pith. sign in

arxiv: 2405.17770 · v2 · pith:NFRFZLVGnew · submitted 2024-05-28 · 💱 q-fin.MF · q-fin.PR

Risk-Neutral Generative Networks

Zhonghao Xian , Xing Yan , Cheuk Hang Leung , Qi Wu This is my paper

Pith reviewed 2026-05-24 01:28 UTC · model grok-4.3

classification 💱 q-fin.MF q-fin.PR
keywords risk neutral densitygenerative neural networkoption pricingno arbitrageterm structureskewnesskurtosis
0
0 comments X

The pith

A neural network generative model extracts no-arbitrage risk-neutral densities from option prices.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a generative neural network that models the distribution of log-returns across different maturities by drawing from a standard normal and using networks to set the term structure of mean, variance, skewness, and kurtosis. Stringent training conditions are added to enforce no-arbitrage across all strikes and maturities. This setup allows fast sampling to price any option and to recover the implied risk-neutral density. The resulting densities can take many different shapes, and the model fits market data more accurately and stably than a range of parametric and stochastic process alternatives. Readers would care if they want a flexible, data-driven way to capture the market's view of future price distributions without building in specific assumptions about volatility dynamics.

Core claim

By representing the term structures of the location, scale, and higher-order moments of log-returns with neural networks and training them under no-arbitrage constraints, the generative model produces samples that can price options across strikes and maturities while extracting risk-neutral densities of diverse shapes, outperforming three parametric models and nine stochastic process models in accuracy and stability.

What carries the argument

The generative mapping from standard normal variates to log-returns, parameterized by neural networks that output maturity-dependent moments, subject to no-arbitrage learning constraints.

If this is right

  • Option prices for arbitrary strikes and maturities can be obtained by generating many samples from the model.
  • Risk-neutral densities can be recovered that display a wide variety of shapes due to flexible higher moments.
  • The approach yields flexible term structures for risk-neutral skewness and kurtosis.
  • The accuracy and stability exceed those of common parametric and stochastic models.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The model could be extended to price path-dependent claims by generating full trajectories.
  • Updates to market option prices could allow near real-time recalibration of the densities.
  • Similar generative structures might apply to other derivatives markets with sparse data.

Load-bearing premise

Imposing stringent conditions on the neural network training process is enough to ensure that generated prices remain arbitrage-free for every strike and maturity while still matching observed market prices.

What would settle it

Generate prices for a dense grid of strikes and maturities and check whether any butterfly spread or calendar spread violates no-arbitrage bounds; if violations occur, the claim fails.

Figures

Figures reproduced from arXiv: 2405.17770 by Cheuk Hang Leung, Qi Wu, Xing Yan, Zhonghao Xian.

Figure 1
Figure 1. Figure 1: The true Heston RND and the recovered RNDs by our proposed models. Overall, RN-DMLP gives the closest recovered RND. Subfigure (a) displays the left-skewed RND generated by the Heston model with parameters: ν0 = 0.05, ϑ = 0.25, κ = 0.15, ξ = 0.35, ρ = −0.9; Subfigure (b) displays the likely-normal RND generated by the Heston model with parameters: ν0 = 0.05, ϑ = 0.25, κ = 0.15, ξ = 0.25, ρ = −0.2; Subfigur… view at source ↗
Figure 2
Figure 2. Figure 2: Pricing performance under the single τ setting. The log10 MSE (averaged in every quarter) on the testing set is plotted against varying time for each model. The shaded areas represent the financial crisis periods. Overall, RN-DMLP achieves the best performance. and the out-of-sample prediction/evaluation is conducted using both MSE and relative MSE, the latter of which is computed by L(Ci , Cˆ i) = (Cˆ i/C… view at source ↗
Figure 3
Figure 3. Figure 3: Pricing performance under the multiple τ setting. The log10 MSE (averaged in every quarter) on the testing set is plotted against varying time for each model. The shaded areas represent the financial crisis periods. Overall, RN-DMLP achieves the best performance and RN-MLP is the second best. They also demonstrate evident superiority to the 9 classical methods, achieving MSE lev￾els in the order of approxi… view at source ↗
Figure 4
Figure 4. Figure 4: Original and the perturbed RNDs of the underlying price at the maturity date estimated by our model RN-DMLP, which is obtained directly from the estimated RNDs of the log-return [PITH_FULL_IMAGE:figures/full_fig_p027_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: This figure displays the monthly risk-neutral volatility (RNM2) extracted by the RN-DMLP model and the VIX index from January 1996 to February 2023, as well as their correlation coefficient. The RNM2 is computed from the RND extracted using option contracts with time-to-maturities between 25 and 35 days. The shaded grey areas represent the financial crisis periods. In the figure, both series have been stan… view at source ↗
Figure 6
Figure 6. Figure 6: Risk-neutral moments’ summary statistics and the representative RNDs with different time￾to-maturities between 1 week (1W) and 1 year (1Y). Subfigure (a) presents the box plots of the three risk-neutral moments RNM2 – RNM4 in every time-to-maturity group. We observe increasing patterns in the second and fourth-order moments, and a decreasing pattern in the third moment as the time-to￾maturity increases. Su… view at source ↗
Figure 7
Figure 7. Figure 7: The term structures of risk-neutral moments (RNM2 – RNM4) produced by our RN-DMLP model (a, c, and e) and by the Heston model (b, d, and f), with the time horizon increasing from 1 week to 1 year. We select five trading days for RN-DMLP and three parameter settings for the Heston. The term structures of RNM3 and RNM4 produced by RN-DMLP demonstrate more flexible patterns, such as the strictly monotonic RNM… view at source ↗
read the original abstract

