SeasonFuser: A Parallel Hybrid Network With Differentiable Seasonal Decomposition for Financial Time Series Forecasting

Abstract

Financial time series forecasting is challenging because market dynamics can change over time and useful temporal patterns may be weak. Existing hybrid models often combine multiple encoders in serial architectures or process decomposed components with a single encoder. We propose SeasonFuser, a parallel hybrid network that combines differentiable Holt–Winters decomposition, component-specific encoders, and adaptive fusion. The decomposition separates the input series into level, trend, and seasonal components and uses trainable seasonal initialization to support sequences with limited historical observations. Each component is processed by a dedicated encoder: a bidirectional long short-term memory (BiLSTM) network for the level component, a Transformer for the trend component, and a temporal convolutional network (TCN) for the seasonal component. The three pathways operate in parallel and are combined using an input-dependent gating mechanism. We evaluate the model on three stock indices—the Shanghai Stock Exchange Composite Index (SSE), the Hang Seng Index (HSI), and the NASDAQ Composite Index—and on Bitcoin (BTC). SeasonFuser achieves the lowest RMSE among the evaluated models on the four datasets and lower RMSE than the persistence baseline on all four datasets. A zero-shot experiment evaluates a Bitcoin-trained model on fifteen other cryptocurrencies without fine-tuning, with $R^{2}\gt 0.90$ for all target assets. The Diebold–Mariano test indicates statistically significant differences between SeasonFuser and the neural baselines, whereas the difference from persistence on Bitcoin is not statistically significant.