A heterogeneous encoding disentangled representation network for financial time series forecasting.

Area of Science:

• Quantitative Finance
• Machine Learning
• Time Series Analysis

Background:

• Financial time-series forecasting faces challenges from non-stationarity and market noise.
• Current models often use single-stream architectures, limiting their ability to capture diverse temporal patterns.

Purpose of the Study:

• To propose HEDR-Net, a novel Heterogeneous Encoding Disentangled Representation Network for financial time-series forecasting.
• To address limitations in disentangling heterogeneous temporal patterns and maintaining contextual coherence in existing models.

Main Methods:

• HEDR-Net employs a structural feature decoupling strategy to model trend, fluctuation, and raw signals.
• Wavelet-guided decomposition separates sequences into distinct channels, encoded by specialized subnetworks (Mamba, TCN, iTransformer).
• A dual cross-attention mechanism and Sparse Mixture of Feature Experts module enhance inter-branch interaction and adaptive fusion.

Main Results:

• HEDR-Net consistently outperformed advanced models like PatchTST and iTransformer on stock market benchmarks.
• The model demonstrated superior forecasting accuracy, robustness, and cross-market generalisation.
• Results validate the effectiveness of structural decoupling and heterogeneous fusion for complex financial forecasting.

Conclusions:

• The proposed HEDR-Net architecture significantly enhances financial time-series forecasting.
• Structural decoupling and heterogeneous fusion are key to improving predictive performance and interpretability.
• HEDR-Net offers a robust solution for complex financial market prediction challenges.