Universal Diffusion Models for Implied Volatility Surfaces: Learning Shared Dynamics Across Stocks

Battle Count: 0

Highly relevant for derivatives desks and risk management teams. The ability to generate realistic, arbitrage-free IVS scenarios across multiple stocks using a single model (transfer learning) significantly reduces computational overhead for Monte Carlo simulations used in pricing and hedging complex option portfolios.

Moderate to High. Requires implementing a conditional diffusion model with FiLM conditioning, handling high-dimensional IVS data (99-dim vectors), and integrating complex finite-difference arbitrage penalties into the loss function or post-processing steps. Training on pooled cross-sectional data adds complexity regarding normalization and batch construction.

The paper provides detailed architectural specifications (FiLM MLP, 8 residual blocks, hidden width 256), training hyperparameters (AdamW, lr 6e-4, batch size 4096, 1000 epochs), and data splitting protocols (50 in-sample/50 out-of-sample stocks, 2010-2022 train/2023-2024 test). It references specific benchmarks (VolGAN) and defines the grid structure explicitly.