Applied Neural Network Methods for Option Pricing

This thesis investigates whether neural networks can be used for option pricing while still respecting basic financial rules. It compares four models: two feed-forward neural networks and two physics-informed neural networks using synthetic option prices generated with the Heston model and the COS method. The models are evaluated not only by pricing accuracy, but also by no-arbitrage checks such as price bounds, monotonicity, convexity, maturity consistency, and put-call parity. The results show that the call-only FFNN gives the most accurate prices, while the hard-bounds PINN successfully removes price-bound violations. At the same time, convexity violations remain a challenge, indicating that accurate predictions do not automatically lead to financially consistent option surfaces.