The result a trading desk should care about is 94.53% coverage on names the model never saw. Using unweighted scenarios, the arbitrage-penalized diffusion places realized daily returns inside its nominal 95% interval for the 50 out-of-sample stocks. Coverage on the 50 training stocks is 94.38%. The 0.15 points separating in-sample from zero-shot coverage is the result worth having here. It survives scenario reweighting, at 94.60% out-of-sample against 94.50% in-sample.
A clean replication still runs into four missing details: the window used to tune each stock's smoothing bandwidth, the rate and dividend convention behind the arbitrage penalties, the scenario reweighting function, and the ticker list.
We ran nothing against this paper. Our option Greeks and implied vols begin around 2020, leaving the 2010 to 2022 training split beyond our reach.
Yang, Wang, Zhang and Li train a conditional DDPM to draw tomorrow's return together with tomorrow's volatility-surface move. Its target is a 100-dimensional vector, formed from the next day's log-return and the 99-dimensional increment of the log implied-volatility surface. The surface uses a fixed grid with 11 moneyness nodes, spanning 0.6 to 1.4, and 9 maturities from one day to one year.
The conditioning state has 102 dimensions. Three scalars join yesterday's full log-surface: two lagged daily returns and a 21-day annualized realized volatility. A FiLM-modulated MLP serves as the denoiser, with hidden width 256, eight residual blocks and 1000 diffusion steps under a cosine schedule. The pooled training sample contains 50 stocks from 2010 to 2022. Testing covers 2023 to 2024, with 1000 scenarios drawn for each conditioning date. Another 50 stocks, selected at random, never enter training.
This is scenario infrastructure. An overnight option-book revaluation needs joint (return, surface) draws, yet per-name fits leave thin-history option names poorly served. Pooling across the cross-section gives the model a chance to cover those names without retraining.
The paper compares four objectives: MSE only, MSE plus surface smoothness, MSE plus static-arbitrage penalties, and both. VolGAN is the benchmark, retrained by the authors on the same pool with recalibrated smoothness penalties. Gradient matching determines the penalty weights. For the hybrid model, calendar receives 0.2739, vertical spread 0.0671, butterfly 0.0166, and the two smoothness terms 0.0086 and 0.0011.
Replication breaks in the preprocessing
The network and optimizer are documented almost exhaustively, a welcome choice. Training uses AdamW at 6e-4, batch 4096, 1000 epochs and EMA 0.999. Z-score parameters are estimated separately for each dimension using only the in-sample training set. The smoother is also identified: a Vega-weighted Nadaraya-Watson fit applied to out-of-the-money quotes. Equations (1) to (4) contain the penalty formulas.
Two omissions directly affect the level of the headline metric. Four gaps matter for anyone trying to build the model:
- Equations (2) to (4) express the static-arbitrage penalties entirely through call prices c(m, tau). The paper uses Black-Scholes inversion to construct the surfaces, yet gives no rate, dividend or forward convention for converting a generated surface back into prices. Every reported level in the arbitrage table depends on that choice. The authors already say the penalties should be normalized by stock-specific baselines, and implementation comparisons face the same problem.
- Stock-specific bandwidths for the Nadaraya-Watson smoother are described as "tuned to minimize static arbitrage." We did not find the tuning window. A full-sample window would let test-period information enter the preprocessing behind the headline metric before model training begins.
- Cont and Vuletić are cited for scenario reweighting, while the function itself is omitted. Beta is 50 and effective sample size remains above 995 of 1000, leaving an implementer to infer the functional form.
- We did not find either the ticker list or a named data vendor. The 100-name universe therefore cannot be reproduced exactly.
Both the smoothness penalty in equations (15) and (16) and the arbitrage penalties are computed on the reconstructed clean target from equation (13). Near k = 1000, that reconstruction approaches noise. We did not find any statement limiting penalty terms to low-noise steps, so an implementation must choose how to handle them.
How much does the 0.15-point gap establish?
Only part of the shared-dynamics case rests on that 0.15-point gap. The authors acknowledge the limitation. Out-of-sample stocks have an observed arbitrage baseline of 0.56e-3, compared with 1.40e-3 in-sample, which prevents penalty levels from serving as a clean transfer measure. Their stated conclusion is that performance "does not collapse under" zero-shot application and that the advantage is "not driven by overfitting."
