acceptodds
Under review as a conference paper at ICLR 2027

Finance-Informed Operator Learning for Option Pricing with Quantum-Compatible Realizations

Abstract

Pricing European options under local volatility requires solving a partial differential equation whose coefficients change with every recalibration, and practitioners repeatedly need not only prices but also their sensitivities to the underlying across spot–time surfaces for many strike and maturity configurations. Neural surrogates are an efficient alternative, approximating the solution operator directly so that model evaluation replaces repeated numerical solves. Near expiry, however, the solution loses regularity, and a network trained on prices alone can misstate the curvature needed for hedging or return prices outside the no-arbitrage bounds. We propose a finance-informed Deep Operator Network (FI-DeepONet) that decomposes the solution operator into a closed-form reference price and an additive learned correction. The framework is finance-informed in that financial structure enters the representation and the output map rather than the loss: the reference uses the strike-line integrated variance—the time integral of the squared local volatility evaluated at the strike, which captures the leading near-expiry curvature singularity, and a smooth monotone admissibility layer enforces pointwise no-arbitrage price bounds, at the cost of piecewise rather than global smoothness in the spot variable. A deterministic post-processing step then corrects the predicted grid after inference and is evaluated separately from the learned operator. We establish a short-maturity estimate and differentiated asymptotics for the exact correction, together with identities relating errors in the learned correction to pricing and sensitivity errors, as well as PDE-residual errors. On in-distribution (ID) tests, the model reduces the global relative price error by nearly an order of magnitude over vanilla and physics-informed DeepONet baselines across independent initializations. On out-of-distribution (OOD) tests, generalization is moderate when the input parameters lie outside the training range but within the same parametric class of local-volatility functions. We also test the trained model on real market data using index option quotes and find that it remains accurate without market-specific network retraining, though it does not improve on the analytic formula given the same volatility input. We further give a quantum-compatible realization of the trained operator, in which selected linear maps are either compiled exactly or approximated within a restricted diagonal–orthogonal family before supervised adaptation.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
neural operators, DeepONet, scientific machine learning, finance-informed learning, physics-informed learning, option pricing, local volatility, quantitative finance, quantum neural operators

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_QUVaOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies operator learning for pricing call options under a local volatility model. The model learns a rescaled offset RθR_\theta relative to a carrier CrefC^{\mathrm{ref}}, defined as the Black–Scholes price with volatility frozen at the strike. It then applies a nonlinear monotone transformation (equation 2.5) to map the result into the no-arbitrage interval. In addition to introducing the methodology, the paper presents theoretical results on the asymptotics of the correction term, numerical experiments, and a section on Quantum-Compatible Realizations.

Critical strengths
  • Operator learning for option pricing is an important topic.
  • Incorporating prior financial structure into the model is a sensible approach.
Critical weaknesses
  1. Writing.

The paper is extremely dense and reads more like an audit log than a research paper. It introduces unexplained terms throughout, which makes it very hard to read, to understand, and to properly assess the contribution.

  • For example, Section 4 on numerical experiments begins as follows: Principal comparisons use pre-PP outputs (FI after FAL), FD Greeks and PDE diagnostics, and five-seed means; sample standard deviations are reported in Appendix C.2. Mechanism ablations use pre-FAL outputs; the PDE-loss factorial uses automatic-differentiation (AD) diagnostics. The text then jumps directly to Section 4.1 on benchmarks and metrics.

  • A second example, from the bottom of p. 7, is reproduced below. Similar issues occur throughout the paper. Fidelity and adaptation. QO reproduces parent predictions to numerical precision in idealized evaluation. At the selected C94 coverage, QDO roughly doubles global price error; near-strike price p95 changes little, Delta p95 worsens, and Gamma p95 improves (Table 2A,B). The selected C94 instance fails the fixed ID/OOD accuracy-preservation criteria (Appendix D.5). These synthetic evaluations use AD Greeks under a protocol separate from the principal FD study. In the three-seed matched C94 study, updating only substituted maps and their biases achieves almost all frozen-to-full error recovery (Figure 3a). Across paired C49–C99 configurations, mean price errors remain near 1.8×10−31.8 \times 10^{-3} without a monotone coverage trend (Figure 3b; Appendix D.6). The term C94 is used and discussed in detail without ever being properly introduced. The passage then refers to a three-seed matched C94 study, which is not introduced either. As a result, it is very hard to follow what is being evaluated.

