The profitability, costs and systematic risk of the post-earnings-announcement-drift trading strategy

Abstract

This paper re-examines the profitability of the post-earnings-announcement-drift (PEAD) trading strategy using a practical simulation approach that aligns with a fund manager’s investment perspective. It allows us to calculate the break-even transaction costs of following a PEAD strategy, and permits the explicit incorporation of transaction costs. Using US data from 1974 to 2007, we show that the traditional event-study method understates the risk and overstates the abnormal return of the PEAD strategy. Accounting for transaction costs in a practical simulation framework, we show there is no abnormal return (alpha) from the PEAD strategy in multi-factor asset pricing regression analyses. These results are robust to sub-period analyses and alternative transaction cost measures. The effects of intraday timing and information risk on the PEAD strategy are also explored. Overall, our study shows that the practical aspects of implementing the PEAD strategy are vitally important to evaluating the risk and return of the strategy. We provide a practical, analytical tool that can be directly adopted by fund managers to study the PEAD strategy with their institutional parameters of transaction costs and market timing.

Appendices

Appendix 1: Classification of earnings surprise

While Foster et al. (1984) construct two time-series models to calculate earnings surprise (standardized unexpected earnings—SUE) and measure the significance of PEAD, they also find that a simple seasoned random-walk model performs just as well. Because of the simplicity of the calculation, this model has been widely used by subsequent studies (for example Bernard and Thomas 1989; Chan et al. 1996) and is also adopted here as our main earnings classification method.

For each calendar quarter, stocks have been divided into 10 deciles based on the SUE, which is calculated as follows:

$$SUE_{i,t} = \frac{{e_{i,q} - e_{i,q - 4} }}{{\sigma_{i,t} }} $$

(7)

where e _i,q is the quarterly earnings observation before extraordinary items and discontinued operations for stock i at quarter q, and e _i,q−4 is the observation four quarters prior. σ _i,t is the standard deviation of (e _i,q –e _i,q−4 ) over the preceding eight quarters.

Once the SUE has been calculated for each calendar quarter, the cut-off points used to classify SUE into 10 deciles are based on the ranking of the previous quarter’s SUE. Foster et al. (1984) first used this method to resolve the cut-off point problem. At each calendar quarter, earnings announcements with the most negative surprise SUE values are assigned to Decile 1 and those with the most positive surprise SUE values are assigned to Decile 10.

Appendix 2: The limited dependent variable model for transaction costs

The limited dependent variable model for estimating transaction costs is introduced by Lesmond et al. (1999). The basic intuition of the model is that marginal investors will not choose to trade if the benefit of trading does not exceed transaction costs, and this will result in a zero return for a given day. Marginal investors can be informed traders or uninformed traders. For informed traders, they compare transaction costs with their information, and for uninformed traders, they compare transaction costs with their needs for liquidity. Therefore, when transaction costs stop marginal investors trading, we will observe a realized zero return which deviates from the true return generating process. Lesmond et al. (1999) use a market model to measure the true returns process as follows,

$$R_{jt}^{*} = \beta_{j} R_{mt} + \varepsilon_{jt} , $$

(8)

where $R_{jt}^{*}$ is the true return for security j at day t, and R _mt is the market return at day t. The relation between the observed return R _jt and the true return $R_{jt}^{*}$ is given by following equations,

$$\begin{array}{*{20}c} {R_{jt} = R_{jt}^{*} - \alpha_{1j} } & {if} & {R_{jt}^{*} < \alpha_{1j} } \\ {R_{jt} = 0} & {if} & {\alpha_{1j} < R_{jt}^{*} < \alpha_{2j} } \\ {R_{jt} = R_{jt}^{*} - \alpha_{2j} } & {if} & {R_{jt}^{*} > \alpha_{2j} } \\ \end{array} $$

(9)

where α _1j is the transaction cost threshold for trades on negative information (selling), and α _2j is the threshold for trades on positive information (buying). α _2j – α _1j is the measure for round trip transaction costs.

By assuming the error term follows a normal distribution N(0, σ _j ), the maximum likelihood method can be used to estimate the model with following log likelihood function,

$$\begin{aligned} Ln(L) & = \sum\limits_{1} {\ln \frac{1}{{(2\pi \sigma_{j}^{2} )^{1/2} }} - } \sum\limits_{1} {\frac{1}{{2\sigma_{j}^{2} }}} (R_{jt} + \alpha_{1j} - \beta_{j} R_{mt} )^{2} \\ & \quad + \sum\limits_{2} {\ln \frac{1}{{(2\pi \sigma_{j}^{2} )^{1/2} }} - } \sum\limits_{2} {\frac{1}{{2\sigma_{j}^{2} }}} (R_{jt} + \alpha_{2j} - \beta_{j} R_{mt} )^{2} \\ & \quad + \sum\limits_{0} {\ln (\Upphi_{2} \left( {\frac{{\alpha_{2j} - \beta_{j} R_{mt} }}{{\sigma_{j} }}} \right) - \Upphi_{1} \left( {\frac{{\alpha_{1j} - \beta_{j} R_{mt} }}{{\sigma_{j} }}} \right)} \\ \end{aligned} $$

(10)

where 0, 1, 2 represent the subsample of zero, negative and positive observed returns, respectively. Φ₁ and Φ₂ are standard normal distribution functions.