Deep Learning-Enhanced Calibration of the Heston Model: A Unified Framework

Arman Zadgar; Farshid Mehrdoust; Juan E. Trinidad Segovia; Somayeh Fallah

arxiv: 2510.24074 · v2 · submitted 2025-10-28 · 🧮 math.AP · cs.LG

Deep Learning-Enhanced Calibration of the Heston Model: A Unified Framework

Arman Zadgar , Somayeh Fallah , Farshid Mehrdoust , Juan E. Trinidad Segovia This is my paper

Pith reviewed 2026-05-18 03:42 UTC · model grok-4.3

classification 🧮 math.AP cs.LG

keywords deep learningHeston modelcalibrationoption pricingneural networksstochastic volatilityS&P 500financial engineering

0 comments

The pith

A hybrid deep learning approach with two neural networks enhances the calibration of the Heston model for better accuracy and speed.

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

This paper develops a unified framework that uses deep learning to make calibrating the Heston stochastic volatility model faster and more reliable for pricing European options. The method trains one neural network to approximate the option price surface using strike and moneyness, and a second network to correct the model's consistent pricing mistakes. Traditional calibration struggles with high computational cost and local minima in the parameter space, so this hybrid solution seeks to address those issues directly. Tests on actual S&P 500 option market data indicate lower errors and stronger performance both inside and outside the training data compared to standard methods.

Core claim

The central claim is that combining the Heston model with a Price Approximator Network and a Calibration Correction Network produces a calibration procedure that is computationally efficient and yields more accurate parameter estimates than conventional nonlinear optimization techniques, as validated by experiments on real S&P 500 option data.

What carries the argument

The Price Approximator Network (PAN) which approximates the option price surface from strike and moneyness inputs, and the Calibration Correction Network (CCN) which refines the Heston output by correcting systematic pricing errors.

Load-bearing premise

The two supervised feedforward neural networks can be effectively trained to approximate the option price surface and correct systematic pricing errors in the Heston model using strike and moneyness inputs.

What would settle it

Conducting the calibration experiments on a different time period or set of S&P 500 options and observing no advantage in error reduction or convergence speed for the deep learning method would falsify the reported superiority.

Figures

Figures reproduced from arXiv: 2510.24074 by Arman Zadgar, Farshid Mehrdoust, Juan E. Trinidad Segovia, Somayeh Fallah.

**Figure 2.** Figure 2: shows the architecture and data flow within the CCN [PITH_FULL_IMAGE:figures/full_fig_p023_2.png] view at source ↗

**Figure 3.** Figure 3: Results of the calibration of the Heston’s option pricing model on S&P 500 option data. The PAN is first employed to estimate a smooth pricing curve for the last traded prices of S&P 500 options, using in-sample data. The results of this estimation are depicted in [PITH_FULL_IMAGE:figures/full_fig_p026_3.png] view at source ↗

**Figure 4.** Figure 4: Approximation of a line using the Price Approximator Network (PAN) model on S&P 500 option [PITH_FULL_IMAGE:figures/full_fig_p027_4.png] view at source ↗

**Figure 5.** Figure 5: Results of improving the Heston’s option pricing model calibration using deep learning for S&P 500 data. 27 [PITH_FULL_IMAGE:figures/full_fig_p027_5.png] view at source ↗

**Figure 6.** Figure 6: Results of the calibration of the Heston’s option pricing model on S&P 500 Mini option data. Next, the PAN is applied to the S&P 500 Mini data set to learn a smoothed representation of the observed option price surface (see [PITH_FULL_IMAGE:figures/full_fig_p029_6.png] view at source ↗

**Figure 7.** Figure 7: Approximation of a line using the Price Approximator Network (PAN) model on S&P 500 mini option data [PITH_FULL_IMAGE:figures/full_fig_p030_7.png] view at source ↗

