arxiv: 2605.06604 · v1 · submitted 2026-05-07 · 💱 q-fin.CP · stat.ML

Recognition: unknown

A Geometry-Aware Residual Correction of Hagan's SABR Implied Volatility Formula

Adil Reghai, Alex Lipton, G\'erard Biau, Lama Tarsissi

Pith reviewed 2026-05-08 02:59 UTC · model grok-4.3

classification 💱 q-fin.CP stat.ML

keywords SABR modelimplied volatilityHagan approximationresidual correctionneural networkgeometry-aware featureshybrid methodstochastic volatility

0 comments

The pith

A neural network trained on the residual error of Hagan's SABR formula, supplied with geometry-aware inputs from the model's SDEs, improves implied volatility accuracy and robustness.

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

The paper establishes that Hagan's analytical approximation for SABR implied volatility can be strengthened by a neural network that learns only the difference from the true value. Geometry-aware variables drawn from the SABR stochastic differential equations are added to the network inputs to supply structural information about the dynamics. The resulting hybrid retains the original formula as its backbone rather than replacing it, so that the network focuses on higher-order terms absent from the asymptotic expansion. A reader would care because the method promises more reliable prices and calibrations in real trading settings while preserving interpretability and computational lightness.

Core claim

The central claim is that augmenting the input space of a neural network with geometry-aware features derived from the SABR SDEs and training the network to predict the residual between exact implied volatility and Hagan's closed-form expression produces a correction that is both more accurate and more robust than either the standalone analytical formula or a conventional neural network trained directly on volatility values.

What carries the argument

The geometry-aware residual correction, in which variables that reflect intrinsic properties of the SABR stochastic differential equations are concatenated to the network inputs while the training target is defined as the difference from Hagan's approximation.

If this is right

The hybrid model delivers higher accuracy and robustness than Hagan's formula alone or than standard neural networks across realistic and stressed parameter domains.
The correction remains lightweight and structurally consistent with the SABR model, preserving the analytical backbone.
The framework supports real-time pricing and calibration tasks in practical trading environments.

Where Pith is reading between the lines

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

The same residual-correction pattern could be applied to other asymptotic expansions that leave systematic higher-order errors.
Training directly on observed market quotes instead of simulated paths might further align the correction with live conditions.
Systematic testing on parameter values deliberately chosen to violate the geometry assumptions would map the practical limits of the approach.

Load-bearing premise

The geometry-aware variables extracted from the SABR stochastic equations encode the higher-order effects missing from the asymptotic expansion well enough for the residual network to generalize reliably outside the training parameter ranges.

What would settle it

If, on parameter sets drawn from stressed regimes lying outside the training distribution, the hybrid model's absolute error in implied volatility exceeds the error of the uncorrected Hagan formula, the claim of improved robustness would be falsified.

Figures

Figures reproduced from arXiv: 2605.06604 by Adil Reghai, Alex Lipton, G\'erard Biau, Lama Tarsissi.

**Figure 1.** Figure 1: SABR implied volatility smile: comparison between the Hagan formula and Monte view at source ↗

**Figure 2.** Figure 2: Minimalist pipeline of the proposed hybrid SABR approximation. Monte Carlo (with view at source ↗

**Figure 3.** Figure 3: Geometric transformations underlying the SABR model. The forward variable is view at source ↗

**Figure 4.** Figure 4: While the training loss rapidly decreases to approximately view at source ↗

**Figure 4.** Figure 4: NDN learning dynamics. Training and validation loss evolution. The validation curve stabilizes near 0.43, indicating severe overfitting and a persistent generalization gap. 19 view at source ↗

**Figure 5.** Figure 5: NDN correlation diagnostics. Correlation between neural predictions and Monte Carlo implied volatilities across ITM, ATM, and OTM regions. Significant dispersion is observed, particularly near ATM. 20 view at source ↗

**Figure 6.** Figure 6: NDN under stress. Model behavior under stressed SABR parameters. The architecture fails to preserve skew and convexity in extreme regimes. 21 view at source ↗

**Figure 7.** Figure 7: NDN smile slices across maturities. Comparison of neural and Monte Carlo implied volatilities for multiple maturities. Long-maturity convexity is progressively underestimated. 22 view at source ↗

**Figure 8.** Figure 8: GeoNN learning dynamics. Training and validation loss evolution. The validation loss stabilizes near 0.08, indicating improved generalization relative to NDN view at source ↗

