Stock market prediction and trading strategy optimization with S4-integrated deep learning architectures: a comparative study of hybrid fusion mechanisms
Ru Xia, Xing-Yi Wang, Shu-Yuan Du, Yan Xiao, Xiao-Fei Li, Jun-Jie Du
Abstract
The accurate prediction of stock market movements remains a challenging task due to the non-stationary, noisy and highly volatile nature of financial time series. While deep learning models such as Long Short-Term Memory (LSTM) networks, Multilayer Perceptron (MLP) and Transformer have demonstrated promising results, they often struggle to capture both long-term dependencies and short-term fluctuations efficiently. To address these limitations, this paper introduces al hybrid framework that integrates the Structured State Space Sequence (S4) model with three fundamental deep learning architectures via tailored fusion mechanisms. Specifically, we propose (1) LSTM-S4, a parallel architecture with attention-based fusion and skip connections; (2) MLP-S4, a serial embedding model that enables temporal modeling for MLP; and (3) Transformer-S4, a tandem structure that combines global attention with local temporal pattern extraction. Extensive experiments are conducted based on real-world data from six Chinese stocks covering diverse sectors and market conditions. The results demonstrate that the S4-integrated hybrid models outperform their base counterparts and existing baseline models in terms of predictive accuracy measured by RMSE, MSE, MAE and $Rˆ2$, as well as simulated investment returns. Notably, we design a practical trading strategy and verify its effectiveness through statistical testing. Paired-sample t-tests confirm that the improvement in returns is statistically significant. Compared to their base counterparts, the LSTM-S4, MLP-S4 and Transformer-S4 models yield t-statistics of −3.72, −3.96 and −2.97, respectively, with all p-values below 0.05 and the minimum p-value reaching 0.01. This conclusion confirms that after incorporating the S4 module, the three models not only enhance prediction performance but also enable the adoption of more profitable trading methods, effectively capturing complex patterns in financial time series. This underscores the significant potential of S4 to enhance temporal feature extraction in financial time series forecasting, offering a valuable tool for practical market prediction.
Source: semanticscholar
Read the AI summary and key takeaways for traders on WOBR Quant Research.