Latent Flow Matching for Arbitrage-Aware Implied Volatility Surface Generation

Battle Count: 72

Highly relevant for quantitative trading in derivatives markets. The framework enables realistic scenario generation for option portfolio risk management, stress testing, and volatility surface calibration. The arbitrage-aware generation ensures synthetic surfaces are financially valid for pricing and hedging. However, it is unconditional (not a forecasting tool) and focuses on surface generation rather than direct trading signal extraction. Most useful for risk managers and derivatives desks rather than high-frequency trading.

Two-stage pipeline requiring careful coordination: (1) VAE training with multiple arbitrage penalty terms requiring differentiable implementations of Gatheral-Jacquier criterion and discrete derivative approximations; (2) Flow matching with ODE integration (100 Euler steps). Architecture is moderate (residual MLPs, U-Net-style), but the arbitrage penalty engineering (calendar, butterfly, call-spread) and SVI interpolation preprocessing add significant complexity. Hyperparameter tuning across multiple loss weights (β, λ_cal, λ_but, λ_call) and latent dimension requires careful experimentation.

GitHub repository provided (https://github.com/DusBaja/ivs-generative-benchmark). Detailed hyperparameters in Table 1. Data source specified (OptionsDX SPX options, Jan 2020 - Dec 2023). Architecture details, training procedures, and evaluation metrics fully described. Anonymous repository noted, which may limit long-term accessibility. Five independent runs reported with mean±std.