REVIEW 4 major objections 4 minor 14 references
Empirical Models of the Time Evolution of SPX Option Prices
T0 review · 4 major / 4 minor · reviewed 2026-08-15 · deepseek-v4-flash
Pith's one-line read A tiny neural net beats Black-Scholes on SPX puts.
desk verdict Solid empirical benchmark undermined by an overclaimed arbitrage-free result that its own tests refute. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is the feedforward pricing function $f_\theta(X)$: a network with two hidden layers of four ReLU neurons that maps option characteristics (strike, moneyness, time to maturity) and market state variables (SPX level, dividend yield, risk-free rate, GARCH(1,1) forecast volatility) to a put price. The GARCH(1,1) volatility estimate, re-estimated daily on a rolling 252-day window, supplies time-varying volatility information without relying on an implied-volatility surface. Training minimizes the Huber loss, which keeps the network from being dominated by a few high-priced options and preserves accuracy on the cheap deep-OTM contracts; the design that includes or excludes the Black-Scholes price as an input isolates how much of each model's performance is borrowed from the theoretical benchmark. SHAP values and a principal-component comparison are then used to argue that the network relies on strike, maturity, and index level in the way option theory would predict.
What would settle it
On a fixed test date, compute the network's prices on a dense strike-maturity grid and search for a static portfolio of puts, such as a butterfly or a calendar spread, with nonnegative payoffs and negative cost; if such a portfolio exists on more than a negligible fraction of dates, the claim that the network delivers arbitrage-free prices without constraints is false.
Extended reading notes
Core claim
On its own terms, the paper claims that an off-the-shelf two-layer feedforward network with four ReLU neurons per layer, trained with Adam, weight decay, early stopping, and the Huber loss, approximates the European put pricing function for SPX more accurately than Black-Scholes-Merton out of sample across 1996–2022. The advantage is largest for deep out-of-the-money puts and during heightened-volatility periods, and it persists whether or not the Black-Scholes price is included as a feature. The paper also claims that the trained network yields prices consistent with no-arbitrage conditions in the large majority of perturbed cases—about 93.5 percent monotone in strike, 95.1 percent convex in strike, and 82.9 percent monotone in time to maturity—without these conditions being enforced during training.
Load-bearing premise
The load-bearing premise is that European put prices must increase with time to maturity, which the paper treats as a no-arbitrage condition; with nonzero interest rates and dividend yields this monotonicity does not generally hold, so the violation statistics do not by themselves establish economic consistency.
Editorial extensions
If this is right
- A single small network can replace a daily recalibrated implied-volatility surface for pricing the SPX put cross-section and remains accurate out of sample.
- Because options are priced directly from state variables, the model can produce simulated option prices along GARCH-simulated SPX paths without interpolating or extrapolating a volatility surface.
- The neural network does not need the Black-Scholes price as an input, so it can be used in regimes where the theoretical model is misspecified.
- The random forest degrades sharply without the Black-Scholes feature and linear regression generally stays below the Black-Scholes baseline, so the network is the only tested model that is both nonlinear and self-sufficient.
- Residual no-arbitrage violations appear when inputs are perturbed far from the original point, so deployment would still benefit from regularization or post-hoc correction.
Reading between the lines
- The paper's monotonicity-in-maturity test is not a valid arbitrage bound for European puts when interest rates and dividend yields are nonzero, so the 17.1 percent violation rate is better interpreted as a property of the test than as evidence that the model admits arbitrage.
- A proper economic-consistency check would test calendar-spread and butterfly arbitrage on a dense strike-maturity grid; the paper's local perturbation procedure only probes neighborhoods of observed points.
- A natural extension is to train the same architecture on SPX calls or on a multi-asset index to see whether the accuracy and no-arbitrage findings generalize beyond puts.