We present a generative approach to price options and extract risk-neutral densities from the market. Specifically, we model the underlying log-returns on the time-to-maturity continuum as a generative model from standard normal. Neural nets are used to represent the term structures of the location, the scale, and the higher-order moments. We impose stringent conditions on the learning process to ensure no arbitrage. This model allows for the efficient generation of samples to price options across strikes and maturities. We have validated the effectiveness of this approach by benchmarking it against a comprehensive set of baseline models. Experiments show that the extracted risk-neutral densities accommodate a diverse range of shapes. Its accuracy significantly outperforms the extensive set of baseline models--including three parametric models and nine stochastic process models--in terms of accuracy and stability. The success of this approach is attributed to its capacity to offer flexible term structures for risk-neutral skewness and kurtosis.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes a generative neural-network model for option pricing and risk-neutral density extraction. Log-returns are generated from a standard normal via neural nets that parameterize the term structures of location, scale, skewness and kurtosis; stringent regularization and moment-matching constraints are imposed during training to enforce no-arbitrage. The approach is benchmarked against three parametric models and nine stochastic-process models on held-out strikes and maturities, with claims of superior accuracy, stability, and the ability to produce diverse density shapes.

Significance. If the no-arbitrage constraints are shown to hold rigorously and the reported outperformance is reproducible on standard option datasets, the method would supply a flexible, data-driven alternative to classical parametric and process-based models for risk-neutral density estimation.

major comments (2)
  1. [Methodology / training objective] The section describing the training objective (regularization terms on the neural-net outputs for location/scale/skew/kurtosis together with the explicit moment-matching constraints) must include the precise mathematical statements of those constraints and a proof or numerical verification that they eliminate static arbitrage for arbitrary strikes and maturities.
  2. [Experiments / benchmark results] Table or figure reporting the benchmark results: the outperformance claim is central, yet the manuscript must state the exact dataset (underlying, date range, number of options), the train/test split, and the precise error metric (e.g., RMSE on implied volatility or price) used for each baseline; without these the numerical superiority cannot be assessed.
minor comments (2)
  1. [Abstract] The abstract asserts 'significantly outperforms' without any numerical values; adding one or two headline metrics would improve readability.
  2. [Notation] Notation for the neural-network outputs (location, scale, skewness, kurtosis) should be introduced once and used consistently throughout the text and equations.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback on our manuscript. We address each major comment below and will revise the paper accordingly to improve clarity and rigor.

read point-by-point responses

  1. Referee: [Methodology / training objective] The section describing the training objective (regularization terms on the neural-net outputs for location/scale/skew/kurtosis together with the explicit moment-matching constraints) must include the precise mathematical statements of those constraints and a proof or numerical verification that they eliminate static arbitrage for arbitrary strikes and maturities.

    Authors: We agree that a more rigorous presentation is required. The manuscript describes the constraints in Section 3 but does not provide their full mathematical formulation or a verification of no-arbitrage. In the revision we will add the exact mathematical statements of the regularization and moment-matching terms together with either a short proof or a comprehensive numerical verification that static arbitrage is eliminated for arbitrary strikes and maturities. revision: yes

  2. Referee: [Experiments / benchmark results] Table or figure reporting the benchmark results: the outperformance claim is central, yet the manuscript must state the exact dataset (underlying, date range, number of options), the train/test split, and the precise error metric (e.g., RMSE on implied volatility or price) used for each baseline; without these the numerical superiority cannot be assessed.

