DeepSVM: Learning Stochastic Volatility Models with Physics-Informed Deep Operator Networks

Battle Count: 73

Highly relevant for quantitative trading applications requiring real-time option pricing and calibration. The O(1) inference cost after training removes the computational bottleneck of repeated PDE solves during calibration. Directly applicable to volatility surface fitting, risk management, and hedging. However, the noise in Greeks (Delta, Gamma) in the ATM regime limits immediate production use for precise hedging. The framework's true value lies in models without closed-form solutions (Rough Vol, LSV), which are increasingly used in practice.

Moderate to high complexity. Requires implementing a DeepONet architecture with branch/trunk MLPs, hard-constrained ansatz, PDE residual computation via automatic differentiation (including mixed partial derivatives), Residual-based Adaptive Refinement (RAR), Sobol sequence warping with tanh mapping, two-stage training (Adam + L-BFGS), and careful handling of the Feller condition. The 5D parameter space + 3D domain space (8D total) adds complexity. Libraries like DeepXDE or custom PyTorch/TensorFlow implementations would be needed.

The paper provides detailed architecture specifications (4 hidden layers, GELU activations, residual connections, Softplus correction), training hyperparameters (Adam lr=1e-4, 10000 iterations, L-BFGS 5000 iterations, RAR every 500 steps), collocation set sizes (200,000 points), and parameter ranges. However, no code repository is mentioned, and specific random seeds or exact network widths are not fully detailed. The Feller condition enforcement and Sobol sequence warping are described but implementation details may vary.