pith. sign in

arxiv: 2409.06551 · v5 · submitted 2024-09-10 · 💱 q-fin.CP

Robust financial calibration: a Bayesian approach for neural SDEs

Christa Cuchiero , Eva Flonner , Kevin Kurt This is my paper

Pith reviewed 2026-05-23 20:51 UTC · model grok-4.3

classification 💱 q-fin.CP
keywords neural SDEsBayesian calibrationimplied volatility boundsrobust calibrationchange of measurefinancial time seriesLangevin sampling
0
0 comments X

The pith

A Bayesian posterior over neural SDE weights mixes models to deliver robust bounds on the implied volatility surface.

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

The paper sets up a prior on the weights of a neural stochastic differential equation and pairs it with a likelihood that draws on both historical price paths and observed option prices. The resulting posterior is treated as a mixture across different neural SDE specifications, which supplies bounds on the implied volatility surface that remain stable even when the exact model form is uncertain. A global universal approximation result based on Barron-type estimates is stated to justify the flexibility of the neural SDE class. The change of measure between the historical and risk-neutral probabilities is learned jointly from the two data sources. Posterior samples are obtained with a Langevin-type algorithm that makes the optimization tractable.

Core claim

By placing a prior on neural-network weights inside a neural SDE and defining a likelihood from combined historical time series and option prices, the induced posterior can be read as a mixture of neural SDE models; this mixture in turn produces robust bounds on the implied volatility surface, while a Barron-type universal approximation theorem guarantees that the neural SDE class is sufficiently rich.

What carries the argument

The posterior distribution over neural-network weights, interpreted as a mixture of neural SDE models that yields robust bounds on the implied volatility surface.

If this is right

  • Robust bounds on implied volatility are obtained directly from the posterior without selecting any single neural SDE specification.
  • Historical time series and option prices are used together once the change of measure is learned inside the same posterior.
  • The global universal approximation theorem for neural SDEs justifies treating the class as flexible enough to cover a wide range of dynamics.
  • Langevin-type sampling provides a practical route to drawing from the posterior and thereby to the robust bounds.

Where Pith is reading between the lines

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

  • The same posterior-mixture construction could be applied to other calibration tasks such as local-volatility surfaces or rough-volatility models.
  • If the mixture bounds prove tighter than classical model-risk intervals, practitioners might replace separate model-selection steps with a single Bayesian calibration run.
  • The approach supplies a natural route to uncertainty-aware pricing that folds both parameter and model uncertainty into one object.

Load-bearing premise

The chosen prior on network weights together with the likelihood produces a posterior whose mixture interpretation actually supplies meaningful and robust bounds, and that the change of measure between historical and risk-neutral probabilities can be recovered from the joint data.

What would settle it

Compute the posterior mixture bounds on a held-out set of option prices and check whether the observed market implied volatilities fall outside those bounds for a substantial fraction of strikes and maturities.

Figures

Figures reproduced from arXiv: 2409.06551 by Christa Cuchiero, Eva Flonner, Kevin Kurt.

Figure 1
Figure 1. Figure 1: Robust calibration bounds on the implied volatility surface in the rough Bergomi model, exemplified with 2 [PITH_FULL_IMAGE:figures/full_fig_p010_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Robust calibration bounds on the implied volatility surface in Heston model, where the model is calibrated [PITH_FULL_IMAGE:figures/full_fig_p015_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Robust price bounds for a floating lookback put option. [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Sensitivity analysis of bounds on the implied volatility surface obtained in the rough Bergomi model. [PITH_FULL_IMAGE:figures/full_fig_p017_4.png] view at source ↗
read the original abstract

The paper presents a Bayesian framework for the calibration of financial models using neural stochastic differential equations (neural SDEs), for which we also formulate a global universal approximation theorem based on Barron-type estimates. The method is based on the specification of a prior distribution on the neural network weights and an adequately chosen likelihood function. The resulting posterior distribution can be seen as a mixture of different classical neural SDE models yielding robust bounds on the implied volatility surface. Both, historical financial time series data and option price data are taken into consideration, which necessitates a methodology to learn the change of measure between the risk-neutral and the historical measure. The key ingredient for a robust numerical optimization of the neural networks is to apply a Langevin-type algorithm, commonly used in the Bayesian approaches to draw posterior samples.

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

1 major / 1 minor

Summary. The manuscript presents a Bayesian framework for calibrating neural SDEs, including a global universal approximation theorem based on Barron-type estimates. A prior is placed on neural-network weights together with a likelihood that incorporates both historical time-series data and option prices; the resulting posterior is interpreted as a mixture of neural SDE models that supplies robust bounds on the implied-volatility surface. The framework also learns the change of measure between the historical and risk-neutral measures and employs a Langevin-type sampler for posterior draws.

