Regime-Gated Return–Risk Fusion for Forward Stock Sharpe Prediction: A Leakage-Audited Evaluation

Forward risk-adjusted return prediction requires separating expected return from risk and adapting both to changing market states. Motivated by the return–volatility multi-task formulation of TFT-ASRO, we develop a causal regime-gated fusion model in which an observable market-volatility gate combines specialized ridge forecasts of future mean and dispersion into a 21-session Sharpe estimate, and we evaluate ten configurations on a seeded 36-asset, three-state market simulator with purged chronological splits. Replication over 20 independent market paths shows that rankings from a single holdout are unstable: random forests lead on the initial path (MAE 2.694) but average 4.053 across paths, whereas the gated model averages 3.764 and regularized ridge regression 3.447. Ablations isolate the effects of the gate, soft weighting and the return–risk decomposition, and identify the volatility proxy and the ratio reconstruction as the components that govern accuracy. The study establishes a leakage-audited protocol and concrete design guidance for regime-aware Sharpe forecasting.