Leveraging multi-time-span sequences and feature correlations for improved stock trend prediction

Yawen Li, Mengyu Zhuang, Junheng Wang, Jinyi Zhou

Abstract: Accurate stock trend prediction is critical for informed investment decisions and the stability of financial markets. However, existing methodologies often overlook fine-grained stock price volatility and fail to incorporate a comprehensive spectrum of technical indicators, inadequately capturing the complex interrelationships fundamental to technical analysis. This paper proposes MSFCE, a novel framework for stock market trend prediction that enhances feature correlations across multi-time-span sequences. Specifically, MSFCE designs a multi-scale feature encoder to capture both intraday and daily features, which are processed through a Transformer-based dimensionally adaptive encoder. Furthermore, the framework leverages higher-order interactions among technical indicators via a graph attention network, dynamically modeling their interdependencies to improve prediction robustness in dynamic markets. Extensive experiments on the SSE50 and CSI300 datasets demonstrate that MSFCE significantly outperforms existing state-of-the-art methods, consistently exhibiting superior performance across multiple test periods and market conditions. Its strong prediction accuracy and risk management suggest practical applicability in trading strategies, yielding significant excess returns in empirical backtests.