Significance. If the mixture interpretation, the explicit likelihood, the Girsanov kernel, and the no-arbitrage consistency can be established and validated numerically, the work would supply a principled route to uncertainty-aware calibration that combines time-series and cross-sectional information. The Barron-type universal-approximation result, if fully proved, would be a useful theoretical addition for neural SDEs in finance.

major comments (1)
  1. [Abstract] Abstract: the central claim that the posterior 'can be seen as a mixture of different classical neural SDE models yielding robust bounds on the implied volatility surface' is load-bearing yet unsupported. Without an explicit joint likelihood on the combined data or an explicit Girsanov kernel, it is impossible to verify that posterior mass concentrates on arbitrage-free risk-neutral dynamics or that the learned Radon-Nikodym derivative is a true martingale; the mixture may therefore reflect prior volume rather than economically meaningful robust bounds.
minor comments (1)
  1. [Abstract] The phrase 'adequately chosen likelihood function' is used without stating the concrete functional form or the conditions that make the choice adequate for the mixture interpretation.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the careful reading and constructive comments on our manuscript. We appreciate the recognition of the framework's potential and address the major comment point by point below, committing to revisions that enhance explicitness and verifiability.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central claim that the posterior 'can be seen as a mixture of different classical neural SDE models yielding robust bounds on the implied volatility surface' is load-bearing yet unsupported. Without an explicit joint likelihood on the combined data or an explicit Girsanov kernel, it is impossible to verify that posterior mass concentrates on arbitrage-free risk-neutral dynamics or that the learned Radon-Nikodym derivative is a true martingale; the mixture may therefore reflect prior volume rather than economically meaningful robust bounds.

    Authors: We agree that the abstract claim requires stronger and more explicit support to be fully verifiable from the text. The manuscript defines a joint likelihood that combines historical time-series data (under the physical measure) with option prices (under the risk-neutral measure) and parameterizes a Girsanov kernel to learn the measure change. However, to directly address the concern, the revised manuscript will: (i) explicitly write the joint likelihood function, (ii) detail the Girsanov kernel and include a verification (analytic or numerical) that the Radon-Nikodym derivative is a martingale, and (iii) add numerical experiments confirming no-arbitrage consistency and that the posterior mixture produces data-driven robust bounds. These changes will make the mixture interpretation transparent and economically grounded rather than prior-driven. revision: yes

Circularity Check

0 steps flagged

No circularity; posterior defined by standard Bayesian update with independent mixture interpretation

full rationale

The derivation specifies a prior on neural-network weights and an adequately chosen likelihood, then forms the posterior by standard Bayes rule; the mixture-of-models reading follows directly as the posterior is a distribution over neural SDE parameters. No equation reduces a claimed prediction to a fitted quantity by construction, no uniqueness theorem is imported from self-citation, and the universal-approximation result is stated as formulated from Barron estimates rather than smuggled in. The change-of-measure step is described as a required methodology, not as a derived output that collapses to the input data. The central claim therefore remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides no explicit free parameters, axioms, or invented entities beyond standard neural-network weights and SDE components.

pith-pipeline@v0.9.0 · 5659 in / 1029 out tokens · 36994 ms · 2026-05-23T20:51:07.493240+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

48 extracted references · 48 canonical work pages · 2 internal anchors

  1. [1]

    Modeling and implementation of local volatility surfaces in Bayesian framework

    Animoku, Abdulwahab and U˘gur, Ömür and Yolcu-Okur, Yeliz. “Modeling and implementation of local volatility surfaces in Bayesian framework”. In: Computational Management Science 15 (2018), pp. 239–258

    work page 2018

  2. [2]

    Importance sampling for option pricing with feedforward neural networks

    Aleksandar Arandjelovi´c, Thorsten Rheinländer, and Pavel V Shevchenko. “Importance sampling for option pricing with feedforward neural networks”. In: Finance and Stochastics (2024), pp. 1–45

    work page 2024

  3. [3]

    Universal approximation bounds for superpositions of a sigmoidal function

    Andrew R Barron. “Universal approximation bounds for superpositions of a sigmoidal function”. In: IEEE Transactions on Information theory 39.3 (1993), pp. 930–945

    work page 1993

  4. [4]

    Pricing under rough volatility

    Christian Bayer, Peter Friz, and Jim Gatheral. “Pricing under rough volatility”. In: Quantitative Finance 16.6 (2016), pp. 887–904

    work page 2016

  5. [5]

