Volatility-Surface Encoders for Regime-Robust Realized Volatility Forecasting Models

Abstract

The accurate forecasting of realized volatility is a fundamental task in quantitative finance, with extensive applications in derivative pricing, portfolio optimization, and risk management. Traditional econometric models and standard machine learning approaches frequently fail to maintain predictive accuracy during abrupt market regime shifts, such as transitions from calm periods to financial crises. This paper introduces a novel deep learning framework that incorporates a Volatility-Surface Encoder to achieve regime-robust forecasting of realized volatility. By treating the entire implied volatility surface as a two-dimensional spatial input, the proposed model employs convolutional autoencoder architectures to extract high-dimensional, forward-looking latent representations of market expectations. These spatial representations are subsequently integrated with historical high-frequency realized volatility measurements using temporal neural networks. The comprehensive empirical analysis utilizes a large-scale dataset of equity index options and high-frequency asset returns, evaluating performance across distinct market regimes. The findings indicate that the inclusion of spatially encoded implied volatility features significantly reduces out-of-sample forecasting errors, particularly during periods characterized by structural breaks and elevated market stress. The proposed methodology provides a robust alternative to purely backward-looking autoregressive models, demonstrating the profound informational value embedded in the structural dynamics of option pricing surfaces.