**Figure 9.** Figure 9: GeoNN correlation diagnostics. Correlation between neural predictions and Monte Carlo implied volatilities across ITM, ATM, and OTM regions. Dispersion is significantly reduced relative to NDN. 23 view at source ↗

**Figure 10.** Figure 10: GeoNN under stress. Model behavior under stressed SABR parameters. Skew and central curvature are more robustly preserved compared to the naive model. 24 view at source ↗

**Figure 11.** Figure 11: GeoNN smile slices across maturities. Comparison of neural and Monte Carlo implied volatilities for multiple maturities. Curvature is well captured for short and intermediate maturities, with mild wing underestimation at longer horizons. 25 view at source ↗

**Figure 12.** Figure 12: ResNN learning dynamics. Training and validation loss evolution. The validation loss stabilizes near 0.113, indicating stable convergence and reduced overfitting relative to NDN view at source ↗

**Figure 13.** Figure 13: ResNN correlation diagnostics. Correlation between neural predictions and Monte Carlo implied volatilities across ITM, ATM, and OTM regions. Dispersion is reduced relative to NDN but remains higher than GeoNN in certain regions. 26 view at source ↗

**Figure 14.** Figure 14: ResNN under stress. Model behavior under stressed SABR parameters. Central regions remain stable, though residual wing bias may persist in extreme configurations. 27 view at source ↗

**Figure 15.** Figure 15: ResNN smile slices across maturities. Comparison of neural and Monte Carlo implied volatilities for multiple maturities. Asymptotic scaling is preserved, though analytical bias is not fully eliminated in long-maturity wings. 28 view at source ↗

**Figure 16.** Figure 16: GeoResNN learning dynamics. Training and validation loss evolution. The validation loss stabilizes near 0.072, indicating the strongest generalization among all architectures view at source ↗

**Figure 17.** Figure 17: GeoResNN correlation diagnostics. Correlation between neural predictions and Monte Carlo implied volatilities across ITM, ATM, and OTM regions. Dispersion is minimal and alignment with Monte Carlo is strongest among all models. 29 view at source ↗

**Figure 18.** Figure 18: GeoResNN under stress. Model behavior under stressed SABR parameters. Curvature and skew remain stable even in high vol-of-vol regimes. 30 view at source ↗

**Figure 19.** Figure 19: GeoResNN smile slices across maturities. Comparison of neural and Monte Carlo implied volatilities for multiple maturities. The model preserves curvature and asymptotic scaling across the full maturity range. 31 view at source ↗

read the original abstract

This paper proposes a hybrid methodology to improve the approximation of SABR (Stochastic Alpha Beta Rho) implied volatility by combining analytical structure with machine learning. The approach augments the neural-network input representation with geometric features derived from the stochastic differential equations of the SABR model. Unlike approaches that fully replace analytical formulas with black-box models, the proposed framework preserves the analytical backbone of the model. The hybridization operates along two complementary dimensions. First, geometry-aware variables reflecting intrinsic properties of the SABR dynamics are used as structured inputs to the network. Second, the neural network is trained to learn the residual error relative to Hagan's closed-form approximation rather than implied volatility directly. The resulting model acts as a structured residual correction to the analytical formula, retaining interpretability while capturing higher-order effects that are not included in the asymptotic expansion. Numerical experiments conducted over realistic parameter domains, as well as stressed environments, show that the method improves accuracy and robustness compared with both analytical approximations and standard neural-network approaches. Because the correction remains lightweight and structurally consistent with the underlying model, the framework is well suited for real-time pricing and calibration in practical trading environments.

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

This paper adds geometry-aware SABR SDE features to a residual neural net that corrects Hagan's formula, but the experimental claims lack the concrete metrics and details needed to judge real gains.

read the letter

The core idea is straightforward: keep Hagan's closed-form SABR implied volatility as the base, then train a network to predict only the residual error while feeding it extra inputs drawn from geometric properties of the model's SDEs. This is not a full replacement of the analytics, which is the main practical point in its favor for trading systems that need speed and some interpretability left intact. The geometry features are presented as a way to inject structure that generic networks miss, and the authors argue this helps with higher-order effects in both normal and stressed parameter ranges. That framing is reasonable and avoids the usual black-box pitfalls in finance ML work. The claim of better accuracy and robustness over pure analytics and standard nets is the part that would matter most to practitioners if it holds. The main weakness is that the abstract gives no numbers, no error tables, no description of how the geometry variables are actually computed, and no information on training versus test domains or out-of-distribution checks. Without those, it is impossible to tell whether the residual correction generalizes or whether the reported improvements are tied to specific data choices. The stress-test concern about feature sufficiency is fair on the current evidence. This is aimed at quants who already run SABR for options pricing and risk and are open to small, structured ML tweaks rather than wholesale model changes. A reader working on hybrid methods might extract the residual-plus-geometry pattern as a template worth testing. The paper deserves peer review because the underlying approach is coherent and relevant to deployed models, even though the current version needs the missing experimental substance to be convincing.

