acceptodds
Under review as a conference paper at ICLR 2027

NISE: Neural-Informed Shrinkage Estimation for High-Dimensional Derivative Pricing under Financial Models

Abstract

High-dimensional derivative pricing commonly relies on Monte Carlo (MC) simulation or neural partial differential equation (PDE) solvers. However, MC methods may require a large simulation budget to achieve low estimation error, while neural predictions may become inaccurate under distribution shift. To address these limitations, we propose Neural-Informed Shrinkage Estimation (\NISE), a hybrid pricing method that uses predictions from an independently trained neural PDE solver as shrinkage targets in a James–Stein estimator. MC samples determine the shrinkage strength, allowing the estimator to exploit informative neural predictions while reducing their influence when they disagree with simulated prices. On the theoretical side, we establish strict dominance over MC under Gaussianity. We then extend the analysis beyond the Gaussian setting and derive a finite-sample risk bound under mild assumptions. The bound explicitly depends on prediction accuracy and yields an upper bound on the excess risk relative to MC, where denotes the simulation budget; moreover, when the MC budget and neural prediction accuracy lie within an appropriate regime, we recover strict dominance over MC without assuming Gaussianity. On the empirical side, experiments on basket options, credit default swaps (CDS), and option pricing under a Heston model calibrated to real-world market data demonstrate performance improvements.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Neural PDE solvers, Monte Carlo estimation, James–Stein estimation, derivative pricing

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_cWgHOpenReview ↗
Rating
3
Confidence
4(+1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Neural-Informed Shrinkage Estimation (NISE) for jointly estimating a vector of derivative prices. An independently trained neural PDE solver supplies a prediction vector, and a positive-part James–Stein rule shrinks the Monte Carlo (MC) estimate toward that prediction. The shrinkage factor depends on the Mahalanobis discrepancy between the two vectors, reducing reliance on the neural target when this discrepancy is large. With exact covariance, the paper establishes strict dominance over MC in expected Mahalanobis loss under Gaussian estimation errors for at least three jointly estimated prices. Under a finite standardized third-moment assumption, it also derives a non-Gaussian finite-sample excess-risk bound with an explicit dependence on target accuracy and an additive term of order N−1/4N^{-1/4}, for fixed dimension and payoff distribution. Experiments cover geometric-basket options, first-to-default basket CDS, and SPX options under market-calibrated Heston dynamics. They show reductions in joint estimation loss relative to MC; the Heston evaluation measures numerical error under a calibrated, discretized model rather than error against market quotes.

Critical strengths
  1. A simple and useful mechanism for combining a learned prediction with simulation.

NISE requires only the neural price vector at the valuation state, rather than neural-gradient evaluations along every simulated path. Its closed-form shrinkage factor responds to standardized disagreement with MC. The GBM and CDS mismatch experiments show that the method reduces reliance on inaccurate predictions (Figures 4–5). This makes the approach potentially useful when a pretrained solver is available but its evaluation-time accuracy is uncertain.

  1. A substantive non-Gaussian analysis supported by several relevant diagnostics.

Theorem 2 and Appendix B.3 provide an explicit finite-sample risk bound under a finite third-moment condition, rather than relying only on a CLT approximation. The empirical study also includes exact/pilot/same-sample covariance comparisons, pilot-size and dimension sweeps, and reference-price robustness checks (Appendix D). In the GBM study, an informative target achieves a Mahalanobis risk ratio of approximately 0.43, whereas zero and grand-mean targets improve risk by less than 0.4% (Appendix D.5). These results support the claim that informative learned targets can produce meaningful gains over MC.

Critical weaknesses
  1. The contribution beyond existing shrinkage methods and the practical advantage over stronger simulation baselines are not sufficiently established.

Equation (12) is the classical positive-part James–Stein estimator applied after centering at an independent neural prediction and whitening; Theorem 1 is a corresponding classical consequence. The non-Gaussian bound is therefore an important part of the additional contribution. However, Appendix A, lines 693–699, inaccurately groups Green & Strawderman (1991) with work lacking rigorous statistical analysis: that paper's publisher abstract explicitly states a quadratic-loss dominance result. Its assumptions need to be distinguished carefully from the present results. Green & Strawderman, publisher abstract.