  • Similarly, Section 4.1 refers to parameters β\beta, γ\gamma, and σ0\sigma_0, which are only properly introduced in Appendix B.2. Only there does the reader learn that the local volatility functions were generated from a parametric family with these three parameters (see equation (B.1)).

  • Section 5 does not explain the role of the quantum component. It refers to QO and QDO without defining these abbreviations, and it does not explain why a quantum-compatible version is of interest for option pricing.

  1. Methodological inconsistencies.

The methodological choices made in the paper do not support the claim that an operator (a,K,T,r)↦V(a, K, T, r) \mapsto V is learned (as stated in lines 139–140, Section 2).

  • The paper formally uses a DeepONet, but all data are generated from a three-parameter volatility model (B.1). The network also receives these parameters (σ0,γ,β)(\sigma_0, \gamma, \beta) as inputs. This makes the volatility input itself redundant and reduces the task to a standard feedforward regression problem.
  • The local volatility functions in (B.1) depend on KK. This is inconsistent with how local volatility models are used and with the operator learning motivation: a single local volatility function should price options consistently across a range of strikes.
  • Along the same lines, the carrier appears to be computed from the generating formula rather than from the sampled volatility functions (lines 152–156).
Questions for authors
  • Could the authors evaluate the model without providing (σ0,γ,β)(\sigma_0, \gamma, \beta) as inputs?
  • Could the authors evaluate the model with the carrier computed from the sampled volatility functions rather than from the generating formula?
  • What is the relevance of the theoretical results? Since they are asymptotic statements, how do they support the paper’s central claims?
5 comments
AuthorsOpenReview ↗

Response to Reviewer QUVa (Part 1/2)

Presentation, definitions, and the quantum component (CW1).

We will reorganize the discussion from the pricing problem and observed volatility function to the analytic carrier, the learned correction, and finally the numerical evaluation. In particular, the generator parameters (σ0,β,γ\sigma _{0},\beta,\gamma) and the strike-dependent benchmark formula will be introduced before the experiments, rather than only in an appendix. We will distinguish predictions before and after the Financial Admissibility Layer (FAL) from outputs following the optional post-processing (PP) step; finite-difference (FD) Greeks will be separated from automatic-differentiation (AD) diagnostics. The quantum section will first identify its representation problem and then define QO (idealized exact factorization), QDO (restricted diagonal–orthogonal substitution with adaptation), and C94 (the selected roughly 94% eligible-operation-coverage configuration) before showing their results.

Does the submitted benchmark test a genuine pricing operator? (CW2).

The reviewer correctly identifies a substantive modeling distinction: Eq. (B.1) of the submitted benchmark uses a coefficient generator that depends on the contract's strike K, and the network has access to the three generating parameters. Thus the original data describe a controlled parameterized, per-contract pricing benchmark; the observed parameter-OOD performance alone does not establish that a single, contract-independent volatility field supports pricing across independent K,T queries.

Our supplementary experiment targets precisely this distinction. Each field a(S,t) is generated before choosing contract terms, held fixed when K and T vary, and supplied to the network as 225 sampled local-variance values together with the contract query. On a pre-specified subset of 42 independent moderate/strong fields (three near-the-money strikes by four maturities per field, 504 queries, T ≥ 0.25 years), the comparison is:

Table 4: Fixed-field, contract-independent pricing test (price MAE).

  • Vanilla DeepONet: 0.13030
  • Residual MLP: 0.04310
  • Analytic carrier: 0.05557
  • FI residual model (retrained, pre-FAL): 0.01238
  • Forward Dupire PDE: 0.003831

The separately trained FI model reduces the carrier's MAE by 77.7%; the three-seed, field-paired 95% interval for MAEFI−MAEcarrier\mathrm{MAE} _{\mathrm{FI}}-\mathrm{MAE} _{\mathrm{carrier}} is [-0.0639,-0.0252]. The observed-field Dupire solver is more accurate, but the FI residual model improves on the matched neural and analytic baselines. The evidence specifically establishes function-observation-based, independent-contract pricing on the tested field class.

AuthorsOpenReview ↗

Response to Reviewer QUVa (Part 2/2)

Can the network price without (σ0,γ,β)(\sigma _{0},\gamma,\beta)? (Q1).

