Macroeconomic Uncertainty and Nonlinear Information Flow in Stock-Return Predictability

A central problem in cross-sectional stock-return prediction is how to balance two competing forecasts built from the same firm characteristics: a stable long-horizon forecast that averages predictive signals over many years and a responsive short-horizon forecast that adapts quickly to recent market conditions. Existing combination rules set this balance using purely statistical criteria or a linear projection on macroeconomic uncertainty, implicitly assuming that uncertainty shifts the optimal balance by the same amount in all market states. This study examines whether uncertainty instead carries nonlinear and state-dependent information about the optimal balance, and whether exploiting that information improves out-of-sample prediction. Using monthly returns and 94 firm characteristics for U.S. common stocks from January 1990 to December 2025, together with five uncertainty indices covering policy, financial, real, macroeconomic, and market-implied uncertainty, we measure nonlinear and directional dependence with mutual information and transfer entropy and develop a nearest-neighbor forecasting rule, MI–kNN–LASSO, which sets the balance by comparing the current uncertainty environment with historically similar states. All five indices relate to the optimal balance more strongly than linear correlations suggest, with macroeconomic uncertainty providing the most reliable predictive signal. The proposed rule forecasts the combination state far more accurately than the linear benchmark and earns a long–short return of 63 basis points per month with a smaller maximum drawdown. Our method offers investors an interpretable and adaptive way to combine return forecasts.