    On deep calibration of (rough) stochastic volatility models

    Christian Bayer et al. “On deep calibration of (rough) stochastic volatility models”. In: arXiv preprint arXiv:1908.08806 (2019)

    work page arXiv 1908

  6. [6]

    Recovering V olatility from Option Prices by Evolutionary Optimization

    Sana Ben Hamida and Rama Cont. “Recovering V olatility from Option Prices by Evolutionary Optimization”. In: The Journal of Computational Finance 8 (Jan. 2005), pp. 43–76. DOI: 10.21314/JCF.2005.130

    work page doi:10.21314/jcf.2005.130 2005

  7. [7]

    Hybrid scheme for Brownian semistationary processes

    Mikkel Bennedsen, Asger Lunde, and Mikko S Pakkanen. “Hybrid scheme for Brownian semistationary processes”. In: Finance and Stochastics 21 (2017), pp. 931–965

    work page 2017

  8. [8]

    The Pricing of Options and Corporate Liabilities

    Fischer Black and Myron Scholes. “The Pricing of Options and Corporate Liabilities”. In: Journal of Political Economy 81.3 (1973), pp. 637–654. ISSN : 00223808, 1537534X. URL: http://www.jstor.org/stable/ 1831029 (visited on 01/17/2024)

    work page 1973

  9. [9]

    Local volatility under rough volatility

    Florian Bourgey et al. “Local volatility under rough volatility”. In: Mathematical Finance 33.4 (2023), pp. 1119–1145

    work page 2023

  10. [10]

    Learning in modular systems

    David M Bradley. Learning in modular systems. Carnegie Mellon University, 2010

    work page 2010

  11. [11]

    Efficient Quasi-Bayesian Estimation of Affine Option Pricing Models Using Risk-Neutral Cumulants

    Riccardo Brignone, Luca Gonzato, and Eva Lütkebohmert. “Efficient Quasi-Bayesian Estimation of Affine Option Pricing Models Using Risk-Neutral Cumulants”. In: Journal of Banking & Finance 148 (Dec. 2022), p. 106745. DOI: 10.1016/j.jbankfin.2022.106745

    work page doi:10.1016/j.jbankfin.2022.106745 2022

  12. [12]

    Deep local volatility

    Marc Chataigner, Stéphane Crépey, and Matthew Dixon. “Deep local volatility”. In: Risks 8.3 (2020), p. 82

    work page 2020

  13. [13]

    FuNV ol: multi-asset implied volatility market simulator using functional principal components and neural SDEs

    Vedant Choudhary, Sebastian Jaimungal, and Maxime Bergeron. “FuNV ol: multi-asset implied volatility market simulator using functional principal components and neural SDEs”. In: Quantitative Finance 24.8 (2024), pp. 1077–1103

    work page 2024

  14. [14]

    Arbitrage-free neural-SDE market models

    Samuel N Cohen, Christoph Reisinger, and Sheng Wang. “Arbitrage-free neural-SDE market models”. In:Applied Mathematical Finance 30.1 (2023), pp. 1–46

    work page 2023

  15. [15]

    Estimating risks of European option books using neural stochastic differential equation market models

    Samuel N Cohen, Christoph Reisinger, and Sheng Wang. “Estimating risks of European option books using neural stochastic differential equation market models”. In: Journal of Computational Finance 26.3 (2022)

    work page 2022

  16. [16]

    Signature-based models: Theory and calibration

    Christa Cuchiero, Guido Gazzani, and Sara Svaluto-Ferro. “Signature-based models: Theory and calibration”. In: SIAM journal on financial mathematics 14.3 (2023), pp. 910–957

    work page 2023

  17. [17]

    A Generative Adversarial Network Approach to Calibration of Local Stochastic V olatility Models

    Christa Cuchiero, Wahid Khosrawi, and Josef Teichmann. “A Generative Adversarial Network Approach to Calibration of Local Stochastic V olatility Models”. In:Risks 8.4 (Sept. 2020), p. 101. ISSN : 2227-9091. DOI: 10.3390/risks8040101. URL: http://dx.doi.org/10.3390/risks8040101

    work page doi:10.3390/risks8040101 2020

  18. [18]

    Global universal approximation of functional input maps on weighted spaces

    Christa Cuchiero, Philipp Schmocker, and Josef Teichmann. Global universal approximation of functional input maps on weighted spaces. 2023. arXiv: 2306.03303 [stat.ML]

    work page arXiv 2023

  19. [19]

    Joint calibration to SPX and VIX options with signature-based models

    Christa Cuchiero et al. “Joint calibration to SPX and VIX options with signature-based models”. In:Mathematical Finance 35.1 (2025), pp. 161–213

    work page 2025

  20. [20]