Yes. We performed a matched 2 × 2 ablation within the submitted parameterized family. The first factor is whether the network receives the explicit generating parameters; the second is whether the carrier uses the generating formula or is reconstructed from sampled volatility. Each of the four configurations was separately trained with three paired seeds. The same 32 in-distribution (ID) and 320 same-family parameter-OOD price surfaces were evaluated:

Table 5: Direct input and carrier ablation; entries are price RelL2 in percent.

Network inputsCarrier constructionID (%)OOD (%)
Parameters includedGenerating formula0.09910.1539
Parameters omittedGenerating formula0.10610.1668
Parameters includedSampled volatility0.09880.1445
Parameters omittedSampled volatility0.10130.1593

ID: 32 surfaces; OOD: 320 parameter-shifted surfaces from the same family. Post-FAL/pre-PP outputs. Every OOD configuration uses exact terminal payoff at τ = 0, leaving positive-maturity predictions unchanged. Values are percentages, so 0.1013% = 0.001013 in dimensionless RelL2.

Without either generator inputs or the formula-based carrier, the model retains ID Price RelL2 0.1013%, close to 0.0991% for the fully generator-informed comparison; the matched same-family OOD values are 0.1593% and 0.1539%. Thus the explicit parameter shortcut is not required for accurate pricing on the tested family and sensor layout.

Can the carrier be reconstructed from sampled volatility? (Q2).

Yes. For this ablation, we use 231 sampled volatilities on a 21 × 11 spatial–calendar-time grid, distinct from the 225 sampled variances used in the contract-independent test. Interpolation supplies the strike-line volatility, whose integrated square determines the Black–Scholes carrier:

=\int _{0}^{\tau}\widehat{\sigma} _{\mathrm{sens}}^{\\,2}(K,T-u)\\,du.$$ The sample-derived and formula-derived carriers differ by global relative L² of only 1.11 × 10⁻⁵ on 32 ID surfaces and 9.69 × 10⁻⁶ on 320 same-family OOD surfaces. With a sampled carrier and otherwise parameter-informed network, ID Price RelL2 is 0.0988%; removing explicit network parameters as well yields 0.1013%. Hence the analytic carrier can be constructed from the observed field, without accessing the generating formula, at the tested sensor resolution. **How the asymptotic theory supports the method (Q3).** Propositions 3.1–3.2 analyze the *exact pricing correction* $W=V-C^{\mathrm{ref}}$. Proposition 3.1 gives a conditional O(τ) value bound; Proposition 3.2 identifies the correction's moving near-the-money profile and associated derivative behavior under stronger assumptions. This directly motivates isolating the strike-line Black–Scholes carrier and learning the remaining scaled correction. The connection is exact at the representation level. Let $\chi=\tau/T$, let $R^{\ast}=W/\chi$ denote the ideal scaled correction, and consider the *pre-FAL* network output using the same exact carrier. Then $$V=C^{\mathrm{ref}}+\chi R^{\ast},\qquad \widehat V=C^{\mathrm{ref}}+\chi R _{\theta} \quad\Longrightarrow\quad V-\widehat V=\chi(R^{\ast}-R _{\theta}).$$ For fixed T > 0, the O(τ) result controls the magnitude of R* near expiry, identifying a mathematically natural target for residual learning. The independent-field pricing experiment shows that this theory-motivated representation can be useful on a specified *finite-maturity* task.

I thank the authors for their comprehensive responses.

Presentation: The authors indicate that they intend to restructure the paper and improve the structure. In the absence of a revised manuscript, I cannot determine whether these changes would indeed address the readability problems, which are pervasive throughout the paper. The underlying question of why a quantum-compatible realization is relevant to option pricing also remains unresolved.

Operator learning: The authors acknowledge that the experiments in the submitted manuscript constitute a “controlled parameterized, per-contract pricing benchmark” and do not demonstrate pricing based on a contract-independent volatility field, which would be the practically relevant application. This was my primary concern, and it applies to every experiment presented in the paper. The new ablation demonstrates that the parameters can be excluded with only minor loss of accuracy. However, within a three-parameter family, the 231 sensor values essentially uniquely identify (σ₀, β, γ), meaning that the problem still amounts to regression over a three-dimensional manifold. This ablation therefore provides no evidence that a map defined on a function space has been learned. For the new fixed-field experiment, the field generator, training data, sensor layout and carrier construction are all left unspecified; it appears that the latency comparison against forward Dupire is measured not at equal grid size.