- The small network size suggests an implicit smoothness prior; pricing on a dense grid and inspecting the implied volatility smile would reveal whether the learned surface is itself arbitrage-free.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper develops statistical models for pricing S&P 500 index put options, using a GARCH(1,1) volatility estimate and a 1996–2022 OptionMetrics dataset. It compares a small feedforward neural network, a random forest, and linear regression under expanding and rolling training windows, with and without the Black-Scholes price as an input feature, reporting MAPE over test periods and moneyness segments. It also presents SHAP-based interpretability and a Section 7 evaluation of no-arbitrage conditions. The headline claim is that a two-hidden-layer, four-neuron-per-layer neural network trained with minimal tuning performs well against Black-Scholes-Merton and 'delivers arbitrage-free option prices without requiring these conditions to be imposed.'
Significance. If the central claims held, the paper would offer a practically useful result: a small, explainable network that both improves on Black-Scholes pricing and automatically satisfies no-arbitrage constraints. The empirical scope is a strength: roughly seven million filtered options over 27 years, two training-window schemes, three model classes, and extensive appendix tables. The SHAP and PCA analyses are also a genuine effort to connect the fitted models to economic intuition. However, the headline arbitrage-free claim is contradicted by the paper's own Section 7 results, and the performance comparisons are reported without uncertainty quantification. The paper's value as an empirical benchmark is diminished by these issues, and the central abstract claim is not supported as stated.
major comments (4)
- [Abstract; Section 7; Section 9] The abstract states that the neural network 'delivers arbitrage-free option prices without requiring these conditions to be imposed,' but Section 7's own tests report that, on 123,587 options, only 93.51% of predictions respect strike monotonicity, 95.09% respect strike convexity, and 82.92% respect time-to-maturity monotonicity. Strike monotonicity and convexity are model-free no-arbitrage bounds for European put prices, so the reported 6.49% and 4.91% violation rates directly contradict an unconditional arbitrage-free claim. Even if the TTM condition is set aside, the two valid conditions are still violated. The conclusion's phrase 'respect key no-arbitrage requirements' is therefore internally inconsistent with the evidence in Section 7, and the claim must either be removed or explicitly qualified as approximate/tolerance-based.
- [Section 5.3; Tables 6, 12, 18] The claim that the neural network 'consistently outperforms both linear regression and random forest benchmarks' is not supported by the appendix tables for in-the-money options. For example, in Table 6 the expanding-window ITM test MAPE for NN+ in the first window is 34.5% versus 5.0% for BS and 4.2% for RF+ in Table 18, and several later ITM windows show NN+ above both the BS baseline and RF+/LR+. Section 5.3's statement that the ITM analysis 'largely mirrors' the OTM findings is contradicted by these numbers. The comparisons are reported as point estimates with no confidence intervals or paired significance tests, so even the OTM outperformance claims are not statistically established.
- [Section 1; Section 9] The stated objective is a framework that 'enables the simulation of joint paths for asset prices and corresponding option prices,' and the conclusion describes the paper as providing a simulation framework. However, no simulation experiment appears anywhere in the manuscript: there are no generated paths, no distributional or tail-risk diagnostics, no stress-test application, and no code. The paper evaluates static out-of-sample pricing accuracy only, so the simulation capability is asserted rather than demonstrated.
- [Section 7] The time-to-maturity monotonicity test is not a valid no-arbitrage test for European put options when interest rates and dividend yields are nonzero, since put prices need not be increasing in maturity under those conditions. Counting the 17.1% TTM violation rate as evidence against economic consistency is therefore misleading. The Section 7 test design also perturbs only one input at a time and uses a $0.05 tolerance, so it is neither a sufficient nor a fully reported test of static arbitrage; a complete check would need simultaneous constraints across strikes and maturities, such as calendar-spread and butterfly conditions.
minor comments (4)
- [Figure 10 caption] The caption for Figure 10 describes the panels as 'NN trained with Black-Scholes information' and 'NN trained without,' but the figure panels are labeled RF+ and RF-; the caption should refer to the random forest model.
