Application of machine learning in option pricing

This thesis investigates the application of machine learning methods— particularly feedforward neural networks—for pricing exotic options, traditionally evaluated via Monte Carlo simulation. A modified implementation of the Volatility Feature Approach (VFA) is proposed, wherein Multilayer Perceptrons (MLPs) are trained to approximate option prices directly from contract parameters and a discretized volatility smile surface. Synthetic datasets covering Asian, barrier, and lookback options are generated using the Bates stochastic volatility model with jumps, with all model parameters randomized within realistic bounds. The resulting neural networks, trained on 3000 samples per option type and evaluated against Monte Carlo ground truth, achieve substantial improvements in computational efficiency while maintaining high predictive accuracy. The methodology supports rapid, arbitrage-free approximations of exotic option prices and offers a practical alternative for real-time pricing scenarios where traditional calibration methods are computationally prohibitive.