Relevance of the theory. The response still does not clarify how the asymptotic bounds translate into actual accuracy or learnability.

Based on these points, the rebuttal does not remove my fundamental concerns. I therefore maintain my original rating.

Reviewer_2CR9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes a finance-informed Deep Operator Network (FI-DeepONet) that decomposes the solution operator into a closed-form reference price and an additive learned correction. Synthetic data was used for majority of the trainings and evaluations. Nasdaq-100 (NDX) call quotes from 13 May to 7 August 2026 was used for evaluation and comparison.

Critical strengths
  1. The mathematics are solid, which lays a good foundation for the numerical experiments.
  2. The idea of having Financial Admissibility Layer is innovative.
  3. The model achieves substantial improvements in pricing accuracy and Greek estimation.
Critical weaknesses
  1. The paper doesn’t report wall-clock time.
  2. In 5.2, the input setup is \beta =\gamma =0 and uses constant-volatility sensors. The experiments cannot show the methods has any value on real data.
  3. The paper mentions ‘The reported benchmark fixes T = 1 and uses scalar features (K, r, σ0, β, γ) for the contract and volatility family’. The synthetic data is very narrow.
  4. The Financial Admissibility Layer only enforces pointwise price bounds rather than full no-arbitrage conditions.
Questions for authors
  1. Why is T fixed at 1? Can the model handle other maturities? What the performance will be?
  2. In table 9, the performance of Normalized DeepONet is worse than DeepONet. Is there any reasons behind it?
  3. Can the authors provide a comparison against conventional numerical PDE solvers for the cost of generating training data? What practical advantage does the quantum-compatible realization offer?
3 comments
AuthorsOpenReview ↗

Response to Reviewer 2CR9 (Part 1/2)

CPU wall-clock time and its measurement scope (CW1).

The submitted paper already reports CPU wall-clock timings in Appendix C.7, Table 26. Across five training seeds, the recorded FI-DeepONet training procedure took 437.6 ± 42.9 s. Once model weights were loaded and the analytic carrier cached, inference on 32 held-out spot–time pricing surfaces (106,272 price evaluations) took 0.1266 ± 0.0021 s. Combining separately timed inference and optional post-processing (PP) components gives 0.2055 ± 0.0055 s.

Constant-volatility inputs in the market study (CW2).

Section 5.2 uses market-derived constant-volatility sensors, corresponding to β=γ=0\beta=\gamma=0. For this same-input model, Black–Scholes already gives the exact analytic reference price. The experiment instead measures agreement with observed index-option quotes under the stipulated input. On 2,586 held-out quotes, market relative RMSE was 1.526% for FI-DeepONet and 0.980% for Black–Scholes. Thus the original market study is a quote-level consistency assessment. We use the controlled nonconstant-field comparisons below to assess the correction mechanism.

Breadth of the volatility inputs and the synthetic family (CW3).

The submitted benchmark fixes total maturity T = 1 and uses an explicit three-parameter volatility generator that depends on the queried strike K. To address this distinction, the supplementary study first specifies an absolute local-variance field a(S,t), independently of option contracts, and then varies the query (K,T) while keeping that field fixed. The input comprises 225 sampled local-variance values and the contract query, without generator parameters.

On the pre-specified main subset of 42 moderate/strong nonseparable fields, with 504 near-the-money queries, the separately trained FI residual model attains Price MAE 0.01238. The matched Residual MLP, analytic carrier, and vanilla DeepONet have MAEs 0.04310, 0.05557, and 0.13030, respectively; the observed-field forward Dupire solver achieves 0.003831.

Why T = 1 was used and performance at other maturities (Q1).

