Equity premium forecasting with reliability-screened forward-looking signals.

Forecasting the equity risk premium is challenging because predictive relationships are unstable out of sample. We propose a two-stage framework that generates forward-looking signals from standard macro-financial predictors and admits them only when they satisfy predictor-level reliability criteria. In Stage 1, each predictor is forecast one step ahead to obtain an expected movement and an uncertainty proxy. In Stage 2, lagged predictors are augmented with the admitted signals and mapped into next-period excess returns using random forests, optionally combined with SHAP-guided screening and dimension reduction. Using monthly U.S. data from 1952 to 2024, we evaluate both benchmark-relative out-of-sample accuracy and the economic value of forecasts in a constrained mean-variance allocation. Selective admission improves out-of-sample accuracy and, in selected specifications, yields economically meaningful gains in risk-adjusted performance and drawdown control. Tail-conditional diagnostics show that these gains are concentrated disproportionately in downside market states, helping explain when statistical improvements translate into economic value. The main qualitative patterns remain robust across alternative learners, across the S&P 500 and the CRSP value-weighted market index, and under net of transaction cost evaluation and alternative portfolio volatility checks. Taken together, the findings suggest that forward-looking predictor information is most useful when admitted selectively on the basis of predictor-level reliability, with its practical value lying less in universal forecast improvements than in more robust and downside-sensitive equity premium forecasting.