**Figure 8.** Figure 8: Results of improving the Heston’s option pricing model calibration using deep learning for S&P 500 mini data. 30 [PITH_FULL_IMAGE:figures/full_fig_p030_8.png] view at source ↗

read the original abstract

The Heston stochastic volatility model is a widely used tool in financial mathematics for pricing European options. However, its calibration remains computationally intensive and sensitive to local minima due to the model's nonlinear structure and high-dimensional parameter space. This paper introduces a hybrid deep learning-based framework that enhances both the computational efficiency and the accuracy of the calibration procedure. The proposed approach integrates two supervised feedforward neural networks: the Price Approximator Network (PAN), which approximates the option price surface based on strike and moneyness inputs, and the Calibration Correction Network (CCN), which refines the Heston model's output by correcting systematic pricing errors. Experimental results on real S\&P 500 option data demonstrate that the deep learning approach outperforms traditional calibration techniques across multiple error metrics, achieving faster convergence and superior generalization in both in-sample and out-of-sample settings. This framework offers a practical and robust solution for real-time financial model calibration.

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.

Desk Editor's Note private letter to a colleague

The paper sketches a two-network DL fix for Heston calibration on real data, but the abstract leaves the core mechanics too vague to confirm the claimed speed and accuracy gains.

read the letter

The one thing to know is that this work pairs a price-approximating network with a correction network to speed up and improve Heston calibration, reporting better error metrics and generalization on S&P 500 options than standard methods. That combination is the concrete new piece, even if deep learning for option surfaces is not brand new overall. The paper does a reasonable job naming the real pain points—slow nonlinear optimization and sensitivity to starting points—and shows the hybrid setup can produce usable results on actual market data without inventing new theory. Credit for sticking to a standard model and testing on live quotes rather than just synthetic cases. The soft spots are mostly about missing specifics rather than outright contradictions. The abstract supplies almost no information on network sizes, loss functions, training splits, or how the baselines were coded and timed, so the outperformance numbers cannot be checked yet. The stress-test point on PAN inputs also lands: if the first network really only receives strike and moneyness and never sees the five Heston parameters, it cannot serve as a surrogate inside the calibration loop, which would remove the main efficiency argument. If the full manuscript adds those parameters to the inputs or explains a different role for PAN, that concern disappears; otherwise it is a load-bearing gap. Minor issues include the lack of any statistical significance tests or ablation runs that would show whether both networks are necessary. This is the sort of paper a quant desk or risk team might want to read for practical ideas on accelerating daily calibration runs. A reader who already works with Heston or similar stochastic-vol models and is open to empirical DL tweaks would get the most out of it. It is worth sending for peer review because the underlying problem is genuine, the data are real, and the idea is straightforward enough that referees could quickly clarify the implementation details and the input question. I would not cite it in its current form but would look again after revisions.

Referee Report

1 major / 1 minor

Summary. The manuscript introduces a hybrid deep learning framework to improve calibration of the Heston stochastic volatility model. It combines two supervised feedforward neural networks—the Price Approximator Network (PAN), which approximates the option price surface from strike and moneyness inputs, and the Calibration Correction Network (CCN), which corrects systematic pricing errors in the Heston output. The authors report that the approach outperforms traditional calibration methods on real S&P 500 option data across error metrics while achieving faster convergence and better in-sample and out-of-sample generalization.

Significance. If the technical implementation is sound, the framework could meaningfully reduce the computational burden of Heston calibration, which remains a practical bottleneck in quantitative finance. A reliable deep-learning surrogate for the model's semi-closed-form pricing formula would be a useful contribution to the literature on efficient model calibration.

major comments (1)

[Abstract] Abstract: The PAN is stated to approximate the option price surface 'based on strike and moneyness inputs.' The Heston price is a function of these quantities together with the five model parameters (κ, θ, σ, ρ, v0), maturity, and spot. Without the model parameters among the network inputs, PAN cannot serve as a surrogate evaluator inside the calibration optimizer; the subsequent CCN correction step then lacks a well-defined quantity to correct. This omission directly undermines the claimed gains in speed and accuracy.

minor comments (1)

