Predicting the implied volatility surface via deep learning

Abstract

This thesis explores deep learning techniques for predicting the implied volatility surface (IVS), a critical component in options pricing and risk management. By applying ConvLSTM and Self-Attention mechanisms, the study evaluates their ability to capture spatial and temporal patterns across strikes and maturities. Results show that grid-based ConvLSTM excels in short-term forecasting, while Self-Attention enhances long-term accuracy by capturing global dependencies. The models were retrained and evaluated under volatile regimes, including the COVID-19 crash, testing their robustness in extreme market conditions. The findings contribute to improved IV surface predictions, benefiting strategies like volatility arbitrage and dynamic hedging.