- [Equation (10)] The SHAP formula is written with an awkward summation notation and a garbled subscript; it should be rewritten as a sum over all subsets S not containing i, with the standard Shapley weights.
- [Section 8.1] The sentence 'MAPE increases as options move deeper OTM, MAPE increases significantly across all models' contains a duplicated subject and should be edited for clarity.
- [General] The manuscript does not provide a data availability statement or code repository; given the volume of numerical results, making the preprocessing and training pipelines available would materially improve reproducibility.
Circularity Check
No circularity: the paper's pricing models are fitted to observed market prices and evaluated out-of-sample; the no-arbitrage claim is an empirical test, not a derivation from its own inputs.
full rationale
The models are trained on market option prices and tested on held-out periods, so the predicted prices are not equal to any fitted input by construction. The GARCH parameters are estimated from SPX index returns, which is an external input, and the Black-Scholes price is used only as a feature or baseline; including it does not force the neural network output to equal the Black-Scholes value because the target remains the observed market price. No uniqueness theorem, ansatz-smuggling self-citation, or renamed empirical pattern carries the central claim. The abstract's 'arbitrage-free' statement is contradicted in part by the paper's own Section 7 violation counts, but that is a correctness or internal-consistency concern, not circularity: the claim is testable and, if the numbers are accurate, false as stated, rather than true by definition. The paper also acknowledges in Section 8.2 that the neural network 'can still produce such violations under certain training conditions,' which further confirms the claim is empirical rather than definitionally forced. Self-citations such as Brini and Lenz 2024 appear only in related-work context and are not load-bearing. Overall, the derivation chain is self-contained with respect to the data and benchmark comparisons, so the appropriate circularity score is 0.
Assumptions & free parameters
free parameters (3)
- GARCH(1,1) parameters (mu, a0, a1, b1) =
Re-estimated daily on rolling 252-day window
- NN hyperparameters (2 hidden layers, 4 neurons per layer, Huber delta=1, learning rate 1e-4, weight decay 1e-3, batch… =
Chosen by hand, not tuned
- RF hyperparameters (100 trees, max depth 10) =
Chosen by hand
assumptions (4)
- domain assumption SPX log returns follow a GARCH(1,1) process with standard normal innovations.
- domain assumption No-arbitrage bounds are characterized by monotonicity in strike and maturity and convexity in strike, as tested in Section 7.
- domain assumption OptionMetrics quotes from 1996-2022 with the stated filters are representative of SPX put market prices.
- domain assumption MAPE is an appropriate error metric for comparing option pricing performance across orders of magnitude.
Cite this review
Pith. "Pith review of Empirical Models of the Time Evolution of SPX Option Prices." pith.science (2026). https://pith.science/paper/LL5VZTZP
@misc{pith2026250617511,
author = {Pith},
title = {Pith review of: Empirical Models of the Time Evolution of SPX Option Prices},
year = {2026},
howpublished = {\url{https://pith.science/paper/LL5VZTZP}},
note = {Machine review of arXiv:2506.17511}
}
read the original abstract
The key objective of this paper is to develop an empirical model for pricing SPX options that can be simulated over future paths of the SPX. To accomplish this, we formulate and rigorously evaluate several statistical models, including neural network, random forest, and linear regression. These models use the observed characteristics of the options as inputs -- their price, moneyness and time-to-maturity, as well as a small set of external inputs, such as the SPX and its past history, dividend yield, and the risk-free rate. Model evaluation is performed on historical options data, spanning 30 years of daily observations. Significant effort is given to understanding the data and ensuring explainability for the neural network. A neural network model with two hidden layers and four neurons per layer, trained with minimal hyperparameter tuning, performs well against the theoretical Black-Scholes-Merton model for European options, as well as two other empirical models based on the random forest and the linear regression. It delivers arbitrage-free option prices without requiring these conditions to be imposed.
