Notes: This table reports predictive $R2$⁠, their standard deviation and statistical significance, and the annualized Sharpe ratios (SR) implied by Martin and Wagner’s (2019) and Kadan and Tang’s (2020) theory-based approaches and the five machine learning models. The standard deviation of the $Roos,s2×100$ (Std Dev) is calculated based on the annual test samples. The SR refer to a zero-investment strategy long in the portfolio of stocks with the highest excess return prediction and short in the portfolio of stocks with the lowest excess return prediction. The p-values are associated with a test of the null hypothesis that the respective excess return prediction has no explanatory power over the zero forecast, $E(Roos,s2)≤0$⁠. For Panel A, the investment horizon is one month, and for Panel B, it is one year. The RPE are computed at the monthly (EOM) frequency. The out-of-sample testing period starts in January 1996 and ends in November 2018 (Panel A) or December 2017 (Panel B), respectively. The features are rank-scaled as described in Section 3.6. The machine learning results are obtained using the long training scheme depicted in Figure 2.