Wavelet-Decoupled Spatiotemporal Network for Stock Return Prediction.

Area of Science:

• Quantitative finance
• Time series analysis
• Machine learning for finance

Background:

• Stock markets exhibit complex, noisy time series with both short-term fluctuations and long-term trends.
• Existing prediction models struggle to effectively differentiate these components, leading to interference and reduced accuracy.
• Capturing dynamic temporal dependencies and cross-stock information propagation while maintaining causal structure is crucial but challenging.

Purpose of the Study:

• To develop a novel network architecture for stock price prediction that effectively handles heterogeneous financial time series.
• To address the limitations of existing methods in distinguishing short-term fluctuations from long-term trends.
• To improve the accuracy and reduce uncertainty in stock return forecasting by modeling these components separately.

Main Methods:

• Proposed the Wavelet-Decoupled Spatiotemporal Network (WaveDSTN).
• Utilized wavelet transformation to decompose stock returns into high-frequency (fluctuations) and low-frequency (trends) components.
• Incorporated a Dual-Path Spatiotemporal Encoder to capture temporal dependencies and cross-stock information propagation while preserving causality.

Main Results:

• WaveDSTN demonstrated significant improvements over existing stock prediction methods.
• Explicitly modeling trend and fluctuation components enhanced predictive accuracy.
• The proposed method effectively reduced uncertainty in stock return forecasting.

Conclusions:

• Decomposing financial time series into trend and fluctuation components is beneficial for stock price prediction.
• WaveDSTN offers a robust framework for capturing complex spatiotemporal dynamics in financial markets.
• The findings suggest a promising direction for enhancing quantitative investment strategies through advanced deep learning techniques.