    Authors: We acknowledge that these experimental details must be stated explicitly. While some information appears in Section 4, we will revise the manuscript to clearly report the underlying asset, exact date range, total number of options, train/test split, and the precise error metric (RMSE on implied volatility) applied to every baseline. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper models log-returns via neural nets for term structures of location/scale/skew/kurtosis and imposes regularization plus moment-matching constraints to enforce no-arbitrage. Reported accuracy is evaluated on held-out strikes and maturities against external baselines (parametric models and stochastic processes). No equations reduce any reported prediction to a quantity already fitted inside the same model, no self-citation is load-bearing on the central claim, and the no-arbitrage conditions are stated as explicit training constraints rather than derived from the outputs themselves. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

Review is abstract-only; the ledger therefore records only the high-level modeling choices stated in the abstract. Neural-network weights constitute a large set of free parameters whose number and regularization are unspecified. The generative assumption that log-returns arise from a standard normal is taken as given.

free parameters (1)
  • Neural-network weights for location, scale, skewness and kurtosis term structures
    The model uses neural nets to represent these functions; all weights are fitted to market data.
axioms (2)
  • domain assumption Log-returns on the time-to-maturity continuum can be represented as a generative map from standard normal whose moments are continuous functions of maturity
    Explicitly stated as the modeling premise in the abstract.
  • domain assumption Stringent conditions imposed during learning suffice to eliminate arbitrage opportunities
    Stated as the mechanism that keeps the model inside the no-arbitrage set.

pith-pipeline@v0.9.0 · 5681 in / 1442 out tokens · 33645 ms · 2026-05-24T01:28:38.676541+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

40 extracted references · 40 canonical work pages

  1. [1]

    Ackerer, D., Tagasovska, N., and Vatter, T. (2020). Deep smoothing of the implied volatility surface. Advances in Neural Information Processing Systems , 33:11552--11563

    work page 2020

  2. [2]

    Amilon, H. (2003). A neural network versus black--scholes: a comparison of pricing and hedging performances. Journal of Forecasting , 22(4):317--335

    work page 2003

  3. [3]

    Bahra, B. (1997). Implied risk-neutral probability density functions from option prices: theory and application. Technical report, Bank of England

    work page 1997

  4. [4]

    and Hornik, K

    Baldi, P. and Hornik, K. (1989). Neural networks and principal component analysis: Learning from examples without local minima. Neural Networks , 2(1):53--58

    work page 1989

  5. [5]

    Barndorff-Nielsen, O. E. (1997). Normal inverse gaussian distributions and stochastic volatility modelling. Scandinavian Journal of Statistics , 24(1):1--13

    work page 1997

  6. [6]

    Bates, D. S. (1996). Jumps and stochastic volatility: Exchange rate processes implicit in deutsche mark options. The Review of Financial Studies , 9(1):69--107

    work page 1996

  7. [7]

    M., McDonald, J

    Bookstaber, R. M., McDonald, J. B., et al. (1987). A general distribution for describing security price returns. The Journal of Business , 60(3):401--424

    work page 1987

  8. [8]

    B., and Yor, M

    Carr, P., Geman, H., Madan, D. B., and Yor, M. (2003). Stochastic volatility for l \'e vy processes. Mathematical Finance , 13(3):345--382

    work page 2003

  9. [9]

    and Berger, R

    Casella, G. and Berger, R. (2024). Statistical inference . CRC Press

    work page 2024

  10. [10]

    Cont, R. (1997). Beyond implied volatility: Extracting information from options prices. Econophysics. Dordrecht: Kluwer

    work page 1997

  11. [11]

    Cont, R. (2001). Empirical properties of asset returns: stylized facts and statistical issues. Quantitative Finance , 1(2):223

    work page 2001

  12. [12]

    and Vuleti \'c , M

    Cont, R. and Vuleti \'c , M. (2023). Simulation of arbitrage-free implied volatility surfaces. Applied Mathematical Finance , 30(2):94--121

    work page 2023

  13. [13]

    Cox, J. C. and Ross, S. A. (1976). The valuation of options for alternative stochastic processes. Journal of Financial Economics , 3(1-2):145--166

    work page 1976

  14. [14]

    Davis, M. H. and Hobson, D. G. (2007). The range of traded option prices. Mathematical Finance , 17(1):1--14

    work page 2007

  15. [15]

    Dugas, C., Bengio, Y., B \'e lisle, F., Nadeau, C., and Garcia, R. (2000). Incorporating second-order functional knowledge for better option pricing. Advances in Neural Information Processing Systems , 13

    work page 2000

  16. [16]

    and Raible, S

    Eberlein, E. and Raible, S. (1999). Term structure models driven by general l \'e vy processes. Mathematical Finance , 9(1):31--53

    work page 1999

  17. [17]