The ranking evidence holds. Every diffusion variant beats VolGAN for both stock groups and under both reweighting panels. Absolute penalty levels carry less weight because the groups begin from different data baselines. Coverage and PCA therefore bear most of the transfer claim.
PCA alignment is convincing. For the arbitrage variant, PC1 explains 82.45% in-sample and 83.14% out-of-sample, versus 81.62% and 80.62% in the observed data. The loading shapes recover level, skew and curvature for AAPL, the only name plotted in the paper.
The conditioning state makes transfer easier than the headline may suggest. It includes yesterday's realized volatility and the full prior surface, including the ATM level. Mapping an observed vol level into an interval width should transfer across names with comparable option chains almost mechanically. Comparability remains our assumption because the paper describes only "100 US equities" split randomly 50/50, with no breakdown by size, sector or liquidity.
Even so, the coverage result is real. With the same inputs, VolGAN covers 75.06% of realized returns over the full universe against a 95% nominal interval, roughly 20 points short. One coverage level remains a thin calibration test. The authors themselves place full distributional calibration in future work.
The penalty made measured arbitrage worse
MSE only wins the arbitrage table. Its reweighted full-universe mean penalty is 0.37e-3, compared with 0.54e-3 for the arbitrage-regularized variant and 1.90e-3 for VolGAN.
Adding the economically motivated penalty left more arbitrage than leaving it out.
The authors suggest three possible causes: smoothed data may already lie near an arbitrage-free manifold, gradients may interfere, and clean-target penalties during training may fail to match violations after sampling. They call for component-wise, multi-seed work. Every result uses one random seed, leaving the 0.33e-3 versus 0.49e-3 in-sample gap without an error bar.
The full row gives a less flattering comparison with the market data. On the reweighted full-universe mean, all diffusion variants show fewer measured violations than the smoothed historical surfaces. The data records 0.97e-3, against 0.37, 0.55, 0.54 and 0.55 for the diffusion variants. Without reweighting, the smoothness variant reaches 1.10e-3 and exceeds the data.
Only the mean makes the generator look cleaner than the market, at 0.37e-3 against 0.97e-3. The observed data retains the lowest median, 0.54e-4 compared with 0.88e-4 for MSE-only and 6.86e-4 for VolGAN. We read this split as a right-tail effect; the authors do not make that inference. The paper supplies no per-name decomposition. In-sample, the observed mean is 1.40e-3, alongside a 0.46e-4 median and a 1.90e-3 95th percentile.
Butterfly violations account for most of the total under every model. VolGAN records 18.40e-4 against 3.72e-4 for MSE-only. The same constraint bound in our note on WSVI's shape family.
One result is especially odd. Reweighting with beta = 50 preserves more than 995 of 1000 effective samples, while reducing the full-universe MSE-only mean from 0.81e-3 to 0.37e-3. Such a near-uniform weight vector can halve the mean only when the within-date penalty distribution has a violent right tail.
Deployment evidence remains absent
We can run a daily model for an end-of-day risk application. We cannot supply the 2010 to 2022 training history. Any implementation on our side would have a shorter history, a universe limited to liquid optionable names, and a quote filter and smoother rebuilt from end-of-day fields.
The abstract frames the model around pricing, hedging and risk-managing option portfolios. We did not find a hedging test. We also did not find a P&L evaluation, and transaction costs appear nowhere.
The authors identify five areas for future work: longer horizons, dynamic arbitrage constraints, full distributional calibration, normalized penalties, and multi-seed analysis. Hedging is not among them. The nearest cited work is Han et al. on SPX, titled "Diffusion Models for Dynamic Volatility Surface Generation and Data-Driven Hedging". The paper makes no comparison with it, and the single-index setting would make one awkward. Despite the risk-management framing, no regime or stress-period breakdown is reported for the 2023 to 2024 test window.
A multi-day path test could change our view of the deployment case if it measured hedge errors net of costs on the zero-shot names. Until such evidence appears, this remains a generator with a fully specified network and optimizer, incomplete preprocessing and price conversion, and one result worth carrying forward: 94.53% coverage on fifty never-trained names against a 95% nominal interval.