Arbitrage-Aware Probabilistic Multi-Step Forecasting of Implied Volatility Surfaces: Joint Generation of Surface-Return Trajectories Using Latent Diffusion
The research proposes a conditional latent diffusion framework to jointly generate 30-day trajectories of implied volatility surfaces and underlying asset returns. It uses an autoencoder for surface representation and a diffusion model for evolution, ensuring the generated scenarios are realistic, preserve economic rules (no static arbitrage), and outperform simple persistence models over longer time horizons.