Empirically, the paper does not compare against a neural martingale control variate, despite discussing this directly relevant pricing approach in Appendix A, or against standard antithetic/control-variate baselines under matched computational budgets. The CDS baseline already uses conditional MC, and Appendix D.5 does compare simple shrinkage centers; these are useful controls, but do not resolve the stronger-baseline question. Appendix C.1 explicitly excludes offline training and pilot estimation from the main sample-efficiency comparison, without reporting training time, online overhead, or a measured amortization point. Consequently, the results establish gains over the reported MC baselines at a given number of evaluation paths, but leave the incremental methodological contribution and computational benefit insufficiently resolved for acceptance.

  1. The strongest large-vector experiment lacks details and validation needed to assess estimated-covariance behavior.

Both main theorems require the exact covariance, whereas the Heston study uses an estimated pilot covariance for J=533–721J=533–721. Appendices C.2 and C.5 do not specify the Heston pilot size, the number of MC repetitions within each of the five runs, or the rule for handling ill-conditioned or singular covariance matrices. This matters because the inverse covariance determines both shrinkage and the reported loss. The pilot-size study in Appendix D.2 uses J=30J=30, and the dimension study in D.3 uses analytic covariance, so neither directly validates this high-dimensional estimated-covariance setting. Figure 6 reports only Mahalanobis loss; ordinary price MSE or contract-level errors would clarify whether improvements also hold in the original price coordinates. Without these details and checks, the reliability and practical magnitude of the Heston result are difficult to judge.

Questions for authors
  1. Contribution and prior work:

Please correct the characterization of Green & Strawderman (1991) in Appendix A and identify precisely which assumptions or conclusions of the present work go beyond that result and prediction-powered adaptive shrinkage. Is the principal new statistical result the non-Gaussian finite-sample bound, rather than the estimator or Gaussian dominance?

  1. Computational advantage:

Can you provide a matched-budget comparison against a standard control-variate or antithetic MC baseline and a relevant neural martingale control-variate method, at least in one setting? Please report offline training cost, pilot cost, online latency, and the number of repeated queries needed to amortize the additional cost. Such evidence would materially affect my assessment of practical significance.

  1. Heston reproducibility and robustness:

What are the exact pilot size MM, the number of evaluation repetitions per run, and the covariance inversion/regularization rule? Please report covariance conditioning and sensitivity to pilot size or regularization in the actual J=533–721J=533–721 setting, along with ordinary price MSE and uncertainty for the paired NISE–MC improvement. This would clarify whether the result remains persuasive beyond the chosen estimated Mahalanobis metric.

3 comments

We sincerely thank the reviewer for the detailed assessment. Below are our point-by-point responses.

W1. (i) Appendix mischaracterizes Green & Strawderman (1991). (ii) The contribution beyond existing shrinkage methods is not sufficiently clear.

Q1. Please correct the characterization of Green & Strawderman (1991) in Appendix A and identify precisely which assumptions or conclusions of the present work go beyond that result and prediction-powered adaptive shrinkage. Is the principal new statistical result the non-Gaussian finite-sample bound, rather than the estimator or Gaussian dominance?

(i) We thank the reviewer for pointing this out. We have corrected it in the revised manuscript.

(ii) We thank the reviewer for the opportunity to clarify our contribution. Green and Strawderman (1991) cover only the Gaussian case: the unbiased estimator and the biased estimator are both assumed to be normal vectors, and their dominance result is proved under this assumption. In derivative pricing, the MC estimate is an average of payoffs, which may not be exactly Gaussian. Theorem 2 does not assume Gaussianity. This result is therefore not covered by Green and Strawderman (1991). It is the main new statistical result of the paper, and we appreciate that the reviewer lists it among the strengths. Prediction-powered adaptive shrinkage (Li and Ignatiadis, 2025) is now discussed in the revised manuscript. Its guarantee is asymptotic in the number of estimation problems mm (their Theorem 5.2, as m→∞m\to\infty). In NISE, the number of contracts JJ is fixed and all prices are estimated from the same NN paths, and Theorem 2 is non-asymptotic.

W1, continued. (i) Comparison with additional baselines: a neural martingale control variate, and a standard control-variate or antithetic baseline, under matched budgets. (ii) Computational costs: offline training and pilot estimation, training time, online overhead, and the amortization point.

