.20 1.90 1.10 1.80 1.00 1.70 0.90 1.60 0.80 1.50 0.70 21 42 63 84 105 126 147 168 189 210 231 252 0.60 0 21 42 Monthly Return, Percent High Inverse Price 1.40 2.50 1.30 2.40 1.20 2.30 1.10 2.20 1.00 2.10 0.90 2.00 0.80 1.90 0.70 21 42 63 84 105 126 147 168 189 210 231 252 Days in Compounding Period Figure 3. This figure continues on the following page. 63 84 105 126 147 168 189 210 231 252 Low Inverse Price 2.60 1.80 0 105 126 147 168 189 210 231 252 Loser 2.20 1.40 0 105 126 147 168 189 210 231 252 0.60 0 21 42 63 84 105 126 147 168 189 210 231 252 Days in Compounding Period Monthly Return, Percent High Short-Term Reversal Low Short-Term Reversal 1.60 1.60 1.50 1.50 1.40 1.40 1.30 1.30 1.20 1.20 1.10 1.10 1.00 1.00 0.90 0.90 0.80 0 21 42 63 84 105 126 147 168 189 210 231 252 0.80 0 21 42 63 Monthly Return, Percent High Volatility 1.30 1.60 1.20 1.50 1.10 1.40 1.00 1.30 0.90 1.20 0.80 1.10 0.70 1.00 0.60 21 42 63 84 105 126 147 168 189 210 231 252 0.50 0 21 42 63 84 Monthly Return, Percent Illiquid 1.40 1.40 1.30 1.30 1.20 1.20 1.10 1.10 1.00 1.00 0.90 0.90 0.80 0.80 0.70 21 42 63 84 105 126 147 168 189 210 231 252 0.60 0 21 42 63 Monthly Return, Percent High Z-Score 1.00 1.50 0.90 1.40 0.80 1.30 0.70 1.20 0.60 1.10 0.50 1.00 0.40 0.90 0.30 21 42 63 84 105 126 147 168 189 210 231 252 Days in Compounding Period 84 105 126 147 168 189 210 231 252 Low Z-Score 1.60 0.80 0 105 126 147 168 189 210 231 252 Liquid 1.50 0.70 0 105 126 147 168 189 210 231 252 Low Volatility 1.70 0.90 0 84 0.20 0 21 42 63 84 105 126 147 168 189 210 231 252 Days in Compounding Period Figure 3. This figure plots for different style portfolios average rolling n-day buy-and-hold returns scaled to a monthly equivalent, [E (Rt−n+1 · · · Rt )]21/n − 1. The number of days n in a compounding horizon is shown on the x-axis. The y-axis in each panel is scaled to accommodate a bias of up to 0.80%. A complete description of the portfolios is provided in Section 4 of the text. Computation details are in the Appendix. A. MSCI Country Indices Percent of Observations 20 15 10 5 0 0.50 0.55 0.60 0.65 0.70 0.75 0.80 0.85 0.90 0.95 1.00 0.85 0.90 0.95 1.00 0.90 0.95 1.00 B. MSCI Regional Indices Percent of Observations 40 30 20 10 0 0.50 0.55 0.60 0.70 0.75 0.80 C. Developed Markets Style Portfolios 30 Percent of Observations 0.65 20 10 0 0.50 0.55 0.60 0.65 0.70 0.75 0.80 0.85 Ratio of Average Rescaled Monthly Returns to Average Buy-and-Hold Annual Returns Figure 4. This figure plots histograms of the ratios of average monthly returns mutiplied by 12 (RS) to average buy-and-hold (BH) annual returns for three sets of portfolios: MSCI country index portfolios in Panel A, MSCI regional index portfolios in Panel B, and developed market style portfolios in Panel C.