Fixing the contract horizon in the submitted benchmark isolates the evolution in remaining life τ=T−t\tau=T-t and makes the short-time representation easier to compare across a controlled family. The supplementary fixed-field setting explicitly tests T∈0.125,0.25,0.5,0.75,1T\in\\{0.125,0.25,0.5,0.75,1\\} years, using the model's maturity gate χ=τ/T\chi=\tau/T. With S0=100S _{0}=100, the main near-the-money evaluation uses strikes K∈97.5,100,102.5K\in\\{97.5,100,102.5\\} and T∈0.25,0.5,0.75,1T\in\\{0.25,0.5,0.75,1\\}, yielding 12 contract queries per field. In this region, FI achieves MAE 0.01238 against 0.05557 for the carrier. A separate short-maturity test with T≤0.125T\leq0.125 yields MAE 0.00607 for FI and 0.00152 for the time-averaged geodesic approximation. Thus the additional training supports multi-maturity pricing in a defined range.

The guarantee and empirical scope of FAL (CW4).

The Financial Admissibility Layer (FAL) constrains individual call prices between their pointwise financial bounds. On the submitted 32 fixed-contract spot–time surfaces, post-FAL spot-monotonicity and spot-convexity violation rates were 1.05% and 1.02% under roundoff-aware checks. In a distinct evaluation on 64 fixed, nonseparable local-variance fields, applying smooth FAL to the learned predictions left median three-seed strike-convexity and calendar violation rates of 0.0174% and 0.0397% (discrete threshold 10−610^{-6}).. The optional grid post-processing removed the checked original spot-grid violations within numerical tolerance.

AuthorsOpenReview ↗

Response to Reviewer 2CR9 (Part 2/2)

Why Normalized DeepONet underperformed in Table 9 (Q2).

The submitted normalized baseline changes three factors simultaneously: S/KS/K query coordinates, V/KV/K output representation, and a training loss equivalent to weighting physical squared price errors by 1/K21/K^2. Its ID Price RelL2 is 0.007554, compared with 0.005985 for vanilla DeepONet, and it has a larger error in each of the five submitted seeds. We separated these choices in a matched 232^3 factorial experiment: eight settings, five seeds each, with common training and validation data, matched update budgets, and checkpoint selection based on physical-price validation error. The average marginal effects on the 32 ID test surfaces are:

Table 3: Marginal effects of normalization choices (positive values indicate higher error).

  • Changed component: Query coordinates S/K; Δ Price RelL2: −0.000524; Δ near-strike p95: −0.250.
  • Changed component: Output V/K; Δ Price RelL2: +0.000697; Δ near-strike p95: +0.244.
  • Changed component: Loss weighted by 1/K²; Δ Price RelL2: +0.000795; Δ near-strike p95: +0.286.

The V/K target changes regression scaling and conditioning; the 1/K² loss changes how strikes contribute to the fitting objective and also changes gradient scale. The effect is also distribution-dependent: in the new factorial experiment the fully normalized endpoint worsened the ID score in only three of five seeds, whereas same-family parameter-OOD RelL2 improved from 0.1987 to 0.0618 in all five.

PDE label costs and the practical quantum-compatible result (Q3).

Generating the submitted 512 training price surfaces required 261.6 s of recorded backward-PDE solver time, excluding neural-network training. For a distinct task with one coefficient field and 54 contracts, forward Dupire required 0.094 s, compared with 1.355 s for scalar backward solves; the forward solver reuses one strike–maturity calculation across contracts.

The quantum-compatible study asks whether a trained classical pricing network can be represented by circuit-compatible linear maps, and how much predictive fidelity is lost under restricted substitution. QO concerns idealized exact factorization of selected maps, whereas QDO substitutes restricted diagonal–orthogonal maps followed by supervised adaptation. For one representative parent and its approximate replacement, global Price RelL2 changes from 0.000837 to 0.001770; this parent-specific evaluation is not the five-seed headline score. The demonstrated contribution is structural compatibility together with its explicit fidelity–coverage trade-off.

Thanks for your reply. I still have concerns regarding on the presentation. I will keep my score.

Reviewer_EKNLOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper builds a neural operator for pricing European calls under local volatility: it maps a local-volatility function and contract terms to the full price surface, together with Delta and Gamma. Rather than learning the price directly, the network learns a correction to an analytic reference price. The authors prove that the exact correction is of order τ\tau near maturity, derive its leading near-strike shape to motivate the linear time scaling, and relate errors in the correction to price and Greek errors. On a synthetic five-parameter local-volatility family, the model reduces price error about sixfold relative to vanilla and physics-informed DeepONets. It is also evaluated on Nasdaq-100 quotes, using implied volatilities as constant-volatility inputs.

