A Heterogeneous Encoding Disentangled Representation Network for Financial Time Series Forecasting

Financial time-series forecasting remains highly challenging due to non-stationarity, market noise, and complex dependencies across multiple temporal scales. Existing state-of-the-art models, although effective in long-sequence learning, often rely on single-stream architectures that struggle to disentangle heterogeneous temporal patterns and maintain contextual coherence. To address these limitations, this paper proposes HEDR-Net, a novel Heterogeneous Encoding Disentangled Representation Network that centres on a structural feature decoupling strategy and achieves coordinated modelling of trend, fluctuation, and raw signals. A wavelet-guided decomposition separates the input sequence into three semantically distinct channels, which are then encoded by structurally specialised subnetworks: Mamba for long-term trends, TCN for short-term fluctuations, and iTransformer for contextual dynamics. A dual cross-attention mechanism is introduced to enhance inter-branch interaction, followed by a Sparse Mixture of Feature Experts module that performs high-dimensional representation compression and adaptive fusion. Extensive experiments on multiple stock market benchmarks demonstrate that HEDR-Net consistently outperforms recent advanced models such as PatchTST and iTransformer, achieving superior forecasting accuracy, robustness, and cross-market generalisation. These results confirm the effectiveness of the proposed structural decoupling and heterogeneous fusion design in improving predictive performance and interpretability under complex financial conditions.