Scholarlink

<p>Accurate stock price forecasting is challenging due to the inherent non-stationarity and complex interdependencies of financial markets. Traditional models often struggle to adapt to distribution shifts and frequently neglect the intricate interactions between multivariate stock indicators. To address these limitations, this paper proposes DSEN-STARformer, a novel framework that integrates dynamic distributional calibration with domain-specific sentiment analysis. First, we propose the Dynamic Slice-Exponential Weighted Moving Average Normalization (DSEN) module, which mitigates distribution shifts by adapting to local statistical properties. Second, we design the STARformer architecture, combining a centralized STAR module to efficiently capture inter-channel correlations and a Rformer module with ProbSparse Self-Attention to model long-range temporal dependencies. Furthermore, to incorporate market sentiment without look-ahead bias, we establish a rigorous feature engineering pipeline using the Chinese Financial Generative Pre-trained Model (CFGPT) with strict temporal alignment. Extensive experiments on 25 stocks across five major industries demonstrate the superiority of our approach. The DSEN-STARformer achieves a Mean Squared Error (MSE) of 0.014 and Mean Absolute Error (MAE) of 0.035, significantly outperforming state-of-the-art baselines while maintaining a competitive inference speed of 38&#xa0;ms/iteration.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Handling distribution shifts and integrating sentiment for stock price forecasting

  • 期刊论文

  • 发表日期：

    2026年3月17日

查看全文
  • Xin Wang,
  • Xiaochen Liu,
  • Zhiqing Yu,
  • Fenglong Kan,
  • Zhendong Gao

摘要

Accurate stock price forecasting is challenging due to the inherent non-stationarity and complex interdependencies of financial markets. Traditional models often struggle to adapt to distribution shifts and frequently neglect the intricate interactions between multivariate stock indicators. To address these limitations, this paper proposes DSEN-STARformer, a novel framework that integrates dynamic distributional calibration with domain-specific sentiment analysis. First, we propose the Dynamic Slice-Exponential Weighted Moving Average Normalization (DSEN) module, which mitigates distribution shifts by adapting to local statistical properties. Second, we design the STARformer architecture, combining a centralized STAR module to efficiently capture inter-channel correlations and a Rformer module with ProbSparse Self-Attention to model long-range temporal dependencies. Furthermore, to incorporate market sentiment without look-ahead bias, we establish a rigorous feature engineering pipeline using the Chinese Financial Generative Pre-trained Model (CFGPT) with strict temporal alignment. Extensive experiments on 25 stocks across five major industries demonstrate the superiority of our approach. The DSEN-STARformer achieves a Mean Squared Error (MSE) of 0.014 and Mean Absolute Error (MAE) of 0.035, significantly outperforming state-of-the-art baselines while maintaining a competitive inference speed of 38 ms/iteration.