acceptodds
Under review as a conference paper at ICLR 2027

Arbitrage-Free Implied Volatility SurfaceReconstruction via Additive Martingales

13 views

Abstract

This work focuses on reconstructing a continuous implied volatility (IV) surface from a small number of option quotes, a fundamental task for option pricing and risk management. We propose Additive Martingale (ADDMART), a few-shot IV surface reconstruction framework that satisfies static no-arbitrage conditions by construction. Rather than predicting IVs directly, ADDMART maps sparse quotes to non-negative increment densities of an exponential additive martingale, from which valid risk-neutral marginals and IVs are recovered. By incorporating the no-arbitrage structure directly into its architecture, rather than enforcing it through regularization in the learning objective, ADDMART produces IV surfaces that satisfy static no-arbitrage conditions throughout the domain. Extensive experiments on ten real-option datasets demonstrate that ADDMART achieves competitive reconstruction accuracy with no material static arbitrage violations, while directly enabling the estimation of VaR, CVaR, and crash probabilities.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Implied Volatility Surface, Arbitrage-Free Learning, Amortized Inference, Set Encoder, Financial Machine Learning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_dhBdOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes AddMart (Additive Martingale), a framework for reconstructing implied volatility (IV) surfaces from a small number of option quotes while structurally guaranteeing static no-arbitrage conditions.

Existing few-shot IV reconstruction methods typically predict IV surfaces directly and enforce no-arbitrage constraints through regularization penalties, which may introduce an accuracy–regularization trade-off without providing strict guarantees.

To address this limitation, AddMart learns the increment distributions of an exponential additive martingale conditioned on sparse option quotes, rather than directly predicting IV. These distributions are combined through convolution to construct valid risk-neutral marginal distributions, from which option prices and IV surfaces are recovered. The martingale-based decoder guarantees butterfly and calendar no-arbitrage conditions without requiring additional arbitrage penalties.

Experiments on ten real-world options datasets demonstrate that AddMart achieves competitive IV reconstruction accuracy compared to existing few-shot methods while exhibiting no material static arbitrage violations. The framework also enables downstream risk analysis, such as CVaR estimation, by directly providing valid risk-neutral distributions.

Critical strengths

1. Clear motivation and problem formulation

  • The paper provides a clear and well-motivated problem formulation. It effectively explains the importance of few-shot IV surface reconstruction and the challenges associated with satisfying static no-arbitrage constraints.

2. Evaluation

  • The paper conducts experiments on a large-scale real-world options dataset, comprising approximately 40 million quotes across 70,000 IV surfaces. The use of extensive market data covering multiple assets and sampling frequencies strengthens the empirical evaluation.

3. Ablation Study

  • The ablation study effectively supports the paper’s central claim by demonstrating that replacing exact linear convolution with cropped or circular convolution introduces calendar arbitrage violations (Table 3a). This provides empirical evidence for the importance of the proposed decoder design.

4. Structural No-Arbitrage Guarantee

  • Unlike penalty-based approaches, AddMart structurally guarantees butterfly and calendar no-arbitrage conditions through its martingale-based architecture. This ensures valid risk-neutral distributions and enables downstream risk measures, such as VaR and CVaR, to be computed without post-hoc clipping or renormalization (Figure 3 and Table 8). This is a particularly compelling practical advantage of the proposed method.
Critical weaknesses

1. Novelty

  • AddMart appears to combine existing ideas from HyperIV [1] and Ning et al. [2] and Lin & Liu [3]. However, it is unclear what the key methodological novelty is beyond combining these approaches.
  • Despite their close methodological relevance, Ning et al. [2] and Lin & Liu [3] are not included as baselines. Furthermore, the improvements in IV reconstruction accuracy over HyperIV are relatively modest. Thus, neither the methodological novelty nor the empirical advantages over closely related approaches are sufficiently established.
  • Recommendation: A clearer demonstration of either the methodological novelty beyond existing approaches or meaningful empirical advantages over prior work would strengthen the contribution. In particular, a comparison with closely related stochastic process-based methods under the same few-shot setting would be valuable.