[Abstract] Abstract: No details are supplied on network architectures, loss functions, training/validation splits, baseline implementations, or statistical significance tests. These omissions make it impossible to assess whether the reported outperformance is reproducible or statistically supported.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for their thorough review and valuable feedback on our manuscript. We address the major comment below and have made revisions to improve clarity.

read point-by-point responses

Referee: [Abstract] Abstract: The PAN is stated to approximate the option price surface 'based on strike and moneyness inputs.' The Heston price is a function of these quantities together with the five model parameters (κ, θ, σ, ρ, v0), maturity, and spot. Without the model parameters among the network inputs, PAN cannot serve as a surrogate evaluator inside the calibration optimizer; the subsequent CCN correction step then lacks a well-defined quantity to correct. This omission directly undermines the claimed gains in speed and accuracy.

Authors: We appreciate the referee's careful attention to the abstract's wording. The PAN is in fact trained as a surrogate for the full Heston pricing function and takes as inputs strike, moneyness, maturity, spot price, and the five Heston parameters (κ, θ, σ, ρ, v0). The abstract's reference to 'strike and moneyness inputs' was intended as a concise emphasis on the primary market variables while assuming the model parameters are understood as part of the pricing map; this phrasing was imprecise. The PAN is used precisely as a fast evaluator inside the calibration optimizer, and the CCN corrects residual systematic discrepancies between the approximated and true Heston prices. We have revised the abstract to list all inputs explicitly and added a sentence in Section 3.1 clarifying the network's input vector and its role as a parameter-dependent surrogate. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected in the hybrid DL calibration framework

full rationale

The paper proposes an empirical hybrid method using two supervised feedforward networks (PAN for price surface approximation from strike/moneyness and CCN for error correction) trained on market data, with performance claims presented as experimental outcomes on S&P 500 options rather than any first-principles derivation or prediction. No load-bearing steps reduce by construction to inputs, self-citations, or fitted parameters renamed as results; the framework is validated against external traditional calibration benchmarks and remains self-contained.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 2 invented entities

The central claim depends on the capacity of supervised neural networks to learn the nonlinear mapping from market inputs to corrected Heston prices without requiring traditional numerical optimization at inference time.

free parameters (1)

Neural network weights and biases
Parameters of PAN and CCN are fitted during supervised training on option data to achieve the reported error reductions.

axioms (1)

domain assumption Feedforward neural networks can approximate the complex pricing functions arising from the Heston stochastic volatility model
This assumption justifies replacing or augmenting traditional calibration routines with the two networks.

invented entities (2)

Price Approximator Network (PAN) no independent evidence
purpose: Approximates the option price surface from strike and moneyness inputs
New component introduced to accelerate price evaluation during calibration.
Calibration Correction Network (CCN) no independent evidence
purpose: Refines Heston model outputs by correcting systematic pricing errors
New component introduced to improve accuracy beyond standard calibration.

pith-pipeline@v0.9.0 · 5698 in / 1584 out tokens · 49364 ms · 2026-05-18T03:42:56.824831+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

31 extracted references · 31 canonical work pages · 2 internal anchors

[1]

L.B.G.Andersen, EfficientsimulationoftheHestonstochasticvolatility model, Journal of Computational Finance 11 (3) (2007) 1–42

work page 2007
[2]

Bayer, B

C. Bayer, B. Horvath, A. Muguruza, B. Stemper, M. Tomas, On deep calibration of (rough) stochastic volatility models,arXiv preprint arXiv:1908.08806 (2019)

work page arXiv 1908
[3]

Becker, P

S. Becker, P. Cheridito, A. Jentzen, Pricing and hedging financial deriva- tives using neural networks, SIAM Journal on Financial Mathematics 10 (2) (2019) 747–777

work page 2019
[4]

Black, M

F. Black, M. Scholes, The pricing of options and corporate liabilities, Journal of Political Economy 81 (3) (1973) 637–654

work page 1973
[5]

Broadie, O

M. Broadie, O. Kaya, Exact simulation of stochastic volatility and other affine jump diffusion processes, Operations Research 54 (2) (2006) 217– 231

work page 2006
[6]