Q2. Can you provide a matched-budget comparison against a standard control-variate or antithetic MC baseline and a relevant neural martingale control-variate method, at least in one setting? Please report offline training cost, pilot cost, online latency, and the number of repeated queries needed to amortize the additional cost. Such evidence would materially affect my assessment of practical significance.

(i) We thank the reviewer for this suggestion. We have added an experiment with MC, NISE, a neural martingale control variate (CV) built from the same DBDP solver following Sabate Vidales et al. (2021), an adaptation of PEMC (Li et al., 2024), and CV+NISE, which applies NISE to the neural CV estimate. The setting is the GBM basket of Section 5.1, with five trained solvers, nine evaluation spots, and N=10,000N=10{,}000 payoff paths for every method. The table below reports the Mahalanobis risk of each method relative to MC; values below one indicate lower risk than MC.

Evaluation spot S0S_0NISENeural CVPEMCCV+NISE
801.0014.73911.1824.627
850.8790.8350.7680.626
900.5550.7560.5730.515
950.5790.6980.5200.477
100 (training spot)0.4330.4370.2890.415
1050.4410.5920.5720.382
1100.4540.5600.5670.358
1150.5240.5410.5770.340
1200.7300.5290.5000.329

Used alone, the neural CV and PEMC can have lower risk than NISE at some spots. However, we also note that NISE is orthogonal to control variates and can be applied on top of them; for example CV+NISE has lower risk than MC and neural CV at 8 of 9 spots.

For a standard control variate, we also added a Black–Scholes control variate in the Heston setting of Section 5.3: NISE has lower Mahalanobis loss than this control variate on all twelve dates: the loss of the control variate stays within 0.4% of that of MC, while the loss of NISE is 13% to 84% below that of MC. The full results are in Appendix F.5 (Figure 13) of the revised manuscript.

(ii) We thank the reviewer for this request. We measured the four costs in the GBM setting of Section 5.1 (d=100d=100, J=30J=30) on a single GPU. Each value is the range over five independently trained DBDP solvers.

CostMeasured value
Offline training of one DBDP solver236 to 253 seconds (1,700 epochs of 100,000 fresh samples each)
Sampling cost for the covariance pilotNone in this setting: as stated in Appendix C.1, the covariance is estimated from the same NN evaluation paths
Online latency of one query at N=8,000N=8{,}000 pathsMC: 43 to 44 ms; NISE: 65 to 71 ms
Queries needed to amortize the training cost680 to 4,740 queries

We also note that these times depend on the hardware and the implementation. The amortization point is the number of queries after which the training time is recovered by the time saved per query, relative to MC run with enough paths to reach the same sum-MSE as NISE (21,200 to 80,700 paths, depending on the solver).

W2. (i) The Heston experiment lacks implementation details and validation of the estimated covariance. (ii) Only Mahalanobis loss is reported.

Q3. What are the exact pilot size MM, the number of evaluation repetitions per run, and the covariance inversion/regularization rule? Please report covariance conditioning and sensitivity to pilot size or regularization in the actual J=533J=533–721721 setting, along with ordinary price MSE and uncertainty for the paired NISE–MC improvement. This would clarify whether the result remains persuasive beyond the chosen estimated Mahalanobis metric.

(i) We thank the reviewer for pointing out these missing details. We have added the details to Section 5.3. On each date, the covariance is estimated from an independent pilot of M=106M=10^6 paths. Each of the five runs uses 500 repetitions with N=2,000N=2{,}000 paths per repetition. We use the full J×JJ\times J pilot covariance with all cross-contract entries, both in the shrinkage factor and in the reported loss. It is rescaled to a correlation matrix and factorized by Cholesky, with no regularization; the factorization succeeded on all twelve dates.

We also ran the requested checks in the actual setting with J=533J=533–721721. The condition number of the pilot correlation matrix is between 1.3×1091.3\times10^{9} and 1.0×10101.0\times10^{10} on the twelve dates. We repeated the evaluation with M∈{104,105,106,107}M\in\{10^4,10^5,10^6,10^7\}. NISE has lower mean Mahalanobis loss than MC in all 48 date–pilot combinations, and for M≥105M\ge10^5 its mean loss differs from that at M=106M=10^6 by at most 0.13% (Appendix F.3, Figure 12 of the revised manuscript).