2. Methodology

  • Rather than predicting IV directly, AddMart learns increment distributions and reconstructs IV through convolution, option pricing, and IV inversion. While this enables structural no-arbitrage guarantees, it also introduces additional modeling complexity and potential sources of estimation and numerical errors. The paper does not sufficiently discuss these trade-offs.
  • Although the model generates increment distributions η\eta, it is trained solely using an IV reconstruction loss. Since different increment distributions may produce similar IV surfaces, the learned distributions may not be identifiable or accurately represent the underlying market process. This is particularly relevant given the paper’s emphasis on downstream risk estimation.
  • Recommendation: A more detailed discussion of the potential costs of the density-based parameterization, as well as the reliability and interpretability of the learned increment distributions, would be helpful.

3. Theoretical Guarantee

  • Theorem 1 establishes static no-arbitrage under the stated assumptions, but it does not guarantee that the learned increment distributions accurately represent the underlying market process or that the reconstructed IV surfaces are close to the true market surfaces.
  • Moreover, the theoretical relationship between errors in the estimated increment distributions and the resulting IV reconstruction errors remains unclear. Thus, the theoretical guarantees primarily concern structural validity rather than estimation accuracy.
  • Recommendation: Clarifying the scope of the theoretical guarantees and, if possible, analyzing how distribution estimation errors affect the final IV reconstruction would strengthen the theoretical contribution.

4. Experiments

  • The improvement in IV reconstruction accuracy over HyperIV is relatively modest. In some evaluation settings, HyperIV even performs slightly better. Although AddMart provides structural no-arbitrage guarantees, its empirical advantage in reconstruction accuracy remains limited.
  • The experiments rely on nine systematically selected reference quotes based on predefined maturities and deltas. While this provides a controlled few-shot setting, it may not adequately represent realistic sparse-quote environments, where observations can be irregularly distributed or certain maturities may be missing. The existing ablation on the number of reference quotes does not fully address this issue.
  • The use of proprietary market data also limits independent reproducibility, although the associated licensing restrictions are understandable.
  • Recommendation: Additional evaluation under more realistic sparse-quote settings, such as irregularly sampled quotes and missing maturities, would help establish the model’s generalization ability.

Reference

[1] Yang, Y., Chen, W., Shu, C., & Hospedales, T. (2025, May). Hyperiv: Real-time implied volatility smoothing. In Forty-second International Conference on Machine Learning.

[2] Ning, B., Jaimungal, S., Zhang, X., & Bergeron, M. (2023). Arbitrage-free implied volatility surface generation with variational autoencoders. SIAM Journal on Financial Mathematics, 14(4), 1004-1027.

[3] Lin, J., & Liu, G. (2024, November). Neural Term Structure of Additive Process for Option Pricing. In Proceedings of the 5th ACM International Conference on AI in Finance (pp. 695-702).

Questions for authors

1. Methodological Novelty

  • How does AddMart differ technically from combining HyperIV’s few-shot reconstruction framework with the neural stochastic process modeling approaches of [2] and [3]?
  • What are the key methodological contributions beyond this combination?

2. Comparison with Related Work

  • Given the methodological similarities, why were [2] and [3] not included as baselines?
  • Could the authors clarify whether these approaches can be adapted to the same few-shot reconstruction setting?

3. Density-Based Parameterization

  • What are the potential costs of predicting increment distributions rather than IV directly, particularly in terms of estimation difficulty, numerical errors, and model expressivity?
  • How do these trade-offs affect reconstruction accuracy?

4. Identifiability and Theoretical Guarantees

  • Since the model is trained solely with an IV reconstruction loss, how can the reliability of the learned increment distributions be assessed when different distributions may produce similar IV surfaces?
  • Beyond structural no-arbitrage, what can be said about the accuracy of the learned distributions and their impact on downstream risk estimates?