Buehler, L

H. Buehler, L. Gonon, J. Teichmann, B. Wood, Deep hedging, Quanti- tative Finance 19 (8) (2019) 1271–1291

work page 2019
[7]

S. Chen, P. Glasserman, Estimating the option price surface with a regression-based approach, Mathematical Finance 20 (2) (2010) 267– 296

work page 2010
[8]

Christoffersen, S

P. Christoffersen, S. L. Heston, K. Jacobs, Option valuation with con- ditional skewness in the Heston model, Journal of Financial Economics 98 (3) (2010) 380–404

work page 2010
[9]

J. C. Cox, J. E. Ingersoll, S. A. Ross, A theory of the term structure of interest rates, Econometrica 53 (2) (1985) 385–407

work page 1985
[10]

Duffie, J

D. Duffie, J. Pan, K. Singleton, Transform analysis and asset pricing for affine jump–diffusions, Econometrica 68 (6) (2000) 1343–1376

work page 2000
[11]

Deeply Learning Derivatives

R. Ferguson, A. Green, Deeply learning derivatives,arXiv preprint arXiv:1809.02233 (2018)

work page internal anchor Pith review Pith/arXiv arXiv 2018
[12]

Forde, E

M. Forde, E. Jacquier, A tutorial on the Heston model, Quantitative Finance 11 (11) (2011) 1613–1624. 32

work page 2011
[13]

Gatheral, The volatility surface: a practitioner’s guide, John Wiley & Sons, 2011

J. Gatheral, The volatility surface: a practitioner’s guide, John Wiley & Sons, 2011

work page 2011
[14]

Glorot, Y

X. Glorot, Y. Bengio, Understanding the difficulty of training deep feed- forwardneuralnetworks, in: ProceedingsoftheThirteenthInternational ConferenceonArtificialIntelligenceandStatistics, JMLRWorkshopand Conference Proceedings, 2010, pp. 249–256

work page 2010
[15]

Gneiting, F

T. Gneiting, F. Balabdaoui, A. E. Raftery, Probabilistic forecasts, cal- ibration and sharpness, Journal of the Royal Statistical Society: Series B (Statistical Methodology) 69 (2) (2007) 243–268

work page 2007
[16]

Goodfellow, Y

I. Goodfellow, Y. Bengio, A. Courville, Y. Bengio, Deep learning, vol. 1, no. 2, MIT Press, Cambridge, 2016

work page 2016
[17]

S. Gu, B. Kelly, D. Xiu, Empirical asset pricing via machine learning, The Review of Financial Studies 33 (5) (2020) 2223–2273

work page 2020
[18]

K. He, X. Zhang, S. Ren, J. Sun, Delving deep into rectifiers: surpassing human-level performance on ImageNet classification, in: Proceedings of the IEEE International Conference on Computer Vision, 2015, pp. 1026– 1034

work page 2015
[19]

J.B.Heaton, N.G.Polson, J.H.Witte, Deeplearninginfinance, Annual Review of Financial Economics 9 (2017) 145–165

work page 2017
[20]

S.L.Heston, Aclosed–formsolutionforoptionswithstochasticvolatility withapplicationstobondandcurrencyoptions, TheReviewofFinancial Studies 6 (2) (1993) 327–343

work page 1993
[21]

Horvath, A

B. Horvath, A. Muguruza, M. Tomas, Deep learning volatility: a hybrid approach to pricing and calibration, Quantitative Finance 21 (4) (2021) 657–675

work page 2021
[22]

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

work page 1994
[23]

D. P. Kingma, J. Ba, Adam: A method for stochastic optimization, arXiv preprintarXiv:1412.6980 (2014). 33

work page internal anchor Pith review Pith/arXiv arXiv 2014
[24]

X. Li, L. Xu, Z. Wang, Machine learning for financial engineering: A review, Journal of Computational Finance 25 (4) (2021) 47–96

work page 2021
[25]

S. Liu, A. Borovykh, L. A. Grzelak, C. W. Oosterlee, A neural network- based framework for financial model calibration, Journal of Mathematics in Industry 9 (1) (2019) 9

work page 2019
[26]

