Paper: arXiv 2409.06551

Authors: Christa Cuchiero, Eva Flonner, Kevin Kurt

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.

Complexity vs Empirical Score

  • Math Complexity: 9.0/10
  • Empirical Rigor: 4.0/10
  • Quadrant: Lab Rats — high math complexity, limited empirical testing

Why this score: The paper involves advanced Bayesian methods, neural SDEs, universal approximation theorems, and measure change theory, indicating high mathematical density. However, the excerpt presents a proof-of-concept framework with sensitivity analysis but lacks detailed empirical results, code, or specific backtesting metrics.

Research Flowchart

  flowchart TD
  A["Research Goal: Robust Calibration of Financial Models"] --> B["Methodology: Bayesian Framework"]
  B --> C["Inputs: Options & Historical Data"]
  C --> D["Process: Measure Change & Neural SDE"]
  D --> E["Computation: Langevin Sampling"]
  E --> F["Outcomes: Robust Implied Vol. Bounds & Universal Approximation"]

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes a Bayesian calibration framework for neural stochastic differential equations that jointly calibrates to option and time series data.

Source sentence
the goal of the current paper is to unite the efforts described above and provide a Bayesian calibration framework based on neural SDEs.

Assumes 6

  • Risk‑free rate is set to zero for simplicity. market
    Source sentence
    From now on we assume for notational simplicity that r = 0, unless specified otherwise.
  • Option market price equals model price plus normally distributed noise. statistical
    Source sentence
    the basis point calibration error for the ith option for the true, underlying θ is supposed to be normally distributed as 104 S0 (C(i)mod(θ) −C(i)mkt) ∼N(0, δ2 i ).
  • Prior distribution is placed on neural network weights. model
    Source sentence
    Together with a prior distribution p(θ) on the weights θ of the neural networks (σ, bV , σV ) appearing in the neural SDE system (1),(2), the desired posterior distribution will then be obtained explicitly.
  • Hedging instruments are square‑integrable martingales under the risk‑neutral measure. model
    Source sentence
    consider p traded hedging instruments (Zt)t∈[0,T ] (including in particular the traded assets Sθ) which are square integrable martingales under the risk neutral measure Q(θ) and take values in Rp.
Show 2 more
  • Synthetic historical data follows a geometric Brownian motion with drift 0.05 and diffusion 0.3. data
    Source sentence
    The input data for the historical measure is one trajectory from the logarithm of a geometric Brownian motion with drift 0.05 and diffusion 0.3. dYt = (0.05 −0.5 · 0.32)dt + 0.3dWt, Y0 = 0 simulated via the Euler‑Maruyama scheme.
  • Option data are Black‑Scholes call prices for specified strikes and maturities. data
    Source sentence
    The input for the Q‑measure are call option prices obtained from the Black Scholes formula for risk‑free rate r = 0.025, 21 strikes from 0.8 to 1.2 and maturities 16, 32, 48, 64, 80 and 96.

Data

  • Synthetic GBM trajectory synthetic unknown
    Source sentence
    The input data for the historical measure is one trajectory from the logarithm of a geometric Brownian motion with drift 0.05 and diffusion 0.3.
  • Black‑Scholes call option prices · Black‑Scholes formula options public
    Source sentence
    The input for the Q‑measure are call option prices obtained from the Black Scholes formula for risk‑free rate r = 0.025, 21 strikes from 0.8 to 1.2 and maturities 16, 32, 48, 64, 80 and 96.
  • S&P 500 implied volatility data (1990s) options public
    1990s · S&P 500
    Source sentence
    ... is calibrated to real call options considering empirical S&P 500 implied volatility data with 10 strikes and 4 maturities.

Costs

Not applicable The paper reports no trading or execution result.

Finds 5

  • The posterior distribution provides robust bounds on implied volatility surfaces and exotic option prices. pricing-calibration
    Source sentence
    the model naturally produces a distribution on prices for exotic options, by computing the variance‑reduced Monte Carlo prices of these exotics using the calibrated neural SDE model and Langevin sampling.
  • Algorithm converged around epoch 13, after which draws are from the posterior. performance
    Source sentence
    The trace plots reveal that the algorithm converged around epoch 13. This means that after this epoch we are dealing with approximate draws from the posterior distribution.
  • Calibration is not very sensitive to the bid‑ask spread parameter δ. statistical
    Source sentence
    The results reveal that the calibration to option data, as well as time series data, is not very sensitive with respect to changes in δ.
  • Neural SDE learns the statistical properties of the Heston model, including mean‑reversion in spot variance. theoretical
    Source sentence
    Our comparison aims to evaluate to which extent the neural SDE is capable of replicating the typical characteristics of the Heston model such as mean‑reversion in the spot variance.
  • Smaller prior variance σprior yields more stable calibration results. statistical
    Source sentence
    Numerical experiments have shown that smaller values of σprior yield more stable results.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 34 related papers). Coverage is partial, so this is neither replication nor refutation.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (64)