5. Generalization to Irregular Sparse Quotes

  • How robust is AddMart when reference quotes are irregularly distributed or certain maturities are missing, rather than selected from predefined maturities and deltas?
  • Does the model maintain competitive reconstruction accuracy under such conditions?
Reviewer_H81UOpenReview ↗
Rating
3
Confidence
4(+1)
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper investigates whether a full implied volatility surface can be reconstructed from a small number of option quotes while satisfying static no-arbitrage conditions.

The proposed method, AddMart, predicts the distributions of increments over successive time intervals using nine reference option quotes. These distributions are combined through convolution to construct risk-neutral distributions across maturities, which are subsequently used to compute option prices and implied volatilities. Once trained, AddMart does not require separate optimization for each new volatility surface. The predicted distributions can also be used to estimate tail-risk measures, including Value-at-Risk (VaR) and Conditional Value-at-Risk (CVaR).

The main contribution is the enforcement of no-arbitrage conditions by construction. Under the paper's European-equivalent and forward-normalized pricing assumptions, and within the modeled maturity interval, the authors prove that call prices are:

  • Decreasing in strike;
  • Convex in strike;
  • Non-decreasing in maturity for a fixed normalized strike.

The authors evaluate AddMart on seven daily datasets and three intraday datasets, using out-of-time testing protocols. The method generally achieves competitive reconstruction accuracy, while the audited evaluation grid exhibits no butterfly or calendar arbitrage violations beyond numerical tolerance.

Additional convolution ablation studies investigate the importance of the proposed construction, and the CVaR experiment examines whether the learned risk-neutral distributions provide useful information for tail-risk estimation.

Critical strengths

1. The no-arbitrage guarantee is supported by both theoretical analysis and experimental evidence.

Section 4 and Appendix C.1 establish why the full convolution construction produces call prices that are convex in normalized strike and non-decreasing in maturity at a fixed normalized strike. The ablation results in Table 3 provide empirical support for the importance of this construction: cropped and circular convolution achieve similar implied volatility (IV) reconstruction errors but introduce calendar arbitrage violations. This consistency between the theoretical guarantees and experimental findings is the strongest reason I consider the proposed method a substantive contribution.

2. The empirical evaluation demonstrates that enforcing no-arbitrage does not substantially compromise reconstruction accuracy.

Across seven daily and three intraday datasets, AddMart achieves competitive reconstruction errors using the same nine reference quotes as the amortized baselines, while exhibiting no arbitrage violations beyond numerical tolerance on the audited evaluation grid. Furthermore, Appendix G.1 reports lower option-implied Conditional Value-at-Risk (CVaR) errors in five of six asset–horizon settings compared with a reference distribution fitted using all available quotes. These findings suggest that the constructed arbitrage-free marginal distributions have practical value beyond merely satisfying static no-arbitrage constraints.

Critical weaknesses

1. The scope and practical applicability of the no-arbitrage guarantee require clearer qualification.

Theorem 1 establishes no-arbitrage properties for European-equivalent call prices at a fixed forward-normalized strike, whereas the main empirical datasets contain listed American-style equity and ETF options. Although the paper explains its use of out-of-the-money (OTM) quotes and parity-implied forwards, it does not quantitatively assess the sensitivity of this conversion to dividends, early exercise, or forward estimation errors. This distinction is important because an unqualified claim of static no-arbitrage may be interpreted as a guarantee for the original listed contracts rather than for the European-equivalent representation considered in the theorem. The paper would benefit from explicitly stating these assumptions and limitations in both the abstract and conclusion, together with an asset- and maturity-dependent sensitivity analysis.

2. The empirical trade-off between structural guarantees, reconstruction accuracy, and computational efficiency is not uniformly favorable.

Appendix G.5 identifies several limitations of the learned representation. In particular, AddMart loses substantial skew beyond one year for SPY, produces a flatter implied volatility smile than the market quotes in the 554-day AAPL example, and exhibits a failure case for short-maturity VIX options. Although these observations do not invalidate the main results at shorter maturities, they indicate that the representation has limitations in certain market regimes and maturity ranges.

