Universal Diffusion Models for Implied Volatility Surfaces: Learning Shared Dynamics Across Stocks¶

Bibliographic record. Follow the original-source link for the publication.

Field	Value
Primary domain	Volatility
Other domains	Portfolio Construction Risk Transfer
Methods	Research Methods
Facets	—
Authors	Mingzhi Yang, Sheng Wang, Chao Zhang, Ruikun Li
Published	2026-09-19
Source	arXiv Quantitative Finance History
Identifiers	arxiv:2609.22893
URL	Open original source

Editorial synthesis¶

Why it matters¶

This preprint tests whether a pooled conditional diffusion model can transfer joint one-day return and IVS-increment generation to stocks excluded from training. Its options-research value is a checkable cross-stock, time-OOS and regularization comparison, not evidence of alpha or trading fitness. (full_text:S1, full_text:S15, full_text:S17, full_text:S19, full_text:S23, full_text:S78, full_text:S88, full_text:S147)

Main author claims¶

The authors propose a shared FiLM denoiser for next-day log-return and 99-dimensional log-IVS increments, comparing MSE, smoothness, arbitrage and hybrid losses against VolGAN retrained on the same stock pool. (full_text:S38, full_text:S39, full_text:S40, full_text:S42, full_text:S65, full_text:S80, full_text:S81)
The authors report lower discrete static-arbitrage penalties for diffusion variants than VolGAN, with MSE lowest. For held-out stocks, reweighted mean penalties are 0.42×10^-3 for MSE and 2.00×10^-3 for VolGAN; arbitrage-variant 95% return-interval coverage is 94.60%, versus 76.33% for VolGAN. These are paper-reported values. (full_text:S89, full_text:S92, full_text:S96, full_text:S118)
The authors use PCA variance allocation, AAPL loading surfaces and similar results across training and held-out stocks to support shared dynamics and transfer without retraining. This is their interpretation of one-step generation diagnostics. (full_text:S124, full_text:S125, full_text:S131, full_text:S134, full_text:S144, full_text:S150, full_text:S152, full_text:S194, full_text:S202, full_text:S209)

Data, method, or discussion scope¶

The experiment uses 100 US stocks with a random 50/50 stock split: training uses 2010-2022 data from the first 50, and both groups are tested in 2023-2024, so even 'in-sample stocks' are time-OOS. OTM quotes undergo stock-specific Vega-weighted Nadaraya-Watson smoothing onto an 11×9 grid. Conditions contain two lagged returns, trailing 21-day realized volatility and previous-day log-IVS; targets are next-day return and surface increment. The authors state that z-scores use only the training pool, with one seed and 1000 scenarios per condition. Evaluation covers grid penalties before/after reweighting, equally weighted stock-average 95% return coverage and PCA along realized test-condition sequences, not freely recursive multi-day paths or trading tests. (full_text:S34, full_text:S38, full_text:S39, full_text:S40, full_text:S70, full_text:S77, full_text:S78, full_text:S80, full_text:S82, full_text:S83, full_text:S85, full_text:S88, full_text:S89, full_text:S94, full_text:S118, full_text:S144, full_text:S147)

Main limitations¶

Author-stated limits are one-day generation, a fixed grid, static rather than dynamic no-arbitrage and one nominal coverage level; regularizer interactions need multi-seed study. Reader-inferred untested boundaries: Table 2 penalties remain positive, so reduced grid violations are not a global arbitrage-free guarantee. One seed, one stock split and descriptive summaries do not establish robust universal superiority. Training-only z-scores do not resolve the undisclosed fitting period for stock-specific smoothing bandwidths. Figure 4's mean±1.96 SD measures generated-sequence dispersion, not a sampling CI for performance differences. One PCA-superiority statement conflicts with Table 4: hybrid, not the arbitrage variant, is closest to held-out PC1 by the stated absolute-distance criterion. Mean 95% coverage establishes neither sharpness, conditional calibration nor full tail/joint-distribution fidelity; option hedging, net trading returns and costs are not evaluated. (full_text:S34, full_text:S38, full_text:S39, full_text:S40, full_text:S82, full_text:S85, full_text:S88, full_text:S94, full_text:S96, full_text:S100, full_text:S105, full_text:S118, full_text:S125, full_text:S128, full_text:S131, full_text:S132, full_text:S144, full_text:S147, full_text:S149, full_text:S150, full_text:S212, full_text:S215, full_text:S216)

Relationships¶

None recorded.