This paper proposes MRA-AGRU, a novel dynamic factor gating architecture for stock return forecasting with explicit market regime awareness. Unlike traditional static factor models, the proposed framework introduces a macroeconomically informed gating mechanism that adaptively reweights factors based on regime signals, combined with an attention-enhanced GRU for temporal modeling. Extensive experiments on the CSI 300 and NASDAQ 100 demonstrate improved predictive accuracy and robustness, including higher Sharpe ratio and reduced maximum drawdown. This work is publicly available as a preprint and represents ongoing research in quantitative finance and machine learning.
Wang et al. (Mon,) studied this question.