Diffusion models for dynamic volatility surface generation and data-driven hedging

Battle Count: 80

Highly relevant to quantitative trading. The paper directly addresses derivative hedging, scenario generation for risk management, and tail-risk reduction. The data-driven hedging framework with diffusion-generated scenarios is immediately applicable to options desks. The demonstrated stability during COVID-19 and superior tail-risk metrics (VaR) are critical for practical trading. The framework connects generative modeling to actionable hedging decisions with transaction costs.

High complexity: requires implementing DDPM with Transformer score networks, sequential conditioning on evolving history, LoRA fine-tuning with arbitrage penalties, and LASSO-based hedging optimization. The 200-step denoising sampler and sequential multi-period generation add computational overhead. However, the code is publicly available and the architecture follows established patterns from Cao et al. (2026) and Han et al. (2025).

Code is publicly available on GitHub. Data source (OptionMetrics SPX daily quotes) is specified with exact date ranges. Model architecture (Transformer), training hyperparameters (AdamW, batch size 256, 1000 epochs, cosine schedule, 200 noise steps), and fine-tuning details (LoRA rank-8, KL penalty 0.1, 200 epochs) are fully specified. VolGAN benchmark independently implemented. However, OptionMetrics data requires a subscription.