(ii) We thank the reviewer for this suggestion. We have added a new comparison in price MSE to Appendix F.6 (Figure 14). It reports the sum of squared price errors (sum-MSE) of DBDP, MC, and NISE on the twelve valuation dates, with the DBDP targets trained under the calibration of each date. NISE has lower sum-MSE than MC on all twelve dates, with ratios to MC between 0.08 and 0.34. The gap between NISE and MC is much larger than the variation across the five runs on every date (error bars in Figure 14).

Thank you for the detailed response and revisions. The authors have addressed my main concerns. I maintain my score weak accept and raise my confidence to 4. Good luck!

Reviewer_icj6OpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Neural-Informed Shrinkage Estimation (NISE) for high-dimensional derivative pricing. A pretrained neural PDE solver provides a target, and a James–Stein rule adaptively shrinks Monte Carlo estimates toward it. Assuming exact covariance, the authors establish strict dominance over plain Monte Carlo under Gaussianity in aggregate Mahalanobis risk and derive a finite-sample non-Gaussian risk bound that depends on target accuracy. Experiments on basket options, basket CDS, and SPX options examine the effects of target quality, simulation budgets, distribution shifts, and implementation choices.

Critical strengths
  • I enjoyed reading this paper and found it very clearly written.
  • Personally, I welcome this direction at ICLR. Although derivative pricing is a classical quantitative finance problem, its computational challenges remain a meaningful application area for machine learning.
  • While reading the paper, I wondered how the gains depend on neural prediction accuracy, how they vary with the Monte Carlo budget, and how the method behaves under distribution shift. I appreciated that Results 1–3 in the experimental section directly addressed these questions.
  • I think the extension of the James–Stein risk analysis to the non-Gaussian setting in Theorem 2 a valuable contribution.
  • (Minor suggestion) I suggest emphasizing that, given a specified financial model, the entire NISE framework requires no supervised dataset of ground-truth derivative prices. Both neural PDE training and the Monte Carlo correction rely on simulated samples, avoiding the potentially substantial cost of generating accurate pricing labels.
Critical weaknesses
  • The method is formulated for a fixed collection of maturities and strikes, whereas practical applications require pricing across a continuous range of both. Increasing JJ could provide broader coverage, but a discussion of the resulting trade-offs in computational cost and accuracy would help assess its practicality.
  • My understanding is that Theorems 1 and 2 apply to any target θ~\widetilde{\theta} independent of the Monte Carlo estimator θ^\widehat{\theta}, regardless of whether it is produced by an NPDE solver. In particular, Theorem 1 appears to restate the existing James–Stein guarantee in this setting. While I appreciate the non-Gaussian extension in Theorem 2, it also seems to characterize the shrinkage estimator rather than establish an advantage specific to the proposed NISE construction.
  • The target plays a role conceptually similar to that of a control variate, making a Black–Scholes-based control-variate method a natural and potentially strong baseline. An analytic approximation could be substantially cheaper to construct than a trained NPDE target while still providing useful information for variance reduction. Such a comparison would be particularly informative under non-GBM dynamics, where the Black–Scholes formula is no longer exact.
Questions for authors
  • Is my interpretation of Theorems 1 and 2 correct: that they establish guarantees for shrinkage toward an admissible independent target, rather than an advantage specific to using an NPDE target?

  • Could the authors compare the proposed approach with a BS control variate estimator under non-GBM dynamics? Identifying cases where NISE offers a meaningful advantage, considering both accuracy and computational cost, would strengthen the case for the learned target.

I am positive about the paper overall and would be willing to revise my recommendation to clear acceptance if the authors convincingly address even some of these concerns.

3 comments

We sincerely thank the reviewer for the positive and constructive comments. Below are our point-by-point responses.

W1. The method is formulated for a fixed collection of maturities and strikes, while practice requires a continuous range of both; the trade-off between computational cost and accuracy as JJ increases should be discussed.

We thank the reviewer for raising this point. NISE is applied to a finite set of contracts, but this set does not need to be fixed in advance: it can be any set of contracts that the solver can price. To cover a continuous range, one takes a denser set of contracts, that is, a larger JJ. We discuss the resulting trade-off below.

Accuracy. In Appendix D.3 (Figure 9), we vary the number of contracts from J=3J=3 to J=100J=100 in the GBM experiment. The Mahalanobis risk of NISE relative to MC falls from 0.896 at J=3J=3 to 0.287 at J=30J=30 and 0.132 at J=100J=100. In this experiment, the gain of NISE over MC grows with JJ.

