Differential ML with a Difference

Battle Count: 107

Highly relevant for derivatives desks and risk management teams that need fast approximations to complex pricing models. The paper directly addresses the accuracy of Greeks (delta, gamma) which are critical for hedging and risk management. However, it is more focused on pricing/risk infrastructure than on trading strategy development or alpha generation. The neural network approximation approach is particularly relevant for real-time pricing of exotic derivatives.

Moderate complexity. The core DML framework builds on Huge and Savine (2020) with code available. The main addition is computing LRM labels (score functions) which requires knowledge of the transition density. For standard models (Black-Scholes, Bachelier), this is straightforward. For discretized SDEs (Heston), the Euler scheme score is more involved. The hybrid gamma method adds another layer of complexity. Network architecture and training are standard.

Code is publicly available on GitHub. Detailed hyperparameters are provided (network architecture: 4 hidden layers of 20 units, softplus activations, Adam optimizer, cosine-decay learning rate between 1e-3 and 1e-6, batch sizes 256-512). Training parameters (m, k, lambda) are specified. Numerical examples use standard Black-Scholes and Bachelier models with known closed-form solutions for validation.