Decoupled Probabilistic Forecasting and Arbitrage-Aware Refinement of Implied Volatility Surfaces

Battle Count: 78

Highly relevant to quantitative trading, particularly in options and volatility trading. The framework directly addresses IVS forecasting for option valuation, hedging, and risk management. The probabilistic scenario generation supports scenario-based risk analysis and stress testing. The static no-arbitrage refinement ensures pricing surfaces are usable for derivative valuation without arbitrage violations. The minute-level forecasting capability is directly applicable to high-frequency options trading and intraday risk monitoring. The CSI 300 focus is relevant to Chinese equity derivatives markets. The decoupled design separates scenario generation from surface construction, mirroring practical trading desk workflows.

High complexity due to the two-stage pipeline: (1) Conditional diffusion model with Transformer-based score network requires careful training of forward/reverse processes, noise scheduling, and repeated sampling (B=100 scenarios); (2) SAAM requires coordinate-based attention with learnable encoders, Softplus output, and training on scattered observations plus synthetic collocation grids; (3) Nadaraya-Watson kernel surface reconstruction with cross-validated bandwidths; (4) Structured spatial discretization onto a 15×17 grid; (5) Static no-arbitrage residual computation (calendar, butterfly, wing) on dense synthetic grids; (6) Penalty weight tuning per regime. Requires GPU training (RTX 4090), PyTorch, and careful handling of irregular contract coverage across dates.

Code is available from the corresponding author upon reasonable request but no public repository is provided. Data comes from RiceQuant, a commercial subscription service, limiting independent verification. Model architecture details (64 channels, 8 heads, 4 residual blocks for diffusion; d=64, 3-4 heads for SAAM) and training settings (Adam optimizer, RTX 4090 GPU) are specified. Penalty weight sensitivity is reported in Appendix C. Five independent runs with mean±std are reported. However, exact hyperparameter search grids, random seeds, and full training code are not publicly available.