Arbitrage-Aware Multi-Step Forecasting of Implied Volatility Surfaces: Modelling Surface Trajectories Using Latent Diffusion

Battle Count: 66

Highly relevant for options trading desks, volatility traders, and risk managers. The framework generates realistic probabilistic multi-step IVS scenarios essential for dynamic hedging, portfolio risk assessment, and stress testing. The joint surface-return trajectory generation captures the leverage effect (negative return-volatility dependence). However, the model is not a direct trading signal generator; its primary value is in scenario generation and risk quantification rather than alpha generation. The persistence benchmark confrontation is particularly important for practitioners who need to verify that model complexity adds value over naive approaches. The 30-day horizon aligns with typical risk management and hedging rebalancing cycles.

High complexity due to the multi-stage pipeline: (1) operator deep smoothing with a pretrained GNO for surface construction, (2) RAE-inspired autoencoder with residual CNN encoder and coordinate-based neural field decoder with arbitrage penalty terms, (3) conditional latent diffusion with Transformer denoiser, anchored displacement formulation, v-parameterization, and mean-scaling gate. Requires expertise in diffusion models, financial mathematics (no-arbitrage conditions), neural architecture design, and time-series modeling. Training involves multiple stages with careful hyperparameter tuning. The 500-step DDPM sampling for 1000 trajectories per forecast origin is computationally intensive. However, code is provided on GitHub.

Code is publicly available on GitHub. Data construction pipeline is provided. However, the underlying data (OptionMetrics IvyDB via WRDS) requires a paid institutional license. The paper provides a transparent, reproducible protocol for constructing IVS forecasting datasets. Model architecture, hyperparameters, training details, and evaluation protocol are thoroughly documented. The authors explicitly frame their contribution as establishing a reproducible benchmark.