acceptodds
Under review as a conference paper at ICLR 2027

Toward Distributionally Robust Decision-Focused Learning

Abstract

Decision-focused learning often relies on point predictions or decides under a single learned distribution, leaving downstream decisions vulnerable to model error. We propose a distributionally robust framework for decision-focused learning that learns a conditional distribution and conformally calibrates a Wasserstein ambiguity set around it. We use empirical Sinkhorn divergences as nonconformity scores, with corrections for entropic regularization and finite-sample estimation error. Under exchangeability and suitable concentration assumptions, the resulting Wasserstein-1 and squared Wasserstein-2 ambiguity sets contain the true conditional distribution with finite-sample marginal coverage. On the coverage event, the robust objective upper-bounds the true conditional expected task loss. We derive gradients through calibration and downstream optimization under regularity conditions, enabling end-to-end task-aware training. Experiments on portfolio allocation demonstrate improved decision quality over two-stage and stochastic decision-focused baselines.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
distributionally robust optimization, conformal calibration, decision-focused learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_tjF3OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a distributionally robust end-to-end learning framework, which predicts the probability distribution of uncertain parameters through a conditional generative model and constructs a Wasserstein ambiguity set using Sinkhorn divergence and conformal calibration. Theoretically, by correcting entropy regularization and finite-sample estimation errors, the authors establish a finite-sample marginal coverage guarantee that the ambiguity set contains the true conditional distribution, and utilize KKT implicit differentiation to achieve end-to-end optimization of distribution prediction, ambiguity set calibration, and downstream DRO. Finally, several experiments demonstrate that this method reduces decision regret compared to the selected baseline.

From a methodological design perspective, I believe this paper offers very limited methodological innovation. The methodological modules involved, such as end-to-end learning of fuzzy set parameters, using generative models to provide contextual distribution samples, and using conformal prediction to handle the context shift problem, are all readily available in similar papers. The overall workflow framework is very similar to "Robust Decisions via Generative Wasserstein Distributionally Robust Optimization," with the core difference being the use of the same radius for all contexts and the use of Sinkhorn divergences as divergences as nonconformity scores to provide marginal coverage results. This is the paper's core theoretical innovation.

However, the problem lies in the fact that marginal coverage cannot provide the probability guarantee for the specific context that contextual DRO is most concerned with. The average probability guarantee cannot handle extreme context situations. Furthermore, the introduction of Sinkhorn divergences brings additional conservatism compared to Wasserstein distance. Therefore, compared to contextual SO, the current marginal results do not provide a more meaningful statistical guarantee, and the performance improvement compared to contextual SO is merely empirical. I look forward to the authors providing further theoretical results comparing the performance improvement of contextual SO.

Finally, the analysis of the preconditions for KKT implicit differentiation in the paper is insufficient. The convexity and Slater conditions assumed in the main text alone cannot guarantee that the optimal solution maps to differentiable paths. Furthermore, in the actual derivation in Appendix J, the authors use stronger conditions such as nonsingular KKT matrices, which should be reflected in the main text.

Critical strengths

I believe this paper has only one critical strength: using Sinkhorn divergences as nonconformity scores. This solves the difficulty of differentiating the Wasserstein distance, but also introduces additional conservatism. The marginal coverage result obtained through this innovation is the core theoretical difference compared to similar pipeline papers.

Critical weaknesses

As I mentioned in the paper assessment, the paper's main weakness lies in its limited methodological novelty compared to similar e2e DRO papers, and its conservative approach of using the same radius for ambiguity sets under all contexts. The presented marginal theoretical results do not address the conditional coverage for individual contexts, which is particularly important in contextual decision-making. Furthermore, it does not establish any theoretical advantage in decision performance or regret over contextual stochastic optimization. The reported performance improvements over the contextual SO baselines are supported only by empirical results.

Questions for authors

Could the authors provide a theoretical analysis demonstrating the performance advantages of the proposed method over end-to-end contextual stochastic optimization (SO)? I am concerned that using a context-independent radius for all Wasserstein ambiguity sets may introduce excessive conservatism, potentially undermining the performance gains achieved through end-to-end learning. Can the authors theoretically characterize this trade-off and establish conditions under which the proposed method outperforms contextual SO in terms of decision regret?

4 comments
AuthorsOpenReview ↗

Response to Reviewer tjF3 (Part 1)

We thank the reviewer for the time, the thoughtful comments, and the helpful suggestions for improving our paper. We directly address the questions and the raised weaknesses.

