<p>This study proposes a stock price prediction model based on the Residual-enhanced Channel-Spatial Attention Network (R-CSAN), which integrates channel-spatial adaptive attention mechanisms with residual connections to effectively capture the multidimensional complex patterns in financial time series. The R-CSAN adopts an encoder-decoder architecture, where the encoder extracts feature correlations from historical data through multiple layers of channel-spatial attention modules, and the decoder incorporates a masking mechanism to prevent future information leakage and introduces a cross-attention mechanism to model inter-market correlations. Experiments conducted on four cross-market stock datasets, including Amazon, Maotai, Ping An, and Vanke, demonstrate that R-CSAN significantly outperforms not only traditional baseline models such as ARIMA, LSTM, and CNN-LSTM, but also recent Transformer-based approaches like Informer, Autoformer, and iTransformer on metrics including RMSE, MAE, MAPE, <InlineEquation ID="IEq1"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="41598_2025_6885_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="21" /> </InlineMediaObject> <EquationSource Format="TEX">\(R^2\)</EquationSource> </InlineEquation>, and return on investment. The model reduces RMSE by 17.3–49.3% compared to traditional methods and 6.2–11.6% compared to Transformer variants, with the highest <InlineEquation ID="IEq2"> <InlineMediaObject> <ImageObject Color="BlackWhite" FileRef="41598_2025_6885_Article_IEq1.gif" Format="GIF" Height="16" Rendition="HTML" Resolution="72" Type="Linedraw" Width="21" /> </InlineMediaObject> <EquationSource Format="TEX">\(R^2\)</EquationSource> </InlineEquation> reaching 93.17% and an increase in return on investment to 482.64%. Ablation experiments confirm the critical contributions of each component, with the temporal module removal causing an average increase of 38.6% in RMSE and channel-spatial attention removal resulting in a 21.3% increase. Moreover, the model provides an interpretative analysis of features and temporal dimensions through attention weight visualization, offering insights into both indicator importance and critical time periods for prediction. In practical applications, R-CSAN’s outputs can be integrated into quantitative trading strategies including breakout trading, moving average crossover signals, and portfolio allocation optimization, providing a new paradigm for robust prediction in highly volatile markets.</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

RCSAN residual enhanced channel spatial attention network for stock price forecasting

  • WenJie Sun,
  • Ziyang Liu,
  • ChunHong Yuan,
  • Xiang Zhou,
  • YuTing Pei,
  • Cui Wei

摘要

This study proposes a stock price prediction model based on the Residual-enhanced Channel-Spatial Attention Network (R-CSAN), which integrates channel-spatial adaptive attention mechanisms with residual connections to effectively capture the multidimensional complex patterns in financial time series. The R-CSAN adopts an encoder-decoder architecture, where the encoder extracts feature correlations from historical data through multiple layers of channel-spatial attention modules, and the decoder incorporates a masking mechanism to prevent future information leakage and introduces a cross-attention mechanism to model inter-market correlations. Experiments conducted on four cross-market stock datasets, including Amazon, Maotai, Ping An, and Vanke, demonstrate that R-CSAN significantly outperforms not only traditional baseline models such as ARIMA, LSTM, and CNN-LSTM, but also recent Transformer-based approaches like Informer, Autoformer, and iTransformer on metrics including RMSE, MAE, MAPE, \(R^2\) , and return on investment. The model reduces RMSE by 17.3–49.3% compared to traditional methods and 6.2–11.6% compared to Transformer variants, with the highest \(R^2\) reaching 93.17% and an increase in return on investment to 482.64%. Ablation experiments confirm the critical contributions of each component, with the temporal module removal causing an average increase of 38.6% in RMSE and channel-spatial attention removal resulting in a 21.3% increase. Moreover, the model provides an interpretative analysis of features and temporal dimensions through attention weight visualization, offering insights into both indicator importance and critical time periods for prediction. In practical applications, R-CSAN’s outputs can be integrated into quantitative trading strategies including breakout trading, moving average crossover signals, and portfolio allocation optimization, providing a new paradigm for robust prediction in highly volatile markets.