V. Nair, G. E. Hinton, Rectified linear units improve restricted Boltz- mann machines, in: Proceedings of the 27th International Conference on Machine Learning (ICML), 2010, pp. 807–814

work page 2010
[27]

Paszke, S

A. Paszke, S. Gross, F. Massa, et al., PyTorch: an imperative style, high-performance deep learning library, Advances in Neural Information Processing Systems 32 (2019) 8024–8035

work page 2019
[28]

Platon, V

R. Platon, V. R. Dehkordi, J. Martel, Hourly prediction of a building’s electricity consumption using case-based reasoning, artificial neural net- worksandprincipalcomponentanalysis, EnergyandBuildings92(2015) 10–18

work page 2015
[29]

J. Ruf, W. Wang, Neural networks for option pricing and hedging: a literature review,arXiv preprintarXiv:1911.05620 (2019)

work page arXiv 1911
[30]

Sirignano, R

J. Sirignano, R. Cont, Universal features of price formation in financial markets: perspectives from deep learning, Quantitative Finance 19 (9) (2019) 1449–1459

work page 2019
[31]

E. M. Stein, J. C. Stein, Stock price distributions with stochastic volatil- ity: an analytic approach, The Review of Financial Studies 4 (4) (1991) 727–752. 34

work page 1991

[1] [1]

L.B.G.Andersen, EfficientsimulationoftheHestonstochasticvolatility model, Journal of Computational Finance 11 (3) (2007) 1–42

work page 2007

[2] [2]

Bayer, B

C. Bayer, B. Horvath, A. Muguruza, B. Stemper, M. Tomas, On deep calibration of (rough) stochastic volatility models,arXiv preprint arXiv:1908.08806 (2019)

work page arXiv 1908

[3] [3]

Becker, P

S. Becker, P. Cheridito, A. Jentzen, Pricing and hedging financial deriva- tives using neural networks, SIAM Journal on Financial Mathematics 10 (2) (2019) 747–777

work page 2019

[4] [4]

Black, M

F. Black, M. Scholes, The pricing of options and corporate liabilities, Journal of Political Economy 81 (3) (1973) 637–654

work page 1973

[5] [5]

Broadie, O

M. Broadie, O. Kaya, Exact simulation of stochastic volatility and other affine jump diffusion processes, Operations Research 54 (2) (2006) 217– 231

work page 2006

[6] [6]

Buehler, L

H. Buehler, L. Gonon, J. Teichmann, B. Wood, Deep hedging, Quanti- tative Finance 19 (8) (2019) 1271–1291

work page 2019

[7] [7]

S. Chen, P. Glasserman, Estimating the option price surface with a regression-based approach, Mathematical Finance 20 (2) (2010) 267– 296

work page 2010

[8] [8]

Christoffersen, S

P. Christoffersen, S. L. Heston, K. Jacobs, Option valuation with con- ditional skewness in the Heston model, Journal of Financial Economics 98 (3) (2010) 380–404

work page 2010

[9] [9]

J. C. Cox, J. E. Ingersoll, S. A. Ross, A theory of the term structure of interest rates, Econometrica 53 (2) (1985) 385–407

work page 1985

[10] [10]

Duffie, J

D. Duffie, J. Pan, K. Singleton, Transform analysis and asset pricing for affine jump–diffusions, Econometrica 68 (6) (2000) 1343–1376

work page 2000

[11] [11]

Deeply Learning Derivatives

R. Ferguson, A. Green, Deeply learning derivatives,arXiv preprint arXiv:1809.02233 (2018)

work page internal anchor Pith review Pith/arXiv arXiv 2018

[12] [12]

Forde, E

M. Forde, E. Jacquier, A tutorial on the Heston model, Quantitative Finance 11 (11) (2011) 1613–1624. 32

work page 2011

[13] [13]

Gatheral, The volatility surface: a practitioner’s guide, John Wiley & Sons, 2011