Cost. The simulated paths are shared by all contracts, so a larger JJ does not necessarily require more simulations. The additional cost of NISE is one solver evaluation per contract (at most linear in JJ), and a J×JJ\times J covariance estimate and factorization (up to cubic in JJ). We measured each step in the GBM setting with N=8,000N=8{,}000 paths, with 30 maturities and J/30J/30 strikes per maturity:

JJMC querySolver evaluationCovariance estimateSolve and shrinkage
3043 ms21 ms0.4 ms0.5 ms
9043 ms21 ms1.0 ms0.6 ms
30043 ms21 ms6.3 ms1.8 ms
90043 ms21 ms60 ms7.4 ms

The solver evaluation does not grow with JJ in this table because of how the solver is built: all strikes of one maturity are priced by the same network in a single batched call. It is timed here with networks of the same architecture as in Section 5.1 (Appendix F.4, Table 5 of the revised manuscript).

W2. (i) Theorems 1 and 2 apply to any independent target, not specifically to an NPDE target. (ii) Theorem 1 restates the existing James–Stein guarantee.

Q1. Is my interpretation of Theorems 1 and 2 correct: that they establish guarantees for shrinkage toward an admissible independent target, rather than an advantage specific to using an NPDE target?

(i) We thank the reviewer for the careful reading; this interpretation is correct. The guarantees hold for any target that is independent of the evaluation paths. We see this as a strength of the framework: the guarantee does not depend on the accuracy of the solver, so NISE can be used with a pretrained solver whose accuracy at evaluation time is unknown. What is specific to the NPDE target is the size of the gain. The bound in Theorem 2 improves as the target becomes more accurate, and an NPDE solver provides an accurate target for the whole price vector. For example, a zero or grand-mean target changes the risk by less than 0.4%, while the NPDE target lowers it by 57% (Appendix D.5).

(ii) We thank the reviewer for pointing this out. Theorem 1 restates the classical James–Stein guarantee; in our revised manuscript, we have highlighted this above Theorem 1. We restate this classical result because it is meant to be read together with Theorem 2. Theorem 1 describes the ideal case: if the MC errors were exactly Gaussian, NISE would have strictly lower risk than MC for any independent target. However, in practice, MC errors may not be exactly Gaussian. Theorem 2 shows what is lost without this assumption: strict dominance is no longer guaranteed, and we obtain instead a finite-sample bound on the risk of NISE relative to MC, which also depends on the accuracy of the target. We appreciate that the reviewer lists Theorem 2 among the strengths.

W3. Comparison with an analytic Black–Scholes control variate under non-GBM dynamics.

Q2. Could the authors compare the proposed approach with a BS control variate estimator under non-GBM dynamics? Identifying cases where NISE offers a meaningful advantage, considering both accuracy and computational cost, would strengthen the case for the learned target.

We thank the reviewer for this suggestion. We have added this comparison to Appendix F.5 of the revised manuscript (Figure 13). It uses the Heston setting of Section 5.3, with the same 2,000 paths per repetition for all methods.

NISE has lower Mahalanobis loss than BS CV on all twelve dates. The loss of BS CV stays within 0.4% of that of MC (so it looks the same as the MC method in that figure), while the loss of NISE is 13% to 84% below that of MC.

The time per query is similar for the three methods: at 2,000 paths, the median on one CPU thread is 82 ms for MC, 87 ms for BS CV, and 93 ms for NISE. The methods differ in what must be prepared before any query: BS CV needs a coefficient pilot (3 to 5 seconds per date), while NISE needs a covariance pilot and a trained solver (6 to 10 minutes on one single GPU for this specific experiment). NISE therefore offers a clear advantage when many contracts are priced jointly and the solver is reused across queries.

W1. I understand that although each evaluation involves a finite set of maturities and strikes, this set need not be fixed in advance. Provided that the underlying solver supports continuous queries, the framework is more flexible than I initially understood. Thank you for clarifying this point.

W2. The clarification confirms my understanding of the theoretical results. I therefore view the theoretical contribution as more incremental than I initially expected, although these results still provide relevant theoretical justification for the proposed approach.

A personal thought on the authors' statement, “We see this as a strength of the framework: the guarantee does not depend on the accuracy of the solver, …,” with no response needed: I do not find solver independence itself a convincing strength here. Similar arguments are often made for optimization-agnostic bounds, such as PAC-Bayesian bounds. In option pricing, such guarantees have little practical value if they are too loose to say much about pricing accuracy.