Referee Report

3 major / 1 minor

Summary. This paper presents a hybrid methodology for approximating the implied volatility in the SABR model. It retains Hagan's closed-form asymptotic expansion as the base and employs a neural network to learn and correct the residual error. The network inputs are augmented with geometry-aware features extracted from the SABR stochastic differential equations. The authors claim that this structured residual correction improves accuracy and robustness relative to both the standalone analytical formula and conventional neural network approaches, as evidenced by experiments across realistic and stressed parameter domains. The framework is positioned as suitable for real-time pricing and calibration due to its lightweight nature and retention of analytical structure.

Significance. Should the empirical results be substantiated with detailed metrics and reproducible protocols, this contribution would be significant for the field of computational finance. It exemplifies a principled way to integrate domain knowledge from stochastic processes with machine learning, avoiding the pitfalls of fully black-box models. By focusing on residual correction informed by geometric properties of the dynamics, the approach could lead to more reliable and interpretable enhancements to existing pricing formulas, with potential benefits for practitioners in volatility modeling and risk management.

major comments (3)

Abstract: The assertion that 'Numerical experiments conducted over realistic parameter domains, as well as stressed environments, show that the method improves accuracy and robustness' lacks any accompanying quantitative data, such as error values, tables, or figures with comparisons. This is a load-bearing issue for the central claim, as the improvement cannot be assessed without metrics or baseline results.
Section 3 (Methodology): The geometry-aware variables are stated to reflect 'intrinsic properties of the SABR dynamics' and are used as structured inputs, but the manuscript does not provide their explicit definitions or how they are derived from the SABR SDEs. This omission prevents verification of whether these features adequately address the higher-order effects absent in Hagan's expansion.
Numerical Experiments section: There are no details on the training/test splits, the specific computation of geometry features, error bars, data exclusion rules, or out-of-distribution testing in stressed environments. These are necessary to support the generalization claims.

minor comments (1)

Abstract: The description of the hybrid framework is clear, but adding a brief note on the dimensionality or typical ranges of the geometry-aware inputs would aid reader intuition without altering the technical content.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed feedback on our manuscript. The comments identify key areas where additional clarity and transparency will strengthen the presentation of our hybrid SABR approximation method. We have revised the manuscript accordingly and provide point-by-point responses below.

read point-by-point responses

Referee: Abstract: The assertion that 'Numerical experiments conducted over realistic parameter domains, as well as stressed environments, show that the method improves accuracy and robustness' lacks any accompanying quantitative data, such as error values, tables, or figures with comparisons. This is a load-bearing issue for the central claim, as the improvement cannot be assessed without metrics or baseline results.

Authors: We agree that the abstract claim would be more compelling with explicit quantitative support. Although the Numerical Experiments section contains the relevant error metrics, tables, and baseline comparisons, we have revised the abstract to summarize key quantitative findings, including average error reductions relative to Hagan's formula and standard neural-network baselines across both realistic and stressed parameter regimes, with references to the supporting tables and figures. revision: yes
Referee: Section 3 (Methodology): The geometry-aware variables are stated to reflect 'intrinsic properties of the SABR dynamics' and are used as structured inputs, but the manuscript does not provide their explicit definitions or how they are derived from the SABR SDEs. This omission prevents verification of whether these features adequately address the higher-order effects absent in Hagan's expansion.

Authors: We appreciate the referee highlighting this omission. The geometry-aware inputs are derived from the SABR SDEs via the local volatility expression, the correlation-driven drift terms, and selected higher-order corrections arising from Ito's lemma applied to the forward price and volatility processes. In the revised Section 3 we now supply the explicit mathematical definitions, the step-by-step derivation from the SDEs, and the precise mapping from model parameters and state variables to each network input. This addition enables direct verification of how the features target effects beyond the leading-order terms retained in Hagan's expansion. revision: yes
Referee: Numerical Experiments section: There are no details on the training/test splits, the specific computation of geometry features, error bars, data exclusion rules, or out-of-distribution testing in stressed environments. These are necessary to support the generalization claims.

