Explainable Deep Learning for Financial Volatility Forecasting: An LSTM-Attention-SHAP Framework with Comprehensive Validation

Financial market volatility forecasting is critical for risk management, portfolio optimization, and regulatory compliance. Traditional models such as GARCH and HAR-RV fail to capture non-linear dependencies and extreme tail events, while deep learning methods lack the transparency required under Basel III/IV and SR 11-7 frameworks. This paper presents a hybrid framework combining Long Short-Term Memory (LSTM) networks with a Bahdanau attention mechanism and SHapley Additive Explanations (SHAP) to deliver high predictive accuracy alongside regulator-verifiable interpretability. Using five asset classes from January 2018 to December 2024, the model achieves a 28.6% RMSE reduction over GARCH(1,1), an R² of 0.72, and a 99% VaR violation rate of 1.05%, passing both Kupiec and Christoffersen backtests. All results are statistically significant at p < 0.001 via Diebold-Mariano tests. SHAP analysis identifies the Geopolitical Risk index (mean |SHAP| = 0.45) as the dominant volatility driver, followed by VIX (0.38) and realized volatility (0.32). The attention mechanism complements this by identifying which historical periods most influence current predictions, providing dual-layer interpretability that is both actionable for risk managers and compliant with regulatory audit requirements. This framework demonstrates that predictive accuracy and transparency can be achieved simultaneously in production financial systems.