Differential Machine Learning for 0DTE Options with Stochastic Volatility and Jumps

Battle Count: 81

Highly relevant to quantitative trading, particularly for 0DTE options which now account for a large share of trading volume. The paper directly addresses fast pricing and Greeks computation (critical for intraday rebalancing), delta hedging performance, jump-risk management, and model calibration. The 90x speedup over Fourier-based benchmarks makes it practical for real-time trading systems. The focus on ultra-short maturities and jump dynamics aligns with current market structure challenges in 0DTE options trading.

Moderate to high complexity. Requires: (1) implementing a Fourier-transform pricer for the Bates model as benchmark, (2) designing a twin-network architecture with automatic differentiation for Greeks, (3) implementing a three-stage training schedule with different loss configurations per stage, (4) constructing numerical quadrature proxies for the jump operator, (5) handling numerical instabilities in ultra-short maturity regimes (robust losses, filtering, no-arbitrage constraints), (6) calibration via L-BFGS-B with the trained surrogate. The neural network architecture itself is straightforward (two FC networks), but the training procedure, loss design, and numerical stability considerations add significant complexity.

The paper provides detailed architecture specifications (network width, depth, activation functions, parameter counts), complete loss function formulations with all weights, training hyperparameters (learning rate, batch size, epochs per stage), sampling distributions for training data, and benchmark implementation details (Fourier pricer with 1024-point quadrature). However, no code repository is explicitly linked, and some implementation details (e.g., exact random seeds for all experiments, specific optimizer schedules beyond Adam) may require additional clarification.