Critical strengths
  • Encoding known pricing structure in the output, rather than in the training loss, is a sound design choice, and the ablations attribute the gains to it.
  • The paper is candid about its scope, including failure on off-family volatility perturbations and no improvement over Black–Scholes on market data.
Critical weaknesses
  • Limited pricing novelty. Learning a correction to an analytic price and constraining the output map are established (Hutchinson et al., 1994; Kim et al., 2025); a payoff reference plus a correction scaled linearly in τ\tau appears in Chakhoyan, Kalici and Malandain, "DeepSVM", arXiv:2512.07162, 2025. The distinguishing element, a Black–Scholes price with local variance frozen at the strike, is the leading term of classical short-maturity expansions, and Propositions 3.1–3.2 restate these asymptotics in price form (Berestycki, Busca and Florent, "Asymptotics and calibration of local volatility models", Quantitative Finance 2(1):61–69, 2002; Gatheral et al., 2012). The output layer enforces only pointwise price bounds, not the convexity, monotonicity and calendar conditions that define static no-arbitrage.

  • The case for operator learning is not made. A single forward Dupire PDE solve prices all strikes and maturities for a given local volatility, yet no accuracy–cost comparison with this solver or with classical short-maturity approximations is reported; the only baselines are generic DeepONets.

  • The market experiment does not test the method. With constant-volatility inputs the exact price is the Black–Scholes reference itself, so the network only needs to output a zero correction and cannot improve on Black–Scholes; no local-volatility surface is calibrated.

Questions for authors
  1. What does the construction add beyond Kim et al. (2025) and DeepSVM, and which parts of Propositions 3.1–3.2 are not implied by existing short-maturity expansions?
  2. How does the method compare in accuracy and cost with solvers of the forward Dupire equation and with classical short-maturity approximations?
  3. How often do the predicted surfaces violate convexity, monotonicity or calendar constraints?
  4. How does the method perform on market data with calibrated local-volatility surfaces?
2 comments
AuthorsOpenReview ↗

Response to Reviewer EKNL (Part 1/2)

Novelty and the short-maturity propositions (CW1; Q1).

The contribution is a pricing representation tailored to the local-volatility PDE. The cited predecessors address related, but distinct, objects: Hutchinson et al. (1994) establish learning-network approaches to pricing and hedging; Kim et al. (2025) learn corrections to analytic pricing approximations in stochastic-volatility models; and DeepSVM (Chakhoyan et al., 2025) uses a payoff-anchored, maturity-scaled nonnegative neural correction for physics-informed learning of the Heston operator. Our method instead constructs a field-dependent Black--Scholes carrier from the local variance along the contract's strike, learns a signed correction, and applies the Financial Admissibility Layer (FAL) separately to the reconstructed price:

W=V-C^{\mathrm{ref}},\quad \widehat V_{\mathrm{pre}}=C^{\mathrm{ref}}+\chi R_\theta,\quad \chi=\frac{\tau}{T}.$$ Here $a(S,t)=\sigma^2(S,t)$ and $\tau=T-t$. Thus, unlike a payoff-only anchor, the carrier uses the supplied coefficient field before the learned correction is evaluated. Our propositions make the error relative to the *specific strike-line carrier* the object of analysis. Proposition 3.1 gives a uniform $|W|/K=O(\tau)$ estimate under uniform ellipticity and log-spot regularity. The key cancellation is visible in the correction equation: its source contains $\tfrac12[a(S,t)-a(K,t)]S^2\partial_{SS}C^{\mathrm{ref}}$. Within the $O(\sqrt{\tau})$ near-strike layer, the coefficient difference and the carrier's curvature have compensating short-time orders, yielding an $O(1)$ source and hence an $O(\tau)$ price correction after integration. Proposition 3.2 is more specific: for $S_\tau=K\exp(y\sqrt{\tau})$ and suitable stronger smoothness, $$\frac{W(S_\tau,T-\tau)}{K\tau}\longrightarrow \frac{b_*}{4}\,y\,g_{a_*}(y),\quad a_*=a(K,T),\quad b_*=K\partial_S a(K,T),$$ with corresponding differentiated Delta/Gamma profiles. Where $b _{\ast}\\,y\neq 0$, the limit identifies linear maturity scaling as the critical nontrivial scaling for this correction. Their role is to identify the exact target $R^{\ast}=W/\chi$ for the representation in Eq. (1); for fixed $T>0$, Proposition 3.1 controls its near-expiry magnitude. On the submitted 32-surface test, FI-DeepONet achieved Price RelL2 $0.000965$, versus $0.005985$ for vanilla DeepONet and $0.006890$ for carrier-only (five seeds for the learned models). **Comparison with classical solvers and neural baselines (CW2; Q2).** A forward Dupire solve reuses the computed strike--maturity surface for a fixed coefficient field. We therefore added this financial baseline and classical approximations, alongside learning baselines. Within each comparison, all methods are evaluated on the same contracts against the same numerical pricing references. **Table 1: Price MAE on matched contracts within each experiment (lower is better).**Each pair reports 6 time-homogeneous fields (324 contracts) / 42 nonseparable fields (504 contracts): - FI residual model: 0.17792 / **0.01238** - Vanilla DeepONet: -- / 0.13030 - Residual MLP: -- / 0.04310 - Analytic carrier: 0.04501 / 0.05557 - Classical approximation: 0.01479 / 0.05336 - Forward Dupire PDE: **0.000883** / **0.003831** The FI model in the first setting uses the original smooth FAL configuration; the second is separately retrained with zero-correction initialization and evaluated pre-FAL on a pre-specified near-the-money subset with $T\geq0.25$ years.
AuthorsOpenReview ↗

