The arbitrage control reaches one representative surface. The full set of surfaces drawn by the model passes around it.
Hao and Ji propose a two-stage forecaster for the implied volatility surface of CSI 300 index options. Stage I uses conditional diffusion to predict the next fixed-grid surface from recent history. Stage II applies the Surface Aware Attention Module, or SAAM. This refinement operator maps the forecast onto any requested (moneyness, maturity) coordinate. Soft penalties target calendar and butterfly violations, while a boundary-curvature wing term handles the edge of the domain. The intended split is clear: the generative stage models market stochasticity, and the attention stage imposes shape restrictions.
The data preparation deserves attention. A Nadaraya-Watson kernel smooths the scattered quotes, with bandwidths selected separately for each date through least-squares cross validation. The resulting field is sampled on a fixed 15 by 17 grid, giving 255 nodes across log-forward moneyness and maturity. Its m range is [log 0.6, log 1.5], and tau is at least 10/365. RiceQuant supplies 206,561 pooled daily implied volatility observations from June 10 2020 to September 30 2024.
For the daily protocol, 22 trading days of history feed a next-day surface forecast under a chronological 80/20 split. The minute protocol uses 60 minutes of history to predict the next five one-minute states, divided 70/10/20. Stage I is judged against the kernel-smoothed grid instead of the raw quotes. Stage II is evaluated at observed contract locations.
We could not run this. The paper studies CSI 300 index options, with its strongest findings at minute frequency. Our end-of-day US options data contains neither intraday chains nor CSI 300 coverage. Using SPY or QQQ would replace both the market and the protocol.
Nothing below comes from a run of ours.
A next-period surface can serve as the mark for a relative-value vol book comparing model values with live quotes. Yet the paper evaluates forecasting accuracy, fit to market observations, and finite-grid residual diagnostics. It does not report hedging error or P&L.
The ensemble stops before refinement
At each test origin, the diffusion model produces B = 100 scenarios. Distributional scoring uses the complete ensemble, yielding daily CRPS of 0.0350 ± 0.0018 and minute-level CRPS of 0.0078 ± 0.0005. CRPS is available only for the diffusion ensemble, with N/A against LSTM-Direct. SAAM receives the pointwise median across the 100 draws. Because the model constructs that median from the ensemble, it is a new surface rather than one of the scenarios actually sampled.
Calendar and butterfly penalties apply to this median alone. The 100 sampled surfaces never enter the refinement stage, even though those scenarios would form the natural input for a distribution of future risk surfaces and scenario P&L. The paper also gives no per-scenario residual statistics for the ensemble. Figure 7 offers a visual view of violation regions in unrefined diffusion output, limited to three representative daily samples.
The authors describe this separation directly. The Introduction says the framework 'separates the generation of future risk surface scenarios from the construction of a refined representative surface used for option valuation, risk measurement, and static no arbitrage assessment.' A single refined marking surface is internally coherent. Still, the opening of Stage II acknowledges what remains on the generative side: the diffusion stage 'generates predictive surface scenarios that need not satisfy the static no-arbitrage restrictions in Proposition 1'.
The abstract reaches further when it says SAAM 'controls static no-arbitrage residual violations on the final surface', although every generated scenario remains unrefined. Its own qualification follows in the next clause: 'reduces measured static no-arbitrage residual violations'.
Most of the forecasting gain arrives at high frequency
On the same fixed grid, the diffusion median improves daily RMSE over LSTM-Direct from 0.0288 ± 0.0007 to 0.0263 ± 0.0006, about 8.7%. The minute-level change is much larger, from 0.0178 ± 0.0001 to 0.0114 ± 0.0004, about 36%. MAE falls from 0.0110 ± 0.0001 to 0.0017 ± 0.0001. That final gap, a factor of six, carries the sharpest result.
This is mainly a high-frequency finding, and it runs against the common prior that noisier data should favour the simpler model. Independent replication remains difficult. The paper says the authors cannot publicly redistribute the RiceQuant data, while code is available on request. Reported dispersions are standard deviations across five independent runs. They are not standard errors from a forecast comparison. We did not find a Diebold-Mariano test or another formal test of the RMSE differences.
How much does a butterfly zero establish?
Every Stage II method in the daily regime records L-cal = L-but = 0.0000. The result is identical for Fusion SAAM, Separation SAAM, MLP-Direct and MLP-VAE. On this daily arbitrage table, the attention refiner and the pointwise MLP cannot be separated. Fit provides the distinction. Fusion SAAM reaches RMSE 0.0190 ± 0.0009 and MAPE 8.47% ± 0.40%, compared with 0.0290 ± 0.0026 and 10.16% ± 1.02% for MLP-Direct.
Those zeros carry the authors' stated limits. The residual calculation uses the same 60 by 60 synthetic collocation grid that supplies the training penalty. As the paper says, the objective 'is not an exact projection onto the continuous arbitrage free set'. In the authors' interpretation, small residuals provide numerical evidence that 'the refined surface satisfies the inspected finite domain diagnostics'. They do not prove global static arbitrage freedom. A finite domain cannot verify the tail condition, so the method substitutes a boundary curvature term.
The raw minute-level data has a mean butterfly violation rate of 0.1099, with median 0.1070 and max 0.2841. Calendar violations average 0.0006. Convexity is therefore the live diagnostic here, and the framework inspects it on the grid used to fit the penalty.
Fit carries a price
The minute-level arbitrage ordering reverses. Both MLP refiners achieve exactly 0.0000 butterfly residual. Fusion SAAM records 0.0002 ± 0.0001, while Separation SAAM records 0.0006 ± 0.0002. The authors report the difference and address it directly.
The MLP zeros come with RMSE of 0.0830 ± 0.0008 and 0.1015 ± 0.0007. Fusion SAAM is at 0.0215 ± 0.0009. Their defence is reasonable: grid admissibility has little value when the surface no longer agrees with observed quotes. MLP-Direct finishes 8.3 vol points of RMSE away from the minute-level quotes, and MLP-VAE finishes 10.2 away. Neither surface is usable.
Two disclosed qualifications remain. Each pipeline uses its own grid-searched penalty weight. At the minute level, lambda = 0.01 for SAAM and 1.00 for the MLPs. The authors describe the sensitivity analysis as representative rather than exhaustive.
Inputs also differ across refiners. SAAM works on the diffusion median, while the MLP baselines receive recurrent forecasts. The paper states: 'the comparison is conducted at the pipeline level under a common evaluation protocol, rather than as an input controlled ablation of refinement operators'. A common protocol gives the comparison some standing. Even so, the 0.0215 versus 0.0830 result combines improvements from Stage I and Stage II.
Evidence that would change the verdict
I would want calendar and butterfly residuals for all 100 raw diffusion draws, with the median reported alongside them. A delta-hedged P&L or vega-weighted mispricing test on the refined surface would also connect the statistical exercise to tradable use.
For now, the framework is a well-built constructor of a representative surface, with its arbitrage machinery measured on the same grid where it was trained. We reached a similar conclusion when examining a different diffusion model whose advertised generative feature contributed less than the surrounding conditioning (our note on tempered-stable diffusion).
The attention refiner earns its place through fit: 0.0215 against 0.0830 at the minute level, and 0.0190 against 0.0290 daily. The title promises Arbitrage-Aware Refinement. In the daily panel, both MLPs share its column of zeros. At the minute level, SAAM leaves 0.0002 and 0.0006 while the MLPs report 0.0000.