J. Gatheral, The volatility surface: a practitioner’s guide, John Wiley & Sons, 2011

work page 2011

[14] [14]

Glorot, Y

X. Glorot, Y. Bengio, Understanding the difficulty of training deep feed- forwardneuralnetworks, in: ProceedingsoftheThirteenthInternational ConferenceonArtificialIntelligenceandStatistics, JMLRWorkshopand Conference Proceedings, 2010, pp. 249–256

work page 2010

[15] [15]

Gneiting, F

T. Gneiting, F. Balabdaoui, A. E. Raftery, Probabilistic forecasts, cal- ibration and sharpness, Journal of the Royal Statistical Society: Series B (Statistical Methodology) 69 (2) (2007) 243–268

work page 2007

[16] [16]

Goodfellow, Y

I. Goodfellow, Y. Bengio, A. Courville, Y. Bengio, Deep learning, vol. 1, no. 2, MIT Press, Cambridge, 2016

work page 2016

[17] [17]

S. Gu, B. Kelly, D. Xiu, Empirical asset pricing via machine learning, The Review of Financial Studies 33 (5) (2020) 2223–2273

work page 2020

[18] [18]

K. He, X. Zhang, S. Ren, J. Sun, Delving deep into rectifiers: surpassing human-level performance on ImageNet classification, in: Proceedings of the IEEE International Conference on Computer Vision, 2015, pp. 1026– 1034

work page 2015

[19] [19]

J.B.Heaton, N.G.Polson, J.H.Witte, Deeplearninginfinance, Annual Review of Financial Economics 9 (2017) 145–165

work page 2017

[20] [20]

S.L.Heston, Aclosed–formsolutionforoptionswithstochasticvolatility withapplicationstobondandcurrencyoptions, TheReviewofFinancial Studies 6 (2) (1993) 327–343

work page 1993

[21] [21]

Horvath, A

B. Horvath, A. Muguruza, M. Tomas, Deep learning volatility: a hybrid approach to pricing and calibration, Quantitative Finance 21 (4) (2021) 657–675

work page 2021

[22] [22]

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

work page 1994

[23] [23]

D. P. Kingma, J. Ba, Adam: A method for stochastic optimization, arXiv preprintarXiv:1412.6980 (2014). 33

work page internal anchor Pith review Pith/arXiv arXiv 2014

[24] [24]

X. Li, L. Xu, Z. Wang, Machine learning for financial engineering: A review, Journal of Computational Finance 25 (4) (2021) 47–96

work page 2021

[25] [25]

S. Liu, A. Borovykh, L. A. Grzelak, C. W. Oosterlee, A neural network- based framework for financial model calibration, Journal of Mathematics in Industry 9 (1) (2019) 9

work page 2019

[26] [26]

V. Nair, G. E. Hinton, Rectified linear units improve restricted Boltz- mann machines, in: Proceedings of the 27th International Conference on Machine Learning (ICML), 2010, pp. 807–814

work page 2010

[27] [27]

Paszke, S

A. Paszke, S. Gross, F. Massa, et al., PyTorch: an imperative style, high-performance deep learning library, Advances in Neural Information Processing Systems 32 (2019) 8024–8035

work page 2019

[28] [28]

Platon, V

R. Platon, V. R. Dehkordi, J. Martel, Hourly prediction of a building’s electricity consumption using case-based reasoning, artificial neural net- worksandprincipalcomponentanalysis, EnergyandBuildings92(2015) 10–18

work page 2015

[29] [29]

J. Ruf, W. Wang, Neural networks for option pricing and hedging: a literature review,arXiv preprintarXiv:1911.05620 (2019)

work page arXiv 1911

[30] [30]

Sirignano, R

J. Sirignano, R. Cont, Universal features of price formation in financial markets: perspectives from deep learning, Quantitative Finance 19 (9) (2019) 1449–1459

work page 2019

[31] [31]

E. M. Stein, J. C. Stein, Stock price distributions with stochastic volatil- ity: an analytic approach, The Review of Financial Studies 4 (4) (1991) 727–752. 34

work page 1991