acceptodds
Withdrawn from ICLR 2027

PIVOT: Bridging Price and Implied-Volatility Objectives via Differentiable Black–Scholes Inversion

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Modern option-learning systems operate in two coordinates: *price space*, where markets quote and no-arbitrage constraints are most naturally enforced, and *implied volatility (IV) space*, where volatility surfaces are smoothed, regularized, and evaluated. The bottleneck is interface, not approximation: Jäckel's seminal *Let's Be Rational* (LBR) solver already inverts the Black–Scholes price to machine precision efficiently. What is missing is a differentiable layer that preserves LBR in the forward pass and avoids backpropagating through its branch logic. Such a layer must also confront the unavoidable singularity of the inverse map in the low-vega regime, where diverges as . We close this gap with PIVOT, **P**rice–**I**mplied-**V**olatility **O**bjective **T**ranslator. PIVOT keeps the LBR forward pass intact and supplies the backward pass by implicit differentiation through the smooth Black–Scholes/Black-76 price map, with an explicit gating contract: invalid domains return NaN, well-conditioned rows receive the exact gradient, and low-vega rows are attenuated rather than silently regularized. On a single H100, a fused Triton kernel reaches IV/s at machine precision ( max relative error vs. the reference C solver); end-to-end label generation sustains 48.9M/s on synthetic chains and 16.6M/s on SPX OptionMetrics. In a HyperIV-style one-day reproduction on SPX, PIVOT-augmented objectives Pareto-dominate the baselines, reducing held-out price MAE by up to 43.4% and the strongest three-seed gated objective improving price MAE by 38.8% and IV MAE by 21.3% jointly; cross-asset results on RUT, VIX, and NDX show directional price-MAE gains of 40.1%, 24.2%, and 16.7%, while an ungated IV-roundtrip control collapses to a degenerate near-zero surface, confirming the gate as a correctness contract rather than a tuning knob. Code: https://anonymous.4open.science/r/pivot-anon.

Primary area:
learning on time series and dynamical systems
Keywords:
black-scholes, implied-volatility, option-pricing, financial-derivatives, generative-models, implicit-differentiation

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_2aYtOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PIVOT, a differentiable implied-volatility (IV) inversion framework designed to bridge option-price and IV-based objectives in neural network training. The method builds on the existing Jäckel Let's Be Rational solver through the FAST-VOLLIB library for efficient, GPU-batched IV computation, and introduces a custom backward pass based on implicit differentiation, where the gradient of IV with respect to option price is computed as 1/Vega. It also incorporates a low-Vega gating mechanism to mitigate numerical instability during training. The authors validate the framework through synthetic experiments evaluating numerical accuracy, gradient correctness, and computational throughput. They further integrate PIVOT into neural volatility-surface models, primarily HyperIV, and evaluate it on real options datasets. The experiments report improvements in both price and IV prediction errors over the reproduced baselines.

Critical strengths

S1. The paper addresses a relevant implementation challenge in financial machine learning by combining an accurate existing implied-volatility solver with automatic differentiation. The explicit treatment of low-vega cases is also useful, as these cases can create numerical instability during training. This provides potential practical value as a reusable computational component.

S2. The paper evaluates numerical accuracy, gradient correctness, and computational efficiency, and includes experiments on real options datasets. The comparison between gated and ungated training is particularly informative in illustrating the importance of handling unstable gradients. These results provide useful evidence that the proposed implementation can operate within neural network training pipelines.

Critical weaknesses

W1. The paper combines existing components, and its main new technical contribution is not sufficiently clear. Contribution C1 presents a vectorized Jäckel implementation, but Section 3.1 explicitly delegates the forward solve to FAST-VOLLIB (Saqur, 2026,arXiv:2604.27210), a previously published library providing vectorized Jäckel IV inversion with PyTorch, JAX, and Triton GPU backends. The earlier work also reports the same 0.056 ms Triton benchmark on 100,000 options, corresponding to approximately 1.79×10⁹ IV/s. Thus, a headline throughput result in the abstract and Table 1 appears to come from the existing library rather than the proposed differentiable wrapper. The backward pass applies the standard implicit-function-theorem result that the derivative of IV with respect to price is 1/Vega. Differentiation through numerical solvers using this principle is well established in machine learning (e.g., Blondel et al., Efficient and Modular Implicit Differentiation, NeurIPS 2022) and quantitative finance (e.g., Henrard, Adjoint Algorithmic Differentiation: Calibration and Implicit Function Theorem, Journal of Computational Finance), but these relevant works are not discussed.