Authors: We acknowledge that greater experimental detail is required for reproducibility and to substantiate the generalization claims. The revised Numerical Experiments section now includes: (i) the precise training/test split ratios and stratified sampling procedure over the parameter domains; (ii) explicit formulas together with pseudocode for computing each geometry-aware feature from the SABR state and parameters; (iii) error bars obtained from five independent training runs with distinct random seeds; (iv) data exclusion rules (e.g., removal of samples yielding negative forwards or numerically unstable volatility paths); and (v) a dedicated out-of-distribution evaluation protocol using stressed parameter sets (high volatility-of-volatility, extreme correlations, and low forward prices) withheld from training. These additions directly address the concerns raised. revision: yes

Circularity Check

0 steps flagged

No significant circularity in hybrid analytical-ML residual correction

full rationale

The paper defines a hybrid method that takes Hagan's external closed-form SABR implied-volatility approximation as its analytical backbone and trains a neural network to predict only the residual correction, with additional inputs consisting of geometry-aware variables extracted directly from the SABR SDEs. No equation or claim reduces the target output to a quantity defined by the fitted parameters themselves, nor does any load-bearing step rely on a self-citation chain, uniqueness theorem imported from the authors' prior work, or an ansatz smuggled via citation. The improvement claim rests on numerical experiments conducted over stated parameter domains; these experiments are independent empirical checks rather than tautological predictions. The derivation chain therefore remains non-circular.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The approach assumes the standard SABR stochastic differential equations hold and that their geometric properties provide useful structured features; the neural network parameters are fitted to data but the claim does not depend on any ad-hoc invented entities.

free parameters (1)

Neural network weights and hyperparameters
The network is trained to predict residuals, so its parameters are fitted to data.

axioms (1)

domain assumption SABR dynamics are described by the standard stochastic differential equations
Geometry-aware inputs are derived directly from these SDEs.

pith-pipeline@v0.9.0 · 5514 in / 1269 out tokens · 43309 ms · 2026-05-08T02:59:03.734552+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

15 extracted references · 4 canonical work pages · 1 internal anchor

[1]

and Kumar, Deep and Lesniewski, Andrew S

Hagan, Patrick S. and Kumar, Deep and Lesniewski, Andrew S. and Woodward, Diana E. , title =. Wilmott Magazine , year =
[2]

, title =

McGhee, William A. , title =. Journal of Computational Finance , year =
[3]

2021 , address =

Thorin, Henrik , title =. 2021 , address =

2021
[4]

2021 , address =

Stuijt, Hugo , title =. 2021 , address =

2021
[5]

Incorporating Prior Financial Domain Knowledge into Neural Networks for Implied Volatility Surface Prediction , journal =

Feng, Lingjie and Wu, Xintao and Arik, Sercan. Incorporating Prior Financial Domain Knowledge into Neural Networks for Implied Volatility Surface Prediction , journal =. 2019 , volume =. 1904.12834 , archivePrefix =

work page arXiv 2019
[6]

Fine-Tune Your Smile: Correction to Hagan et al

Ob. Fine-Tune Your Smile: Correction to Hagan et al. , journal =. 2008 , pages =

2008
[7]

arXiv preprint , year =

Paulot, Louis , title =. arXiv preprint , year =. 0906.0658 , archivePrefix =

work page arXiv
[8]

Risk Magazine , year =

Antonov, Alexandre and Konikov, Michael and Spector, Michael , title =. Risk Magazine , year =
[9]

CBS Research Working Paper , year =

Lund, Brian , title =. CBS Research Working Paper , year =
[10]

arXiv preprint , year =

Muguruza, Aitor and Manso, Jose and Serna, German , title =. arXiv preprint , year =. 1901.09647 , archivePrefix =

work page arXiv 1901
[11]

Adam: A Method for Stochastic Optimization

Kingma, Diederik P. and Ba, Jimmy , title =. arXiv preprint , year =. 1412.6980 , archivePrefix =

work page internal anchor Pith review arXiv
[12]

, title =

do Carmo, Manfredo P. , title =
[13]

Rosenberg, Steven , title =
[14]

2020 , doi =

Reghai, Adil and Kettani, Othmane , title =. 2020 , doi =

2020
[15]

Analysis, Geometry, and Modeling in Finance: Advanced Methods in Option Pricing , publisher =

Henry-Labord. Analysis, Geometry, and Modeling in Finance: Advanced Methods in Option Pricing , publisher =