Brooks and Mitchell's repair step makes reconstructed option grids pass discrete static-arbitrage checks, at a cost to volatility accuracy. Their ablation gives most of the credit for denoising to the learned prior, just as the authors do. The repair supplies feasibility. This is useful valuation plumbing for synthetic data, with a considerable distance still to travel before a desk can trade it.

Inside the reconstruction

The model fills gaps in sparse, noisy option grids. Each test surface contains a 12 by 16 array of normalized European call prices. Maturities span 0.15 to 1.50 years, strikes run from 0.78 to 1.22, and spot is set to 1. Ground truth comes from mixing three zero-rate Black-Scholes call surfaces. Weights follow Dirichlet(2,2,2), and volatilities are drawn uniformly from [0.09, 0.55]; this construction makes every true surface arbitrage-free. A grid point is observed with probability 0.30, with Gaussian quote noise of standard deviation 0.003.

Reconstruction proceeds in three stages, starting with linear interpolation to fill the grid. A four-layer residual CNN then predicts a correction, using 3x3 convolutions and 24 channels. During training, it receives extra Gaussian perturbations of random size up to 0.016. An input channel tells it the perturbation size; at inference, that channel is zero. This training device accounts for the entire "diffusion-inspired" label. The authors explicitly say there is no reverse-time sampler.

The final stage repairs the grid through 14 alternating cycles. At each strike, one half takes a cumulative maximum along maturity. The other applies weighted isotonic regression to each slice's 15 strike slopes, clips them to [-1, 0], then refits a bounded intercept. The training set contains 640 surfaces, validation 128 and testing 160. The complete run takes 24.2 s on a CPU.

Who earns the 42.6% MAE improvement?

Against interpolation, the full method cuts MAE by 42.6%, to 0.001886 from 0.003288. RMSE falls to 0.002756 from 0.005059. The four-way comparison makes clear how little of that gain belongs to repair. When applied directly to interpolation, the same 14-cycle procedure worsens MAE from 0.003288 to 0.003520. After the denoiser, repair nudges MAE from 0.001909 to 0.001886, while worsening RMSE from 0.002728 to 0.002756.

The network earns the accuracy gain.

The authors describe the division of labour similarly: most predictive benefit comes from the learned prior, with feasibility supplied by the second stage. Changing the noise level leaves that reading intact. At noise 0.001, repairing the denoiser lowers MAE from 0.001622 to 0.001567. At 0.006, the corresponding change is from 0.002755 to 0.002744.

The paper leaves an important comparison unresolved. Its paired bootstrap compares interpolation with the full method alone, reporting a difference of 0.001402 and a 95% interval of [0.001319, 0.001483]. We did not find an interval for the 0.000023 MAE improvement from repairing the denoiser. That comparison is the one needed to support the second stage's accuracy contribution. The authors also acknowledge that the bootstrap conditions on one fitted checkpoint. Separately, the abstract reports MAE of 0.001679 for the unconstrained denoiser, a figure absent from Table I, where it is 0.001909.

Feasibility is where repair justifies itself. The network alone lowers calendar violations from 7.47% to 0.65%, yet leaves a larger butterfly problem. Its unrepaired output violates butterfly constraints in 17.66% of tested comparisons, against 10.38% for the interpolation fed into it. Repair brings both calendar and butterfly rates to 0.00% at a 1e-5 tolerance.

Feasible grids, worse volatility marks

The authors disclose the volatility tradeoff, though it warrants more attention than they give it. Interpolation has implied-vol MAE of 0.014725. The unconstrained denoiser lowers that to 0.008421, before repair raises it to 0.009248. Feasibility thus costs about 0.0008 of absolute vol, roughly a tenth of the denoiser's error.

The scoring samples differ. Both IV MAEs use only points that invert, covering 98.33% of the denoiser's grid and 99.46% after repair. Points made scoreable only by repair may account for part of the gap.

A relative-value book uses residuals in vol space to rank options, making this the metric that determines trades. Changing those marks changes the cheap and rich ordering. Zero butterfly violations alongside worse vol marks makes the grid more useful to a static-arbitrage filter and less useful for ranking. We discussed related tail damage from forcing surfaces arbitrage-free in an earlier review.

SPY remains a prospective test

We could not run a test on the paper's market because there is no market in the study. It uses synthetic European calls at zero rates. The constraints likewise apply to European calls at zero rates, and the authors say funding and dividends require a forward and discount adjustment. Our version uses listed SPY options, which are American. Before repair, it adjusts end-of-day SPY calls for early-exercise value as well as funding and dividends. This adjustment introduces another source of model risk.

The learned prior creates a separate difficulty. Training uses three-component lognormal mixtures; the authors put negative skew among future generator extensions. Applying the model to SPY is therefore an out-of-distribution test unsupported by the synthetic results. The authors describe their exercise as "reconstruction under controlled incompleteness rather than forecasting or calibration of historical AAPL surfaces." They also acknowledge that the altered-noise sets do not demonstrate resistance to changes in the structural family.

Our prospective test reconstructs each day's surface after hiding a random subset of quotes. It scores held-out vol residuals separately for repaired and unrepaired surfaces. We lack option bid-ask quotes, so the evaluation can use marks and assumed costs, but cannot validate executable arbitrage profits. A strategy constructed this way is currently being evaluated. The paper's 0.001886 and 0.009248 figures supply no SPY evidence, and we have not tested the method as published.

The abstract's figures do not reconcile with the body

Several reported figures disagree. The abstract gives 1,280 surfaces; the splits in the body total 928. It also claims 55% missing observations. In the body, each entry is retained with probability 0.30 and the four corners are always kept, leaving roughly 70% missing. For the unconstrained denoiser, the abstract reports MAE of 0.001679 against 0.001909 in Table I.

The abstract further claims zero strike-monotonicity and strike-bound violations. Section 6.1 describes the slope-bound checks, and the conclusion says none were detected. No rate appears in a table anywhere.

The references also contain mismatches. Fengler's spline smoothing is attributed to [8], a credit-default paper. The citation for VAE surfaces is [13], a seismic-risk benchmark; denoising score matching is attributed to [16], a tax-anomaly study.

These problems leave the ablation's logic intact on the body's own figures. The CNN looks promising as a reconstruction component. Repair serves as a feasibility filter, with a price paid in vol space. A held-out market test in which the repaired surface beats the unrepaired surface on vol error would change our view.