    Approximation by superpositions of a sigmoidal function

    George V . Cybenko. “Approximation by superpositions of a sigmoidal function”. In:Mathematics of Control, Signals and Systems 2 (1989), pp. 303–314. URL: https://api.semanticscholar.org/CorpusID:3958369

    work page 1989

  21. [21]

    Estimate nothing

    Moritz Duembgen and Leonard Christopher Gordon Rogers. “Estimate nothing”. In: Quantitative Finance 14.12 (2014), pp. 2065–2072. 28 Robust financial calibration: a Bayesian approach for neural SDEs A PREPRINT

    work page 2014

  22. [22]

    Robust pricing and hedging via neural stochastic differential equations

    Patryk Gierjatowicz et al. “Robust pricing and hedging via neural stochastic differential equations”. In: Journal of Computational Finance 26.3 (2022)

    work page 2022

  23. [23]

    Understanding the difficulty of training deep feedforward neural networks

    Xavier Glorot and Yoshua Bengio. “Understanding the difficulty of training deep feedforward neural networks”. In: Proceedings of the thirteenth international conference on artificial intelligence and statistics. JMLR Workshop and Conference Proceedings. 2010, pp. 249–256

    work page 2010

  24. [24]

    Advanced Econometrics: A Bridge to the Literature

    Edward Greenberg. Advanced Econometrics: A Bridge to the Literature. John Wiley & Sons, 1983

    work page 1983

  25. [25]

    Efficient Bayesian inference for multivariate factor stochastic volatility models with leverage

    David Gunawan, Chris Carter, and Robert Kohn. “Efficient Bayesian inference for multivariate factor stochastic volatility models with leverage”. In: arXiv preprint arXiv:1706.03938 (2017)

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  26. [26]

    Robust calibration of financial models using Bayesian estimators

    A. Gupta and C. Reisinger. “Robust calibration of financial models using Bayesian estimators”. In:Journal of Computational Finance 17 (2014), pp. 3–36

    work page 2014

  27. [27]

    Neural joint S&P 500/VIX smile calibration

    Julien Guyon and Scander Mustapha. “Neural joint S&P 500/VIX smile calibration”. In: Available at SSRN 4309576 (2022)

    work page 2022

  28. [28]

    Mimicking the one-dimensional marginal distributions of processes having an Itô differential

    István Gyöngy. “Mimicking the one-dimensional marginal distributions of processes having an Itô differential”. In: Probability theory and related fields71.4 (1986), pp. 501–516

    work page 1986

  29. [29]

    A closed-form solution for options with stochastic volatility with applications to bond and currency options

    Steven L Heston. “A closed-form solution for options with stochastic volatility with applications to bond and currency options”. In: The review of financial studies 6.2 (1993), pp. 327–343

    work page 1993

  30. [30]

    Approximation capabilities of multilayer feedforward networks

    Kurt Hornik. “Approximation capabilities of multilayer feedforward networks”. In:Neural Networks 4.2 (1991), pp. 251–257. ISSN : 0893-6080. DOI: https://doi.org/10.1016/0893-6080(91)90009-T. URL: https: //www.sciencedirect.com/science/article/pii/089360809190009T

    work page doi:10.1016/0893-6080(91)90009-t 1991

  31. [31]

    Modeling Univariate and Multivariate Stochastic V olatility in R with stochvol and factorstochvol

    Darjus Hosszejni and Gregor Kastner. “Modeling Univariate and Multivariate Stochastic V olatility in R with stochvol and factorstochvol”. In: Journal of Statistical Software 100 (2021), pp. 1–34

    work page 2021

  32. [32]

    Efficient Bayesian inference for multivariate factor stochastic volatility models

    Gregor Kastner, Sylvia Frühwirth-Schnatter, and Hedibert Freitas Lopes. “Efficient Bayesian inference for multivariate factor stochastic volatility models”. In: Journal of Computational and Graphical Statistics 26.4 (2017), pp. 905–917

    work page 2017

  33. [33]

    Neural controlled differential equations for irregular time series

    Patrick Kidger et al. “Neural controlled differential equations for irregular time series”. In:Advances in neural information processing systems 33 (2020), pp. 6696–6707

    work page 2020

  34. [34]

    Adam: A Method for Stochastic Optimization

    Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization . 2017. arXiv: 1412.6980 [cs.LG]

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  35. [35]

    Universal approximation property of neural stochastic differential equations

    Anna P Kwossek, David J Prömel, and Josef Teichmann. “Universal approximation property of neural stochastic differential equations”. In: arXiv preprint arXiv:2503.16696 (2025)

    work page arXiv 2025

  36. [36]

