Learning Option Greeks using Differential Machine Learning

Our project studies how differential machine learning (DML) can improve option pricing and Greek estimation under the Heston model, enabling faster and more stable estimation of both value and risk, essential for quantitative finance. We generate training data using Monte Carlo simulation and compute pathwise sensitivities (Delta, Vega, Rho) alongside prices. Neural networks are then trained with both value and differential labels, using a weighted loss to balance accuracy across outputs. We benchmark results against Fourier-based analytical solutions. Compared to standard ML, DML consistently reduces pricing and sensitivity errors. Delta supervision improves both Delta and Gamma, while adding Vega and Rho further enhances performance. The combined model using all three sensitivities achieves the best overall accuracy.