Regime-Conditioned Convolutional Autoencoders with Recurrent Networks for Implied Volatility Surface Forecasting

Accurate forecasting of the Implied Volatility Surface (IVS) is a fundamental challenge in financial engineering, derivatives pricing, hedging, and risk management due to the high-dimensionality of option grids and non-linear market dynamics. Traditional models struggle with complex patterns, while standard deep learning architectures suffer from noise. This paper presents a novel two-stage deep learning framework for IVS forecasting that combines (1) a regime-conditioned convolutional autoencoder that learns latent representations of daily 11x34 IVS grids conditioned on the VIX regime, and (2) temporal sequence models (LSTM, GRU, Transformer) that forecast future latent vectors. Using 1,423 daily S&P 500 IVS grids from OptionMetrics (2020–2025) via WRDS, we conduct a comprehensive 226-configuration hyperparameter sweep. Our best model, a GRU with a 60-day lookback window, achieves out-of-sample RMSE of 0.066 for 1-day, 0.083 for 5-day, and 0.100 for 20-day horizons. This represents a 5.7% RMSE improvement over the persistence baseline at the 20-day horizon. Although the Diebold-Mariano test indicates that this difference is not statistically significant at the 5% level (p=0.646) due to limited test set size, the model demonstrates consistent directional outperformance. The findings show that a 60-day context window is optimal, GRUs outperform Transformers at this data scale, and the autoencoder's reconstruction floor sets a bound on forecasting accuracy. The framework provides interpretable representations that correlate cleanly with market indicators, offering a robust tool for risk management.