  1. Approximation by superpositions of a sigmoidal function (1989) — link
  2. Neural joint S&P 500/VIX smile calibration (2022)
  3. Estimating risks of European option books using neural stochastic differential equation market models (2022)
  4. A Generative Adversarial Network Approach to Calibration of Local Stochastic Volatility Models (2020) — link
  5. Joint calibration to SPX and VIX options with signature-based models — → 2301.13235
  6. Robust pricing and hedging via neural stochastic differential equations (2022)
  7. Learning Partially Known Stochastic Dynamics with Empirical PAC Bayes (2020) — link
  8. Applied Stochastic Differential Equations (2019) — link
and 56 more
  1. A closed-form solution for options with stochastic volatility with applications to bond and currency options (1993)
  2. Bayesian Learning via Stochastic Gradient Langevin Dynamics (2011)
  3. MCMC Estimation of Levy Jump Models Using Stock and Option Prices (2011) — link
  4. Robust calibration of financial models using Bayesian estimators (2014)
  5. Understanding the difficulty of training deep feedforward neural networks (2010)
  6. Recovering Volatility from Option Prices by Evolutionary Optimization (2005) — link
  7. Bayesian Learning via Stochastic Gradient Langevin Dynamics (2011)
  8. Preconditioned Stochastic Gradient Langevin Dynamics for Deep Neural Networks (2015) — link
  9. Computation of Deterministic Volatility Surfaces (1999) — link
  10. Neural SDEs as Infinite-Dimensional GANs (2021) — link
  11. Neural controlled differential equations for irregular time series (2020)
  12. Applied Stochastic Differential Equations (2019)
  13. Adaptive Rejection Sampling for Gibbs Sampling (1992)
  14. Gibbs sampling for Bayesian non‐conjugate and hierarchical models by using auxiliary variables (1999)
  15. Lévy Processes and Infinitely Divisible Distributions (1999)
  16. Advanced Econometrics: A Bridge to the Literature (1983)
  17. Learning in modular systems (2010)
  18. Bayesian Methods in Applied Econometrics, or, Why Econometrics Should Always and Everywhere Be Bayesian (2007)
  19. Approximation capabilities of multilayer feedforward networks (1991) — link
  20. Adam: A Method for Stochastic Optimization (2017) — link
  21. Complete controllability of continuous-time recurrent neural networks (1997) — link
  22. Deep neural networks, generic universal interpolation, and controlled ODEs (2019) — link
  23. Bayesian analysis in inverse problems (1999) — link
  24. Sig-SDEs model for quantitative finance (2020)
  25. Signature-based models: Theory and calibration (2023)
  26. Arbitrage-free neural-SDE market models (2023)
  27. FuNVol: multi-asset implied volatility market simulator using functional principal components and neural SDEs (2024)
  28. Neural networks for option pricing and hedging: a literature review (2020) — link
  29. On deep calibration of (rough) stochastic volatility models (2019) — link
  30. Estimate nothing (2014)
  31. Approximation and Estimation Bounds for Artificial Neural Networks (1994) — link
  32. A Bayesian take on option pricing with Gaussian processes (2021) — link
  33. The Pricing of Options and Corporate Liabilities (1973)
  34. Efficient Quasi-Bayesian Estimation of Affine Option Pricing Models Using Risk-Neutral Cumulants (2022) — link
  35. PyTorch: An Imperative Style, High-Performance Deep Learning Library (2019) — link
  36. Global universal approximation of functional input maps on weighted spaces — → 2306.03303
  37. Deep local volatility (2020)
  38. Exponential convergence of Langevin distributions and their discrete approximations (1996) — link
  39. Robust financial calibration: a Bayesian approach for neural SDEs (2024) — link
  40. Pricing under rough volatility (2016)
  41. Neural Projection Filter: Learning Unknown Dynamics Driven by Noisy Observations (2023) — link
  42. Importance sampling for option pricing with feedforward neural networks (2024)
  43. Hybrid scheme for Brownian semistationary processes (2017)
  44. Turbocharging Monte Carlo pricing for the rough Bergomi model (2018) — link
  45. Stochastic differential equations and applications (2007)
  46. Probabilistic machine learning for local volatility (2021)
  47. Modeling and implementation of local volatility surfaces in Bayesian framework (2018)
  48. Efficient Bayesian inference for multivariate factor stochastic volatility models (2017)
  49. Efficient Bayesian inference for multivariate factor stochastic volatility models with leverage (2017) — link
  50. Modeling Univariate and Multivariate Stochastic Volatility in R with stochvol and factorstochvol (2021)
  51. Universal approximation property of neural stochastic differential equations — → 2503.16696
  52. Mimicking the one-dimensional marginal distributions of processes having an It ^o differential (1986)
  53. Markovian projections for It ^o semimartingales with jumps (2024)
  54. Local volatility under rough volatility (2023)
  55. Robust financial calibration: a Bayesian approach for neural SDEs (2024) — link
  56. Universal approximation bounds for superpositions of a sigmoidal function (1993)

3 of 64 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.