Overall, while the theoretical contribution is somewhat more incremental than I had initially anticipated, I believe the paper meets the acceptance bar for ICLR. I will therefore raise my recommendation to clear acceptance.

Reviewer_G9R7OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The authors study the pricing of financial derivatives that requires computation of discounted expectations under a suitable martingale measure.Two standard approaches are

  • Monte Carlo (MC) simulation averages payoffs over simulated market scenarios. It is unbiased, but needs many scenarios to be accurate.
  • A neural network trained in advance to solve the pricing PDE

NISE combines the two methods by taking a weighted average of the Monte Carlo prices and the network’s prices, with a single weight for the J derivatives. The weight is chosen automatically. When the two sets of prices agree to within the Monte Carlo noise, the average relies mostly, or entirely, on the network. When they disagree, it relies mostly on Monte Carlo.

The authors prove two results.

  • Theorem 1: NISE is on average more accurate than MC, in an error measure where each price error is scaled by its MC noise. This holds if the MC errors are exactly normally distributed and their covariance is known
  • Theorem 2: without the normality assumption, NISE can be worse than MC only by an amount that shrinks as the number of simulated scenarios N grows, at rate N^(−1/4).

The experiments cover three settings:

  • basket options on 100 correlated assets;
  • credit default swaps;
  • S&P 500 index options under a stochastic-volatility (Heston) model calibrated to 2023 market data.
Critical strengths

-The method is simple and can be added on top of any pretrained pricer. Because the weight on the network falls when it disagrees with Monte Carlo (MC), the damage from an inaccurate network is limited, as the distribution-shift experiments show. -The paper is clearly written, and the appendix is candid about what is measured.

Critical weaknesses
  • Limited novelty. Combining an unbiased estimate with an independent, possibly biased one using James–Stein weights is due to Green and Strawderman (1991). Applying this to a Monte Carlo (MC) estimate and a neural prediction is due to Gu et al. (2022). Theorem 1 is the classical James–Stein result applied without change (Lehmann and Casella, 1998). Theorem 2 uses a standard central-limit argument and does not engage with existing non-Gaussian James–Stein theory (Kim and White, 2001; Hansen, 2016).

  • The theory does not cover the method as used, or the error that matters for pricing. Both theorems assume a known covariance, which is estimated in all experiments. Theorem 1 also assumes exactly normal errors. Theorem 2 does not guarantee improvement for a fixed, imperfect network as NN grows: its gain decays as O(N−1)O(N^{-1}), while its excess term decays as O(N−1/4)O(N^{-1/4}). The guarantees use an error measure scaled by MC noise, not price error. Under squared price error, improvement requires tr(Σ)/λmax⁡(Σ)>2\mathrm{tr}(\Sigma)/\lambda_{\max}(\Sigma) > 2 (Bock, 1975). Because a single weight is shared across contracts, individual prices can become worse, and per-contract errors are not reported.

  • No comparison with unbiased neural variance-reduction methods. Sabate Vidales et al. (2021), cited only in Appendix A, build unbiased control variates from the same class of neural PDE solver. Ech-Chafiq et al. (2021), Hinds and Tretyakov, and Li et al. (2024) are neither cited nor compared. NISE uses only the network's price and introduces bias. All test problems have closed-form or semi-analytic prices, and the market-data experiment measures error against model prices, not market quotes.

  • Z. El Filali Ech-Chafiq, J. Lelong, and A. Reghai. Automatic control variates for option pricing using neural networks. Monte Carlo Methods and Applications, 27(2):91–104, 2021.

  • B. E. Hansen. Efficient shrinkage in parametric models. Journal of Econometrics, 190:115–132, 2016.

  • P. D. Hinds and M. V. Tretyakov. Neural variance reduction for stochastic differential equations. Journal of Computational Finance, 27(3), 2023.

  • T.-H. Kim and H. White. James–Stein-type estimators in large samples with application to the least absolute deviations estimator. Journal of the American Statistical Association, 96(454):697–705, 2001.

  • E. L. Lehmann and G. Casella. Theory of Point Estimation, 2nd edition. Springer, 1998.