    Markovian projections for Itô semimartingales with jumps

    Martin Larsson and Shukun Long. “Markovian projections for Itô semimartingales with jumps”. In:Electronic Communications in Probability 29 (2024), pp. 1–13

    work page 2024

  37. [37]

    Stochastic differential equations and applications

    Xuerong Mao. Stochastic differential equations and applications. Elsevier, 2007

    work page 2007

  38. [38]

    Turbocharging Monte Carlo pricing for the rough Bergomi model

    Ryan McCrickerd and Mikko S. Pakkanen. “Turbocharging Monte Carlo pricing for the rough Bergomi model”. In: Quantitative Finance 18.11 (Apr. 2018), pp. 1877–1886. ISSN : 1469-7696. DOI: 10.1080/14697688.2018. 1459812. URL: http://dx.doi.org/10.1080/14697688.2018.1459812

    work page doi:10.1080/14697688.2018 2018

  39. [39]

    PyTorch: An Imperative Style, High-Performance Deep Learn- ing Library

    Adam Paszke et al. “PyTorch: An Imperative Style, High-Performance Deep Learn- ing Library”. In: Advances in Neural Information Processing Systems 32 . Curran Associates, Inc., 2019, pp. 8024–8035. URL: http://papers.neurips.cc/paper/ 9015-pytorch-an-imperative-style-high-performance-deep-learning-library.pdf

    work page 2019

  40. [40]

    Sig-SDEs model for quantitative finance

    Imanol Perez Arribas, Cristopher Salvi, and Lukasz Szpruch. “Sig-SDEs model for quantitative finance”. In: Proceedings of the First ACM International Conference on AI in Finance. 2020, pp. 1–8

    work page 2020

  41. [41]

    Exponential convergence of Langevin distributions and their discrete approximations

    Gareth O. Roberts and Richard L. Tweedie. “Exponential convergence of Langevin distributions and their discrete approximations”. In: Bernoulli 2 (1996), pp. 341–363.URL: https://api.semanticscholar.org/CorpusID: 18787082

    work page 1996

  42. [42]

    Neural networks for option pricing and hedging: a literature review

    Johannes Ruf and Weiguan Wang. Neural networks for option pricing and hedging: a literature review. 2020. arXiv: 1911.05620 [q-fin.CP]

    work page arXiv 2020

  43. [43]

    Applied Stochastic Differential Equations

    Simo Särkkä and Arno Solin. Applied Stochastic Differential Equations. 2019. URL: https://users.aalto. fi/~ssarkka/pub/sde_book.pdf. 29 Robust financial calibration: a Bayesian approach for neural SDEs A PREPRINT

    work page 2019

  44. [44]

    Complete controllability of continuous-time recurrent neural net- works

    Eduardo Sontag and Héctor Sussmann. “Complete controllability of continuous-time recurrent neural net- works”. In: Systems & Control Letters 30.4 (1997), pp. 177–183. ISSN : 0167-6911. DOI: https://doi. org/10.1016/S0167-6911(97)00002-9. URL: https://www.sciencedirect.com/science/article/ pii/S0167691197000029

    work page doi:10.1016/s0167-6911(97)00002-9 1997

  45. [45]

    Neural Projection Filter: Learning Unknown Dynamics Driven by Noisy Observations

    Yangtianze Tao, Jiayi Kang, and Stephen Yau. “Neural Projection Filter: Learning Unknown Dynamics Driven by Noisy Observations”. In: IEEE Transactions on Neural Networks and Learning Systems (Jan. 2023), pp. 1–15. DOI: 10.1109/TNNLS.2022.3233888

    work page doi:10.1109/tnnls.2022.3233888 2023

  46. [46]

    A Bayesian take on option pricing with Gaussian processes

    Martin Tegner and Stephen Roberts. “A Bayesian take on option pricing with Gaussian processes”. In: arXiv preprint arXiv:2112.03718 (2021)

    work page arXiv 2021

  47. [47]

    Probabilistic machine learning for local volatility

    Martin Tegner and Stephen Roberts. “Probabilistic machine learning for local volatility”. In: Journal of Computa- tional Finance (2021)

    work page 2021

  48. [48]

    MCMC Estimation of Levy Jump Models Using Stock and Option Prices

    Cindy Yu, Haitao Li, and Martin Wells. “MCMC Estimation of Levy Jump Models Using Stock and Option Prices”. In: Mathematical Finance 21 (July 2011). DOI: 10.1111/j.1467-9965.2010.00439.x. 30

    work page doi:10.1111/j.1467-9965.2010.00439.x 2011