A ninefold reduction in training-sample pricing error, with the pricing formula held fixed, mostly indicts the benchmark calibration. Across 8,286 contract-day observations from 2 January to 6 August 2024, total RMSE falls from 0.0668 for RS-HHW to 0.0072 for its deep-learning-calibrated version. That objection remains after giving the paper credit for what it builds well, which is quite a lot.

Li starts with a Heston-Hull-White hybrid. A two-state continuous-time Markov chain then controls an additional diffusion loading in the variance equation, lambda_1 or lambda_2 according to the regime. The paper calls the model RS-HHW. Under the T-forward measure, using the Hull-White zero-coupon bond as numéraire, the conditional characteristic function stays affine in variance, rate and log-price. A 2x2 matrix ODE derived from the finite-state Feynman-Kac formula gathers the regime term, and Fourier inversion produces European call prices. Setting lambda_1 = lambda_2 = 0 recovers plain HHW, giving the paper a clean nested benchmark.

The calibration layer marks the departure from standard deep-calibration work. Its network never produces a price directly. Instead, it estimates how the RS-HHW price changes when the parameter vector moves away from a baseline. Li draws 1,000 parameter vectors by Latin hypercube inside a local box around that baseline, computes the semi-analytical price perturbation for each one, and standardizes the result. A three-expert mixture-of-experts network learns those perturbations from the parameter vector and contract features.

A second network turns 40 signature contract prices into an initial parameter guess. During online calibration, annealing followed by fmincon minimizes weighted error against the surrogate. Candidate vectors are then refined under the true RS-HHW objective using up to 1,200 representative training observations. Final prices always come from the original formula. Li says this directly: "The neural network is not used to replace the semi-analytical pricing formula."

That restraint makes sense. The paper's timing table puts a semi-analytical evaluation at 0.0018 seconds, versus 129.9359 seconds for a 500,000-path Monte Carlo in MATLAB R2024a without parallel computing, about 72,000 times faster. There is little reason to learn a substitute for such a cheap formula. Accelerating the search is the useful application.

Can the variance process remain positive?

The additive lambda_X dW_3 term invalidates the usual CIR Feller guarantee. The paper acknowledges the problem: "the standard CIR Feller condition alone is not sufficient to guarantee strict positivity of the variance process." Its remedy combines an admissibility restriction and calibration penalties, while the Monte Carlo applies max{v,0} truncation. The simulation therefore prices a truncated process, whereas the formula uses the untruncated process. Their difference is never measured.

Validation also uses textbook parameters from He and Lin (S0 = 100, K = 100, T = 1, v0 = 0.05), rather than the parameters estimated from the data. With 500,000 paths and S0 = 100, the semi-analytical price is 18.5375 and the Monte Carlo estimate is 18.5467. Across the grid S0 in [90, 110], the maximum relative error reaches 0.9532%. Nine-tenths of a percent on an 18.54 option is roughly 0.18 in price units. The empirical differences, by contrast, are absolute gaps in the third decimal place for options priced below 0.50. Those figures use different units and cannot sensibly be set beside each other.

The dataset contains daily SSE 50 ETF call quotations from Wind for 2024 expiries, covering 2 January to 30 September 2024 and 181 trading dates. Filtering reduces 13,988 initial contract-day records to 9,520. The paper's attrition table attributes the three largest cuts to the trading-date restriction, which removes 3,119, the seven-calendar-day maturity filter, which removes 612, and a no-arbitrage bound check, which removes 542. The first 144 dates form the training sample with 8,286 observations. The remaining 37 dates provide 1,234 test observations.

We ran no test of our own on any of this. SSE 50 ETF options are unavailable on our platform, and these pricing figures would not carry over to SPY or other liquid US equity-index chains. The mechanism itself transfers readily: cross-sectional prices paired with strikes, maturities and a rate input.

The win reveals the calibration problem

Out-of-sample total RMSE is 0.0179 for DL-RS-HHW. The corresponding figures are 0.0645 for RS-HHW, 0.1118 for Heston, 0.1123 for HHW and 0.1242 for Bates. MAE is 0.0097, compared with 0.0570, 0.1024, 0.1002 and 0.1115. Against RS-HHW, the in-sample reduction is ninefold, from 0.0668 to 0.0072. Out of sample, the ratio is 3.6x.

The pricing formula is identical for RS-HHW and DL-RS-HHW.

Only the parameter vector changes.

Cutting in-sample error ninefold through a better starting point and search implies that the rolling-window calibrations finished far from their own optimum. The exercises also differ by construction. Equation 4.3 gives the benchmark parameters as the average of window-specific estimates, calculated over 40-day windows with a 40-day step. DL-RS-HHW instead uses one vector, refined against the true RS-HHW objective on up to 1,200 representative training observations. Pricing at the average of several non-convex window optima answers a different question from pooled-sample minimization. The paper never isolates how much of the 0.0668-to-0.0072 difference comes from that design choice.

