MIST-Net: A market–identity–spatial–temporal network for stock return prediction

Stock return prediction is a challenging task, which is often influenced by systemic market influence and idiosyncratic stock characteristics. Existing models often fail to jointly capture these factors while effectively integrating spatio-temporal information. In this paper, we propose a market–identity–spatial–temporal network (MIST-Net) to model market influence and individual information to enhance the prediction performance. To capture the systemic influence, we design a modulation mechanism based on FiLM that adjusts stock features according to market regimes. Regarding idiosyncratic characteristics, we utilize prior knowledge to constrain the identity embedding, which generate a data-driven adaptive graph. To enable the integration of spatio-temporal information and obtain predictions based on the characteristics of the stocks, our parallel architecture extracts temporal and spatial features independently, fusing them via bidirectional attention to capture instantaneous and cross-temporal correlations. Finally, learned identity embedding serve as personalized queries to aggregate fused features for prediction. Experiments on CSI100, CSI300, and CSI500 show that MIST-Net outperforms evaluated baselines, achieving an information coefficient of 0.058, 0.057, and 0.053 respectively. Investment simulations also show its resilience; it achieves an annualized return of 9.3% on CSI300 and maintains a positive return (1.9%) even on the bearish CSI500 dataset where most baselines suffer losses.