Moreover, within the daily reliable region, HyperIV achieves a slightly lower asset-averaged reconstruction error (1.27 versus 1.29), while Table 10 reports substantially lower single-surface inference latency (1.19 ms versus 7.60 ms). These comparisons suggest that the practical benefit of the structural no-arbitrage guarantee should be evaluated more carefully against its accuracy and computational costs. A maturity-bucket analysis, particularly for long-dated options (LEAPS), and a more precise characterization of the regimes in which AddMart offers a favorable trade-off would strengthen the empirical conclusions.

Questions for authors

1. Baseline selection and hyperparameter settings.

Why were VolNP and VolNP-SRCL excluded from the nine-quote comparison? Appendix F.2 states that GNO was trained without its originally proposed arbitrage penalties and that HyperIV's three penalty weights were all set to 1. Were these settings selected through validation? Could the authors report the reconstruction errors and arbitrage violation rates of GNO and HyperIV using their original or validation-tuned penalty settings?

2. Sensitivity of the distributional evaluation.

In Appendix G.1, how sensitive are the reported CVaR errors and method rankings to the smoothing penalty used to fit the reference distribution? What procedure was used to repair invalid baseline densities, and does the CVaR ranking change under alternative repair methods? Could the authors also report CDF errors or tail-probability errors against the same reference distribution to assess distributional accuracy independently of validity?

3. Contribution of cross-maturity martingale coupling.

Could the authors add an ablation that retains the same encoder and density parameterization but predicts valid marginal distributions independently at each maturity, without convolution-based cross-maturity coupling? How would this model compare with AddMart in terms of implied volatility reconstruction error, calendar arbitrage violations, and single-surface inference latency? This comparison would help isolate the contribution and computational cost of the martingale construction.

Reviewer_3G5POpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The idea of this paper is to reconstruct IV surface satisfying static no-arbitrage conditions by construction through a martingale process. The mathematics is solid and the experiments are large with 10 datasets, including European options, American options, intraday data and daily data.

Critical strengths
  1. The model maps sparse quotes to increment laws of an exponential additive martingale and then decodes them via full-support convolution rather than directly predicting an unconstrained surface. This hard guarantee is practically valuable for risk management.

  2. Lots of experiments with different datasets have been done to evaluate the proposed framework and compared with the existing work.

  3. Its permutation-invariant set encoder and parameter-free martingale decoder provide a structured approach to sparse-data reconstruction.

Critical weaknesses
  1. In section 2 the authors mentioned some neural related papers i.e. Lin & Liu (2024) and then declare ‘ADDMART shares this additive martingale backbone but is quote conditioned and amortized few shot via a permutation invariant set encoder’. However none of them has been used as a baseline.

  2. The baseline comparisons are not fair. In the original HyperIV paper k\in[-1.5, 1.5] however this paper uses k\in[-1.5,0.5]. For GNO replication, the authors declare ‘It is trained only with query-IV MSE; no calendar or butterfly penalty is added.’ In Widemann et al’s paper, penalties were considered. In addition, the author mentioned ‘GNO (Li et al., 2020; Wiedemann et al., 2025). Our GNO baseline uses the published Operator Deep Smoothing GNOVolSmoother architecture’. In fact the term GNOVolSmoother was not mentioned in either paper.

  3. The datasets contains American and ETF options, however the theoretical framework, pricing formulas, and risk-neutral density recovery assume European-style options. In addition, Breeden–Litzenberger identity used to interpret option-price curvature as a risk-neutral density does not directly apply to American option prices.

Questions for authors
  1. How does ADDMART compare against Lin & Liu (2024)?
  2. On page 23, authors mentioned the data-cleaning process. How many contracts were removed? If those are not removed what will Table 1 look like?
  3. Why was penalty removed when replicating GNO?
  4. Why was k narrowed down when replicating HyperIV?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.