Response to Reviewer EKNL (Part 2/2)

The second experiment directly addresses a key operator-learning issue: each absolute local-variance field a(S,t)a(S,t) is specified before contracts are chosen and is then held fixed while KK and TT vary. On 42 independent moderate/strong nonseparable test fields, the trained finance-informed correction reduces MAE relative to its carrier by 77.7%, and also outperforms both matched neural alternatives. The three-seed, field-paired 95% interval for MAEFI−MAEcarrier\mathrm{MAE} _{\mathrm{FI}}-\mathrm{MAE} _{\mathrm{carrier}} is [-0.0639,-0.0252]. Thus the evidence identifies a concrete regime-dependent benefit of the finance-informed representation.

Online accuracy–cost comparison (CW2; Q2).

On a separate workload of eight independently specified fields and 3,024 price queries, prepared-input FI inference took 12.7–13.3 ms, compared with 1.447 s for forward Dupire. Both met the same pre-specified criteria, MAE ≤0.02\leq 0.02 and absolute-error p95 ≤0.05\leq 0.05, although Dupire achieved a lower actual MAE (0.00187 versus 0.0084–0.0108 for FI). This is a measured online latency advantage at a common error tolerance. Thus, the relevant application is repeated pricing across changing coefficient fields under an explicit error budget, with offline expenses amortized over future requests.

Financial constraints and measured violations (CW1; Q3).

FAL maps each call price to pointwise lower and upper bounds. This is a useful output-level guarantee, but it is not a proof of monotonicity, convexity, or calendar consistency over an entire surface. We measured the distinct conditions at their relevant evaluation stages:

Table 2: Discrete price-shape checks; the two settings use different models and grids.

  • Submitted model, 32 spot–time surfaces, post-FAL: spot monotonicity, 1.05% violation.
  • Submitted model, 32 spot–time surfaces, post-FAL: spot convexity, 1.02% violation.
  • Additional fixed-field test, 64 fields, smooth FAL applied: strike convexity, 0.0174% violation.
  • Additional fixed-field test, 64 fields, smooth FAL applied: calendar consistency, 0.0397% violation.

Optional grid post-processing eliminated the checked spot-grid violations within numerical tolerance. Accordingly, we define FAL's guarantee precisely as pointwise admissibility and treat cross-contract shape conditions as distinct empirical criteria.

Market-derived inputs and calibrated local volatility (CW3; Q4).

The submitted index-option test uses market-derived constant-volatility sensors (β=γ=0\beta=\gamma=0). Under that specified input, Black–Scholes is the exact corresponding model price, so the learned correction is not expected to improve the analytic same-input reference. On 2,586 held-out market quotes, relative RMSE was 1.526% for FI versus 0.980% for Black–Scholes. This experiment therefore assesses consistency against quotes under a constant-volatility input. We will make this distinction explicit in the manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.