Li's description supports this interpretation. The abstract concludes that "residual-surrogate-assisted calibration improves the empirical calibration of semi-analytical hybrid option pricing models", while the conclusion says the deep-learning component "was used only to assist calibration". Yet the DL row appears beside Heston and Bates as though it represented a competing model, instead of a new fit of the RS-HHW row immediately above it.

The estimated parameters make the issue harder to ignore. In the DL-RS-HHW row, q12 drops from 7.9542 to 0.1413, roughly a 56x shift in one transition intensity. The same row moves rho from -0.3980 to -0.0126. This fit effectively removes the leverage effect for an equity underlying whose daily returns have skewness 2.7519 and kurtosis 15.3040, while still reducing pricing error by an order of magnitude. We read the objective surface as close to flat in rho and q12. The paper supplies no likelihood profile or identification check that would settle the matter.

The DL row consequently says little about regime dynamics. Evidence for the regime-switching component comes from the RS-HHW row, where rho is -0.3980 and q12 is 7.9542, compared with Heston, Bates and HHW. None of the calibration objections above alters that comparison.

Out of sample, the result thins

The out-of-sample RMSE is 2.5x its in-sample value, rising from 0.0072 to 0.0179. By moneyness, ATM error moves from 0.0086 to 0.0230, while ITM moves from 0.0062 to 0.0234. Those increases are 2.7x and 3.8x. OTM error barely changes, from 0.0069 to 0.0066. Benchmark errors stay nearly flat across the split: RS-HHW goes from 0.0668 to 0.0645, and Heston from 0.1176 to 0.1118.

Li reports the same behavior in the fitted-price plots, observing that the model slightly underestimates the price spike near the sample's end. One parameter vector remains fixed through the entire 37-date test period. No walk-forward re-estimation is reported. The 0.0179 figure therefore records the performance of a frozen vector over 37 dates.

The statistical case is less persuasive than its t-statistics appear. Diebold-Mariano tests use the out-of-sample pricing errors, though the paper does not say how many observations enter each test. The window contains 1,234 contract-day records. Dependence is heavy, with multiple strikes observed on the same date and the same contracts appearing on adjacent dates. The reported statistics of -13.899 against RS-HHW under MSE loss and -17.773 under MAE loss treat each observation as independently informative.

Li already plays down the Pesaran-Timmermann results, calling them "complementary evidence to the RMSE, MAE, and DM results rather than as the sole basis for model ranking". The paper also flags the identical statistic of 4.667 for RS-HHW and HHW. Even that discount looks understated. Through the pricing formula, model price changes mechanically follow the observed underlying and time decay. Every model therefore clears the 1% level, including Bates at 3.667 and Heston at 4.333. It is difficult to see what the test contributes.

No trading result

There is no P&L or hedging test. The reported pricing MAE of 0.0097 is in yuan, while the paper's time-series plot places the daily average option price below 0.50. The error is about 2% of a 0.50 contract and a larger share for cheaper contracts. Tradability depends on quantities left unmeasured: the crossed spread, persistence of the gap, hedge error over the holding period, and the recalibration frequency needed to preserve the signal. Li states this limitation plainly, listing early exercise, transaction costs, bid-ask spreads, liquidity and microstructure among the omitted features. We would recalibrate each date instead of freezing one vector for 37 of them. The paper does not test that approach.

One result survives every reservation. Bates performs worst, with in-sample RMSE of 0.1376 and out-of-sample RMSE of 0.1242. Plain Heston records 0.1176 and 0.1118 on the same sample, despite kurtosis above 15. The fitted jump parameters are lambda_J 0.3063, mu_J 0.2549 and sigma_J 0.1683. Across this chain and these nine months, they contribute nothing to tail fit.

Heston and HHW are indistinguishable under MSE loss (DM 0.719, p=0.4724). Under MAE loss, HHW has the lower average loss (DM -4.019, p=0.0001). Stochastic rates alter the shape of the error distribution without shifting its second moment. These are useful negative findings about the benchmark set.

We have written before about apparent edges that reside in the comparison more than the method: a stress index whose margin over one baseline shrank against a sharper one. The same pattern appears here. The residual surrogate is defensible engineering, and the semi-analytical derivation appears sound.

One table would change our view of the headline. RS-HHW should be recalibrated on the pooled training sample with the same annealing budget and the same 1,200-observation refinement, then priced out of sample. If its gap to 0.0179 largely disappears, the network amounts to a convenience. If the gap remains, Li has something.