A Generative Adaptive Decomposition Hierarchical Transformer for Stock Price Forecasting

Forecasting stock prices remains a critical yet challenging task due to the non-stationary, noisy, and highly volatile nature of financial time series. This study proposes a novel hybrid forecasting framework, called the Generative Adaptive Decomposition Hierarchical Transformer (GADHT), designed to improve predictive accuracy and interpretability. The framework first employs Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) to decompose price signals into intrinsic mode functions (IMFs), effectively separating meaningful patterns from noise. Next, a generative pretraining strategy based on IMF masking and reconstruction is introduced to learn robust spectral-temporal representations without requiring labeled data. Finally, a hierarchical transformer equipped with energy-weighted attention dynamically focuses on the most informative spectral and temporal features to enhance prediction. Comprehensive evaluations on multiple stocks with varying volatility profiles demonstrate GADHT’s robust overall performance. Specifically, for TSLA, it substantially outperforms state-of-the-art baselines, achieving superior predictive accuracy and model reliability. Furthermore, interpretability analysis through spectral and temporal attention maps confirms the model’s ability to align with financial domain knowledge, providing transparent insights into decision-making processes. These results highlight GADHT's potential as a practical and reliable tool for real-world stock forecasting and dynamic portfolio management.