W2. In Table 2, adding the price auxiliary loss alone reduces price MAE by 34.3% (3.822 to 2.512), while adding the gated IV roundtrip brings the total improvement to 38.8% (2.340). The roundtrip therefore provides a further 6.8% reduction relative to the price-only objective, with only three seeds and non-negligible reported variation (±0.277 and ±0.172). More importantly, the price auxiliary loss uses the standard Black–Scholes formula and does not require differentiable IV inversion. The proposed roundtrip first converts the predicted IV into an option price using Black–Scholes, and then converts that price back into IV using PIVOT. For valid, well-conditioned inputs, converting IV to price and then back to IV should recover the original predicted IV. The corresponding gradients also cancel because the derivative of Black–Scholes price with respect to IV is Vega, while the derivative of IV with respect to price is 1/Vega. Their product is therefore 1, as also noted in the paper's Eq. 17. Consequently, in these cases, the roundtrip loss provides the same training signal as directly comparing predicted IV with market IV, apart from the additional gating and loss weighting. These can also be implemented using the closed-form Black–Scholes Vega without PIVOT. Although the paper additionally handles low-Vega cases through gating and sentinel replacement, it does not compare against a directly gated IV loss with equivalent numerical safeguards. Such a baseline is important for determining whether the differentiable inverse provides a distinct benefit. Without this comparison, the experiments do not convincingly establish that the paper's central proposed component is responsible for the reported prediction improvements.

Questions for authors

Q1 (Regarding W1): Could the authors clarify the specific technical contributions of PIVOT beyond the existing FAST-VOLLIB library and established implicit-differentiation techniques? In particular, which reported efficiency results are attributable to the proposed differentiable wrapper rather than the underlying solver?

Q2 (Regarding W2): Could the authors clarify whether the reported downstream improvements depend on differentiable IV inversion, rather than the additional loss terms and gating mechanism? In particular, how would the proposed approach compare with a directly gated IV loss using equivalent weighting and numerical safeguards, but without PIVOT?

Reviewer_CC3SOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies implied volatility (IV), a central concept in quantitative finance that can be viewed as a normalization of option prices. Its goal is to develop an autograd-compatible version of LBR, a widely used method for computing IV. The forward pass is based on existing work (Saqur, 2026), and the backward pass applies the implicit function theorem together with a masking step that excludes NaN values. The paper presents detailed experiments and compares against IV smoothing with a benchmark method, HyperIV.

Critical strengths
  • The paper addresses a relevant practical problem.
  • The forward pass implements a well-established numerical solver.
  • An anonymized code archive is provided.
Critical weaknesses
  1. The contribution is overstated relative to prior work. Contribution C1 presents the implementation of the forward pass as a key contribution of the paper. However, Section 3.1 attributes this component to FAST-VOLLIB (Saqur, 2026), stating that its construction and numerical analysis "are documented in that reference" and that this paper’s contribution is "the autograd-native wrapper specified in §3.2–3.3." Since the forward solver (Section 3.1) is taken from prior work, the methodological contribution that remains (Sections 3.2 and 3.3) is very thin. It amounts to steps 4–6 of Algorithm 1: standard implicit differentiation combined with a masking convention. Differentiating through an iterative solver via the implicit function theorem is a standard technique, covered for instance in the JAX documentation (Custom derivative rules for JAX-transformable Python functions). A more general framework, of which the proposed method is a special case, is developed in Section 2.1 of: Blondel, Berthet, Cuturi, Frostig, Hoyer, Llinares-López, Pedregosa & Vert, "Efficient and Modular Implicit Differentiation," NeurIPS 2022.

  2. The value of the experiments is unclear. Section 4.1 validates the forward pass, which was introduced in prior work. The loss-based experiments do not clearly isolate the paper’s new contribution. The loss in (6) differentiates only the plain Black–Scholes map and is therefore unrelated to PIVOT or to the paper’s contribution. The loss in (7) is formulated in IV space: after substituting the expression for P~i\tilde{P}_i, it reduces to a direct comparison of implied volatilities, and V^\widehat{\mathcal{V}} relies only on the standard Black–Scholes map.

  3. Presentation and broader relevance. The paper does not make clear what the broader relevance of the method is, or whether there are lessons to be drawn beyond this specific application. Even its relevance to practitioners in the finance industry is not well motivated.