  • F. Li, H. Chen, J. Lin, A. Gupta, X. Tan, G. Xu, Y. Nevmyvaka, A. Capponi, and H. Lam. Prediction-enhanced Monte Carlo: a machine learning view on control variate. arXiv:2412.11257, 2024.

Questions for authors
  1. What does NISE, and do Theorems 1 and 2, add beyond Green and Strawderman (1991), Gu et al. (2022), and existing non-Gaussian James–Stein theory?
  2. How does NISE compare, at matched cost, with unbiased neural control variates built from the same solver (Sabate Vidales et al., 2021) and with Li et al. (2024)?
3 comments

We sincerely thank the reviewer for the careful and insightful review. Below are our point-by-point responses.

W1. Some novelty issues. Discuss (i) Green and Strawderman (1991), (ii) Gu et al. (2022), and (iii) existing non-Gaussian James–Stein theory (Kim and White, 2001; Hansen, 2016).

Q1. What does NISE, and do Theorems 1 and 2, add beyond Green and Strawderman (1991), Gu et al. (2022), and existing non-Gaussian James-Stein theory?

We agree that combining James–Stein-type estimation with the Monte Carlo method is not new, and that a neural-network variant has been applied to image rendering by Gu et al. (2022), and the non-Gaussian analysis has been made by Kim and White (2001) and Hansen (2016). We appreciate the opportunity to clarify our contributions relative to these works.

(i) Green and Strawderman (1991) require exact Gaussianity throughout their work. As stated in Section 4.2 of our submission, "finite-sample MC estimates need not be exactly Gaussian". We therefore establish in Theorem 2 a finite-sample risk bound that does not require Gaussian payoff vectors or Gaussian MC estimation errors, a setting outside the scope of Green and Strawderman (1991).

(ii) The method of Gu et al. (2022) is designed specifically for image rendering. It includes steps such as blockwise shrinkage within local pixel windows, a network trained with supervision on reference images to estimate per-pixel variances, and a regression on rendering features to reconstruct the biased input. None of these steps appears in NISE; we use a closed-form shrinkage factor and need no reference prices. In addition, their theoretical discussion relies on the Gaussian bound of Green and Strawderman (1991) under an independence assumption, and they note that this assumption does not hold for denoising because the denoised image correlates with its input. In NISE, there is no such issue, since its target is trained independently of the evaluation paths.

(iii) While Kim and White (2001) and Hansen (2016) both provide non-Gaussian results, they do not cover a finite simulation budget or a target at a fixed distance from the truth. Both results are asymptotic: Kim and White define risk through the limiting distribution, and Hansen takes the limit of a trimmed loss as n→∞n\to\infty. Instead, we establish a finite-sample upper bound on the Mahalanobis risk for every NN and every target, which these asymptotic results do not provide.

We have revised our manuscript to make these distinctions explicit. We thank the reviewer again for these suggestions.

W2. Gap between theory and practice: (i) both theorems assume a known covariance; (ii) Theorem 1 assumes normal errors, while Theorem 2 does not guarantee improvement for a fixed, imperfect network as NN grows; (iii) the guarantees use Mahalanobis loss rather than price error; (iv) a single shared weight can worsen individual prices, and per-contract errors are not reported.

We thank the reviewer for these insightful observations and we address them below.

(i) Thanks for this observation. To address this concern, our submission has included the exact-covariance setting of the theorems and compares it with the estimated versions on identical paths (Appendix D.1, Figure 7). The results are nearly identical: at the training spot S0=100S_0=100, the risk ratios are 0.4325 (exact), 0.4328 (independent pilot), and 0.4309 (same sample). Estimating the covariance therefore has little effect on the method as used.

(ii) We thank the reviewer for this insightful question; this is exactly the message of our two theorems. In the ideal Gaussian setting, NISE always beats MC (Theorem 1). However, in practice, MC errors are not Gaussian; we can no longer prove that NISE always wins. Instead, we can only prove that NISE cannot lose to MC by more than an amount of order N−1/4N^{-1/4}, for every NN and every network (Theorem 2). For parameters beyond our theory, we have validated the performance of NISE using empirical experiments.

