Stochastic Volatility Modelling with LSTM Networks: A Hybrid Approach for S&P 500 Index Volatility Forecasting

Battle Count: 119

Directly addresses volatility forecasting for the S&P 500, a core input for risk management, options pricing, and trading strategies. Includes practical VIX futures trading simulation with transaction costs. The hybrid approach bridges econometric and ML methods relevant to quantitative finance. However, the trading strategy results show underperformance vs. naive benchmarks with conventional loss functions, and the MADL modification only partially addresses this. The paper provides actionable insights for volatility-based signal generation.

Requires implementation of SV model with MCMC (Metropolis-Hastings + ASIS) via stochvol package, LSTM architecture with extensive hyperparameter tuning (25 random search trials × 3 repetitions × 11 windows), rolling window data management, multiple scaling approaches, and statistical testing (Wilcoxon, Diebold-Mariano). The two-stage pipeline (SV → LSTM) adds integration complexity. Approximately 24 hours of tuning time on CPU. Moderate-to-high complexity for full reproduction.

All code and supplementary materials are openly accessible via GitHub (https://github.com/aperekhodko/sv_lstm_hybrid_model). Data sourced from Yahoo Finance (publicly available). The stochvol R package and TensorFlow/Keras are standard tools. However, the extensive hyperparameter tuning across 11 rolling windows and MCMC sampling introduce some stochastic variability. No GPU was used, and results may vary slightly across runs.