Improving option pricing with hybrid deep learning and financial models

We propose a novel hybrid framework that integrates stochastic volatility modeling with deep learning. The main contribution is a dual neural network architecture combining convolutional and LSTM networks (CNN-LSTM). This architecture learns and corrects pricing errors from the Heston model, and can optionally be enriched with GARCH-based time-varying volatility forecasts. The methodology involves three stages. First, the Heston model is calibrated using a genetic algorithm. Second, a GARCH model captures volatility dynamics over time. Third, a hybrid parametric model is built by integrating the CNN-LSTM network with the Heston framework (with or without GARCH predicted volatility). Results show that the hybrid Heston – CNN-LSTM model significantly lowers pricing errors and corrects systematic bias. Adding GARCH further enhances performance, especially for deep out-of-the-money options and under turbulent market conditions.