    Fengler, M. R. (2009). Arbitrage-free smoothing of the implied volatility surface. Quantitative Finance , 9(4):417--428

    work page 2009

  18. [18]

    and Schied, A

    F \"o llmer, H. and Schied, A. (2011). Stochastic finance: an introduction in discrete time . Walter de Gruyter

    work page 2011

  19. [19]

    and Gen c ay, R

    Garcia, R. and Gen c ay, R. (2000). Pricing and hedging derivative securities with neural networks and a homogeneity hint. Journal of Econometrics , 94(1-2):93--115

    work page 2000

  20. [20]

    Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., and Bengio, Y. (2014). Generative adversarial nets. Advances in Neural Information Processing Systems , 27

    work page 2014

  21. [21]

    Hamidieh, K. (2014). Rnd: Risk neutral density extraction package. R package version , 1

    work page 2014

  22. [22]

    Harrison, J. M. and Kreps, D. M. (1979). Martingales and arbitrage in multiperiod securities markets. Journal of Economic Theory , 20(3):381--408

    work page 1979

  23. [23]

    Harrison, J. M. and Pliska, S. R. (1981). Martingales and stochastic integrals in the theory of continuous trading. Stochastic Processes and Their Applications , 11(3):215--260

    work page 1981

  24. [24]

    Heston, S. L. (1993). A closed-form solution for options with stochastic volatility with applications to bond and currency options. The Review of Financial Studies , 6(2):327--343

    work page 1993

  25. [25]

    Hornik, K., Stinchcombe, M., and White, H. (1989). Multilayer feedforward networks are universal approximators. Neural Networks , 2(5):359--366

    work page 1989

  26. [26]

    M., Lo, A

    Hutchinson, J. M., Lo, A. W., and Poggio, T. (1994). A nonparametric approach to pricing and hedging derivative securities via learning networks. The Journal of Finance , 49(3):851--889

    work page 1994

  27. [27]

    and Wetterau, D

    Kienitz, J. and Wetterau, D. (2013). Financial modelling: Theory, implementation and practice with MATLAB source . John Wiley & Sons

    work page 2013

  28. [28]

    and Seneta, E

    Madan, D. and Seneta, E. (1990). The variance gamma (vg) model for share market returns. Journal of Business , 63(4):511--524

    work page 1990

  29. [29]

    Merton, R. C. (1976). Option pricing when underlying stock returns are discontinuous. Journal of Financial Economics , 3(1-2):125--144

    work page 1976

  30. [30]

    Orosi, G. (2015). Estimating option-implied risk-neutral densities: a novel parametric approach. Journal of Derivatives , 23(1):41

    work page 2015

  31. [31]

    Roper, M. (2010). Arbitrage free implied volatility surfaces. preprint

    work page 2010

  32. [32]

    Rubinstein, M. et al. (1998). Edgeworth binomial trees. Journal of Derivatives , 5:20--27

    work page 1998

  33. [33]

    and Westerhoff, F

    Schmitt, N. and Westerhoff, F. (2017). On the bimodality of the distribution of the s&p 500's distortion: Empirical evidence and theoretical explanations. Journal of Economic Dynamics and Control , 80:34--53

    work page 2017

  34. [34]

    and Symens, S

    Schoutens, W. and Symens, S. (2003). The pricing of exotic options by monte--carlo simulations in a l \'e vy market with stochastic volatility. International Journal of Theoretical and Applied Finance , 6(08):839--864

    work page 2003

  35. [35]

    and Wissel, J

    Schweizer, M. and Wissel, J. (2008). Arbitrage-free market models for option prices: The multi-strike case. Finance and Stochastics , 12:469--505

    work page 2008

  36. [36]

    and Xiu, D

    Song, Z. and Xiu, D. (2016). A tale of two option markets: Pricing kernels and volatility risk. Journal of Econometrics , 190(1):176--196

    work page 2016

  37. [37]

    and Yan, X

    Wu, Q. and Yan, X. (2019). Capturing deep tail risk via sequential learning of quantile dynamics. Journal of Economic Dynamics and Control , 109:103771

    work page 2019

  38. [38]

    Yan, X., Wu, Q., and Zhang, W. (2019). Cross-sectional learning of extremal dependence among financial assets. Advances in Neural Information Processing Systems , 32

    work page 2019

  39. [39]

    Yan, X., Zhang, W., Ma, L., Liu, W., and Wu, Q. (2018). Parsimonious quantile regression of financial asset tail dynamics via sequential learning. Advances in Neural Information Processing Systems , 31

    work page 2018

  40. [40]

    Yang, Y., Zheng, Y., and Hospedales, T. (2017). Gated neural networks for option pricing: Rationality by design. In Proceedings of the AAAI Conference on Artificial Intelligence , volume 31

    work page 2017