Predicting Winner and Loser Stocks: A Classification Approach – Roope Rihtamo

We introduce a binary response model framework to predict future winner and loser stocks in the cross-section of US stock returns. Instead of predicting returns as in the traditional Fama-MacBeth regression setup, we predict the sets of winners and losers directly with separate binary response models. We assess the predictive power of several past returns related anomalies and examine the time-variation in such predictive relations. Our out-of-sample forecasting results show that portfolio strategies based on these direct winner and loser predictions consistently outperform conventional benchmarks.  We also find that the well-documented decline in the predictive power of various anomalies since the millennium is less pronounced when applying our approach.