Research on Deep Learning Financial Volatility Prediction Method Based on Signal Decomposition and Data Augmentation

Abstract Accurate forecasting of financial volatility is critical for risk management and investment decision-making. This study proposes a novel three-stage hybrid model (C-WG-BL) integrating Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), Wasserstein Generative Adversarial Network with Gradient Penalty (WGAN-GP), and Bidirectional Long Short-Term Memory (BiLSTM) networks. First, CEEMDAN decomposes the original volatility series into multi-scale Intrinsic Mode Functions (IMFs). Next, WGAN-GP is applied to high-frequency IMFs containing major noise and microstructural fluctuations, generating high-quality synthetic data to expand the training set and enhance the model’s ability to capture complex patterns. Finally, BiLSTM forecasts all IMFs (augmented high-frequency and original low-frequency), and the results are integrated to reconstruct the final prediction. Empirical analysis using 1-minute high-frequency data from the SSE 50, CSI 300, CSI 500, and SSE Composite Index shows that C-WG-BL significantly outperforms mainstream deep learning and ablation models. For the SSE 50, it improves R2 by 5.38% and reduces MAPE from 19.197% to 7.432% (–11.76 percentage points) compared with the best baseline (C-BiLSTM). The model also maintains high accuracy under extreme conditions such as the early COVID-19 outbreak, and Diebold–Mariano tests confirm statistically significant error reductions. This study offers an efficient, robust, and generalizable solution for high-frequency financial volatility prediction.