Comparisons with Gen-WDRO and the impact of conformal radius

To our best knowledge, Gen-WDRO is the only published paper studying a similar problem and using task-learned radius selection. We implemented the framework of Gen-WDRO according to its paper. We compare our DRDFL against it in both synthetic and real-world experiments, which can be found in Table 1 (DRDFL outperforms Gen-WDRO).

Why Gen-WDRO fails?

  • In experiments, the learned radius of Gen-WDRO decreases to a very small value at the end of the end-to-end training, which comes from using the same data to train the CNF and the neural network for predicting the radius in parallel shown in Figure 1 of [3]. This makes Gen-WDRO's performance very close to stochastic DFL. When there is an obvious distribution shift between the training data and the test data, Gen-WDRO and stochastic DFL are unable to capture such uncertainty and shift and perform worse than DRDFL.

  • In Gen-WDRO the radius is merely trained to minimize the task loss rather than to contain the true distribution. For the problem studied in [3], there can be multiple combinations of the learned distribution's mean and the learned radius giving the same objective value, which is also mentioned in [3] by the authors. The ambiguity set with the learned radius in Gen-WDRO cannot be viewed as the true uncertainty.

Why conformal radius? Any other framework training a distribution predictor and a radius predictor in parallel using the same data and the task loss, the 2 problems of Gen-WDRO mentioned above will occur. To resolve these 2 problems, different data from the one for learning the distribution has to be used to quantify the uncertainty of the predicted distribution in each training iteration. Conformal calibration is the simplistic realization of this and can be differentiated. See Appendix O Comparison with Gen-WDRO for more details.

Also, a fixed radius can simply result in very conservative performance. If a user sets a too large fixed radius, whose effect is similar to a too high confidence level, the regularization coming from the ambiguity set will decrease the performance very much as shown in Figure 4 and 5. In contrast, a too small fixed radius cannot capture the uncertainty and the distribution shift.

Marginal coverage and the common radius. Distribution-free coverage conditional on each context is not attainable in finite samples [1], and contextual SO and Gen-WDRO provide no coverage guarantee of any kind. For a given context, the newly added Appendix N gives a guarantee on the decision that is conditional on that context (Theorem N.1); it applies when outcome samples at the context are available. The common radius is a quantile of the predictor's own errors, so it shrinks as the predictor improves, and the confidence level controls its conservatism (Figures 4–5). With this radius, DRDFL has 17% and 47% lower regret than end-to-end contextual SO (SDFL) with the same predictor (Table 1).

Conservatism of the Sinkhorn score.

In Table 2 (Appendix M.1), we compare the Wasserstein distances between 2 distribution, the Sinkhorn divergences with and without correction terms. It shows that the Sinkhorn divergences with 2 correction terms does not exceed the true Wasserstein distance by a large margin with a proper entropic regularization weight. For high-dimensional distributions, the majority of the "overestimation" comes from the plug-in correction not the entropic correction. For real-world experiments, where only finite samples from distributions (no closed forms of the distributions) are available, the best we can target on is the Wasserstein distances using the plug-in estimators (empirical measures averaged over finite samples). Therefore, the plug-in correction is not added and the "overestimation/conservatism" is very little. See Section 5.1 and Appendix M.1 for more details.

AuthorsOpenReview ↗

Response to Reviewer tjF3 (Part 2)

Theoretical comparison with contextual SO

If the learned law is exact and there is no distribution shift, SO is optimal and no method can improve over it. We therefore characterize the trade-off in two regimes and give the condition under which DRDFL outperforms SO.

Without shift (new Remark J.1). Let dθ(x):=W1(P⋆(⋅∣x),Pθ(⋅∣x))d_\theta(x):=W_1(P_\star(\cdot\mid x),P_\theta(\cdot\mid x)) be the model error, which our calibration bounds by ρθ\rho_\theta with probability at least 1−τ1-\tau. For W1W_1, the DRDFL decision zθ∗z_\theta^\ast minimizes the SO objective plus ρθL(z)\rho_\theta L(z), where L(z)=max⁡k∥ak(x,z)∥∗L(z)=\max_k\Vert a_k(x,z)\Vert_\ast is the sensitivity of the loss to cc (Eq. (10), Proposition J.1). The true and predicted expected losses of any decision zz differ by at most dθ(x)L(z)d_\theta(x) L(z). On the coverage event, this gives:

  • Guarantee. The true expected loss of zθ∗z_\theta^\ast is at most the DRO value (Theorem 4.1), and its regret is at most 2ρθL(z⋆)2\rho_\theta L(z_\star), where z⋆z_\star is the true optimal decision, because the DRO value is at most the robust objective at z⋆z_\star. For the SO decision zSOz_{\mathrm{SO}}, the worst-case loss over the calibrated ambiguity set is its predicted loss plus ρθL(zSO)\rho_\theta L(z_{\mathrm{SO}}), and zθ∗z_\theta^\ast is the decision that minimizes this worst case.
  • Cost. If the learned law is exact, the excess loss of DRDFL over SO is at most ρθ [L(zSO)−L(zθ∗)]\rho_\theta\thinspace[L(z_{\mathrm{SO}})-L(z_\theta^\ast)].