Questions for authors
  1. Could the authors clarify the contribution of the paper relative to Saqur (2026)?
  2. Could the authors present an experiment in which the full PIVOT method (rather than only the Black–Scholes map or the forward solver) is essential?
  3. Could the authors discuss the broader relevance of the method beyond this application?
Reviewer_58xZOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces PIVOT, a differentiable, GPU-compatible operator for recovering implied volatility from option prices using Jäckel’s LBR algorithm. Its main contributions are batched IV inversion for heterogeneous option contracts, implicit differentiation through the Black-Scholes/Black-76 pricing map, and low-vega gating that prevents unstable gradients in ill-conditioned regions. The authors validate PIVOT through synthetic numerical-accuracy and throughput experiments, as well as downstream volatility-surface smoothing experiments on SPX options and cross-asset evaluations on RUT, VIX, and NDX. The results show improved price-IV trade-offs for a HyperIV-style model relative to a local vanilla reproduction, while ungated IV-roundtrip training collapses in low-vega regimes. Experiments with a graph neural operator provide supplementary evidence, although the resulting price-MAE improvements are modest.

Critical strengths

1: The paper tackles the difficulty of integrating implied-volatility inversion into end-to-end, GPU-based option-learning pipelines that combine price-space and IV-space objectives. By pairing Jäckel’s fast inversion algorithm with implicit differentiation, PIVOT provides a batched, autograd-compatible operator with PyTorch and JAX backends and a fused Triton forward kernel; the forward inversion remains Jäckel’s LBR routine, so the contribution is an interface layer rather than a learned replacement of the solver. The numerical-accuracy and throughput experiments support its practical feasibility. This is consequential to my overall assessment because the work turns a meaningful infrastructure limitation into a reusable research primitive, providing practical value beyond a purely conceptual application of established mathematics. 2: The paper connects the inverse-vega gradient to numerical instability as vega approaches zero, and introduces a gating contract-a hard mask, a smooth gate, and NaN semantics-to attenuate the contribution of ill-conditioned regions in the IV-roundtrip objective. The ungated experiments exhibit severe training failures on SPX and RUT, whereas the VIX diagnostic, whose evaluated data contain no observations below the extreme low-vega threshold, does not exhibit the same collapse; its modest IV-MAE degradation reflects under-training rather than a singularity event. This contrast supports the proposed explanation for the observed failures and makes the numerical treatment more convincing than a safeguard evaluated only through aggregate performance gains. This matters to my overall assessment because it addresses a problem-specific difficulty central to the method's practical use, although the evidence does not establish that the same safeguard is universally necessary for differentiable IV methods.

Critical weaknesses
  1. PIVOT combines a GPU-batched Jäckel LBR forward pass with an implicit backward pass, but the evaluation does not compare this design against alternative differentiable inversion approaches, such as a vectorized Newton-based solver with implicit differentiation or a vectorized bisection method. The CPU comparison establishes the benefit of GPU vectorization over a scalar implementation, but does not isolate the advantage of the chosen solver design over plausible GPU-native alternatives. This is consequential to my overall assessment because the paper's main contribution is primarily an engineering and systems integration; without such comparisons, it remains unclear how much of the reported benefit is specific to the proposed design rather than to GPU batching and implicit differentiation more generally.

  2. The robustness and external validity of the downstream improvements remain uncertain. The reported gains vary across assets and model families. In the supplementary GNO experiments, price-MAE improvements are below the authors’ stated 5% practical threshold, while on NDX the best price-auxiliary run reduces price MAE by 16.7% but slightly worsens IV MAE. Moreover, the larger HyperIV improvements are evaluated against a local reproduction whose IV MAE differs substantially from the original HyperIV result; differences in data conventions and training budgets complicate direct comparison with the published baseline. There also appear to be inconsistencies in the reported SPX price-MAE improvement across sections: the abstract and cross-asset summary cite 43.4%, whereas the main experimental table reports 48.1% for the best single run, requiring clarification of which configuration each number refers to.

Questions for authors
  1. The paper establishes that GPU vectorization outperforms the scalar CPU implementation of LBR, but does not isolate the advantage of the chosen solver design. Can the authors provide a comparison against a GPU-vectorized Newton-based IV solver with implicit differentiation, or another suitable differentiable inversion baseline, in terms of accuracy and throughput under comparable settings? This would help establish whether the reported performance gains are specific to PIVOT’s solver design or primarily reflect GPU batching and implicit differentiation.
  2. The abstract reports an “up to 43.4%” reduction in price MAE, while the main experimental results appear to report different values under different configurations or aggregation conventions. Could the authors clarify the exact configuration, seed, and aggregation procedure underlying each reported figure, and reconcile the differences across the abstract, main table, and cross-asset summary?
  3. Statistical robustness of downstream gains. The best price-auxiliary run on NDX reduces price MAE by 16.7% but slightly worsens IV MAE (+1.5%), while the supplementary GNO experiments report price-MAE gains below the paper’s stated 5% practical threshold. Can the authors provide multi-seed variability or uncertainty estimates for the augmented and vanilla models across the evaluated assets?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.