Do Anomalies Really Predict Market Returns? New Data and New Evidence
Nusret Cakici, Christian Fieberg, Daniel Metko, Adam Zaremba
Review of Finance, Volume 28, Issue 1, January 2024, Pages 1–44, https://doi.org/10.1093/rof/rfad025
Asset pricing literature typically views return predictability from two perspectives. The first strain examines cross-sectional predictability, exploring whether stock characteristics assist in answering why some stocks outperform others. The other focuses on the time-series dynamics of the market equity premium. Recent research has attempted to link these two seemingly disjointed lines of research, with some showing that cross-sectional anomaly strategies indeed contain information about future aggregate stock returns. In other words, past returns on long-short anomaly portfolios may help to predict the market risk premium.
In our paper, we comprehensively reexamine market return predictability by equity anomalies. Using new data from both U.S. and international markets, we examine hundreds of anomalies in 42 countries around the world. Our baseline sample contains more than 80,000 stocks, 10 million return observations, and one billion monthly stock characteristics. With this data at hand, we take advantage of machine learning models and forecast market equity premia based on anomaly portfolio returns. With a focus on variable selection and dimension reduction techniques, machine learning is well-equipped to tackle challenging prediction problems by reducing degrees of freedom and condensing redundant variation among predictors.
Our findings yield a simple yet unequivocal conclusion: equity anomalies, as such, cannot predict market returns. Any apparent predictability lacks external validity in two critical aspects: stock market selection and anomaly sample. While some evidence may be spotted in individual markets—such as the United States—it originates from a handful of specific anomalies and depends heavily on seemingly unimportant methodological choices.
The return predictability does not extend to international stock markets. Using data from 1990 to 2021, we feed our prediction models with up to 153 prominent anomaly portfolios. We find no robust evidence of market risk premium predictability. The predictive R2 coefficients are typically negative and confirmed by alternative measures, such as utility gains and Sharpe ratios. Further, our results hold for various methodological modifications, including different portfolio breakpoints, training windows, portfolio compositions, and anomaly sets.
Finally, we find that portfolio design affects prediction performance. An identical selection of anomalies supplied to the same forecasting model can have differing prediction accuracy. However, our baseline inference is that anomalies are not reliable predictors of market portfolio returns. Despite the hundreds of factor portfolios and research designs tested, only very few generated predictive abilities.