Figures
Figures from the paper (27 more)
Reference graph
Works this paper leans on
-
[3]
doi: https://doi.org/10.1016/j.jfineco.2021.12.007
ISSN 0304-405X. doi: https://doi.org/10.1016/j.jfineco.2021.12.007. URL https://www.sciencedirect.com/science/article/pii/S0304405X21005249. Jay Cao, Jacky Chen, and John Hull. A neural network approach to understanding implied volatility movements.Quantitative Finance, 20(9):1405–1413,
-
[6]
The us treasury yield curve: 1961 to the present
Refet S G¨ urkaynak, Brian Sack, and Jonathan H Wright. The us treasury yield curve: 1961 to the present. Journal of monetary Economics, 54(8):2291–2304,
work page 1961
-
[8]
Andrey Itkin. Deep learning calibration of option pricing models: some pitfalls and solutions.arXiv preprint arXiv:1906.03507,
arXiv 1906
-
[9]
Scaling laws for neural language models.arXiv preprint arXiv:2001.08361,
Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models.arXiv preprint arXiv:2001.08361,
arXiv 2001
-
[13]
Laurens Van Mieghem, Antonis Papapantoleon, and Jonas Papazoglou-Hennig. Machine learning for option pricing: an empirical investigation of network architectures.arXiv preprint arXiv:2307.07657,
-
[14]
The last six months always serve as the test set
46 A Tables with Performance Results In each table, the Date column indicates the period covered, spanning from the beginning of the training set to the end of the test set. The last six months always serve as the test set. The format YY/MM - YY/MM is used, where the first date marks the start of the training set and the second denotes the end of the six-...
-
[1986]
Tim Bollerslev, Michael Gibson, and Hao Zhou
doi: 10.1016/0304-4076(86)90063-1. Tim Bollerslev, Michael Gibson, and Hao Zhou. Dynamic estimation of volatility risk premia and investor risk aversion from option-implied and realized volatilities.Journal of econometrics, 160(1):235–245,
-
[1994]
Machine learning methods for pricing financial derivatives.arXiv preprint arXiv:2406.00459,
Lei Fan and Justin Sirignano. Machine learning methods for pricing financial derivatives.arXiv preprint arXiv:2406.00459,
Show all 14 references
-
[2002]
Application of machine learning in option pricing: A review
Wenda Li. Application of machine learning in option pricing: A review. In2022 7th International Conference on Social Sciences and Economic Development (ICSSED 2022), pages 209–214. Atlantis Press,
2022
-
[2016]
Black–scholes option pricing using machine learning
Shreyan Sood, Tanmay Jain, Nishant Batra, and HC Taneja. Black–scholes option pricing using machine learning. InProceedings of International Conference on Data Science and Applications: ICDSA 2022, Volume 1, pages 481–493. Springer,
2022
-
[2017]
Dynamics of the implied volatility surface: Theory and empirical evidence
45 Jacinto Marabel Romo. Dynamics of the implied volatility surface: Theory and empirical evidence. Quantitative Finance DOI, 10(14697688.2012):686668,
2012
-
[2020]
Deep learning for exotic option valuation.arXiv preprint arXiv:2103.12551,
Jay Cao, Jacky Chen, John Hull, and Zissis Poulos. Deep learning for exotic option valuation.arXiv preprint arXiv:2103.12551,
-
[2021]
Differential machine learning.arXiv preprint arXiv:2005.02347,
Brian Huge and Antoine Savine. Differential machine learning.arXiv preprint arXiv:2005.02347,
2005 arXiv
-
[2022]
Empirical performance of alternative option pricing models.The Journal of finance, 52(5):2003–2049,
Gurdip Bakshi, Charles Cao, and Zhiwu Chen. Empirical performance of alternative option pricing models.The Journal of finance, 52(5):2003–2049,
2003
Reviewed August 15, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.