(iii) We thank the reviewer for this observation. We agree that our guarantees are stated in the Mahalanobis loss, and that a guarantee under squared price error requires an additional condition, as in Bock (1975). We use the Mahalanobis loss because the JJ contracts are priced from the same paths, so their MC errors are correlated and differ in size across contracts. For the price error itself, we provide empirical evidence. Figure 3 of our submission reports the sum of squared price errors (sum-MSE), and NISE is below MC at every budget for the three spots shown. In the new experiment described below, which covers nine spots, the sum-MSE ratio of NISE to MC is below one at all nine spots, and is 0.63 at the training spot.

(iv) We thank the reviewer for this suggestion. To address this concern, we have added an experiment that reports the RMSE of every contract at all nine spots. At the training spot, NISE lowers the MSE of 24 of the 30 contracts; the other six are the six shortest maturities, where MC is already most accurate. Moreover, at S0=120S_0=120, NISE lowers the MSE of all 30 contracts. We also note that in many applications, the contracts are used through portfolios and not one by one. In such applications, the accuracy of each individual contract is not the main concern. The full results are in Appendix F.1 (Table 2 and Figure 11) of the revised manuscript.

W3. (i) Compare NISE, at matched cost, with the neural control variate (Sabate Vidales et al. (2021)) built from the same solver, and PEMC (Li et al. (2024)). (ii) Cite and discuss related work, including Ech-Chafiq et al. (2021), Hinds and Tretyakov (2023), and Sabate Vidales et al. (2021). (iii) All test problems have closed-form or semi-analytic prices. (iv) The market-data experiment measures error against model prices, not market quotes.

Q2. How does NISE compare, at matched cost, with unbiased neural control variates built from the same solver (Sabate Vidales et al., 2021) and with Li et al. (2024)?

(i) We sincerely thank the reviewer for this valuable suggestion. NISE is orthogonal to control variates: a control variate lowers the variance of the MC estimate, while NISE shrinks an unbiased estimate toward the solver's price, so NISE can be applied directly to a control-variate estimate. We have added an experiment with MC, NISE, a neural control variate (CV) built from the same DBDP solver following Sabate Vidales et al. (2021), an adaptation of PEMC (Li et al., 2024), and CV+NISE, which applies NISE to the neural CV estimate. The table reports the Mahalanobis risk of each method relative to MC (lower is better), with the same number of payoff paths for every method.

Evaluation spot S0S_0NISENeural CVPEMCCV+NISE
801.0014.73911.1824.627
850.8790.8350.7680.626
900.5550.7560.5730.515
950.5790.6980.5200.477
100 (training spot)0.4330.4370.2890.415
1050.4410.5920.5720.382
1100.4540.5600.5670.358
1150.5240.5410.5770.340
1200.7300.5290.5000.329

Used alone, the neural CV and PEMC can have lower risk than NISE at some spots. However, we also note that NISE is orthogonal to control variates and can be applied on top of them; for example CV+NISE has lower risk than MC and neural CV at 8 of 9 spots.

(ii) We thank the reviewer for these references. In the revised manuscript, we now discuss Sabate Vidales et al. (2021) in the related work of the main text (Section 2), together with Ech-Chafiq et al. (2021) and Hinds and Tretyakov (2023). We use the control variate of Sabate Vidales et al. (2021), which can be built from the same solver as NISE, as their representative in the new comparison in (i).

(iii) We thank the reviewer for raising this point. Among our three test problems, the CDS experiment (Section 5.2) has no closed-form or semi-analytic price: its reference prices come from an independent simulation with 10610^6 paths, and NISE still lowers the Mahalanobis risk relative to MC there (Figure 5).

(iv) We thank the reviewer for this comment. NISE is a numerical method whose goal is to compute the price under a given model accurately, so we measure its error against the model price, as stated in Appendix C.5. The gap between the model price and the market quote is a model and calibration error; it depends on which model is chosen and how it is calibrated. In this experiment, the role of real market data is to determine the test problem: it provides realistic model parameters, a realistic set of contracts, and twelve different market conditions; by using the real-market data, the Heston parameters cannot be cherry-picked to favor NISE.

Reviewer_G9R7OpenReview ↗

Adressing boith comments

Thank you for the clarifications and additional experiments. The per-contract results and comparisons with neural control variates and PEMC strengthen the empirical case.

I still see Theorem 2 as a modest technical contribution. The finite-sample bound extends earlier asymptotic work, but assumes known covariance and uses Mahalanobis loss. F

On balance, the additional experiments and the clarification of the finite-sample contribution move me to weak accept. I still view this as a borderline case given the limited methodological novelty.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.