MSTAN: A multi-scale temporal attention network for stock prediction

Stock price prediction remains a challenging task due to the inherent non-stationarity, multi-scale temporal dependencies, and complex cross-asset correlations in financial markets. In this paper, we propose MSTAN, a novel Multi-Scale Temporal Attention Network designed to model these spatiotemporal dependencies explicitly. MSTAN constructs multi-scale representations through a two-dimensional periodic reconstruction strategy and employs a Temporal Hybrid Attention mechanism to jointly learn local fluctuations and global trends jointly. Furthermore, MSTAN employs an adaptive module with channel-wise attention to dynamically capture inter-stock dependencies and integrates multi-scale features through a progressive coarse-to-fine fusion strategy. Extensive experiments across diverse datasets, including Chinese A-shares and the US market, demonstrate that MSTAN consistently outperforms state-of-the-art baselines, achieving MAE reductions of up to 28.6 %. Portfolio backtesting further validates its practical utility, showing superior risk-adjusted returns.