DRDFL is therefore the better choice when the model error is adverse to the SO decision, at a cost of at most the radius times the reduction in sensitivity when the model is exact.

With shift (new Appendix P). DRDFL (ours) performs better than end-to-end contextual SO when there is a distribution shift between the training data and the test data, especially with a large distribution shift. We place SO in its most favorable position: it knows the true conditional distribution of the training data exactly, while DRDFL only has its prediction. Theorem P.1 gives an exact threshold on the W1W_1 shift between the training and the test distribution. Whenever the DRDFL decision is less sensitive than the SO decision (G(x)>0G(x)>0), its worst-case expected loss over all shifts of size at most γ\gamma is smaller than that of SO if and only if γ>γ⋆(x)=Π(x)/G(x)\gamma>\gamma_\star(x)=\Pi(x)/G(x), the loss given up without shift divided by the reduction in sensitivity. When the ambiguity set covers the model error, γ⋆(x)≤2ρθL(zSO)/G(x)\gamma_\star(x)\le 2\rho_\theta L(z_{\mathrm{SO}})/G(x), and γ⋆(x)≤ρθ\gamma_\star(x)\le\rho_\theta if the predictor is exact. The comparison is stated for the worst case because a particular shift may favor either decision.

Context-independent radius. The threshold holds feature by feature, so a common radius does not weaken it. On average over the features, the same threshold holds with averaged quantities, and the features that ρθ\rho_\theta fails to cover only add a term proportional to the miscoverage, which is at most τ\tau on average (Theorem P.1(iii)).

The experiments confirm. DRDFL is better under shift and SDFL is better without shift (Section 5.3).

Marginal Coverage vs. Conditional Coverage

We agree with the reviewer on that a radius with a coverage conditioned on each feature can results in a better performance than a radius with a marginal coverage (averaged over all features).

Post-training conditional coverage. We propose a decision-focused calibration algorithm which can be used after the end-to-end training finishes and gives a coverage guarantee conditioned on each feature (Theorem N.1). See Appendix N Decision-Focused Calibration with a Conditional Coverage Guarantee for details.

Impossibility of conditional coverage in end-to-end training. However, this algorithm is not diferentiable and cannot be integrated in the end-to-end training. We even show that under the current setting having conformal calibration with a conditional coverage in the end-to-end training is impossible. See for Appendix N.1 for details.

Conditions for KKT differentiation

Section 4 now states the conditions under which we differentiate the KKT system (Assumption K.1): linearly independent active constraint gradients, strict complementarity, positive definiteness of H1H_1 on their null space, and smoothness of ∇zΦ1\nabla_z\Phi_1. Under these conditions, the solution is locally unique and differentiable [2].

AuthorsOpenReview ↗

Response to Reviewer tjF3 (Part 3)

Novelty of Our Paper and Framework

Aware of the true uncertainty. Compared to frameworks like Gen-WDRO, whose learnable radius cannot quantify the uncertainty properly, our framework (DRDFL) proposes a true uncertainty-aware end-to-end training of a distribution predictor.

Robust to distribution shift. Compared to stochastic decision-focused learning and Gen-WDRO, whose performance degrades under a distribution shift, ours learns this uncertainty via conformal calibration and performs robustly.

Bounds and coverage. We derive explicit one-sided finite-sample bounds that relate empirical debiased Sinkhorn scores to population W1W_1 and W22W_2^2 discrepancies. Using these bounds, we calibrate Wasserstein ambiguity sets with finite-sample marginal coverage of the true conditional law in an end-to-end decision-focused learning framework.

References

[1] Foygel Barber, Rina, et al. "The limits of distribution-free conditional predictive inference." Information and Inference: A Journal of the IMA 10.2 (2021): 455-482.

[2] Fiacco, Anthony V. "Sensitivity analysis for nonlinear programming using penalty methods." Mathematical programming 10.1 (1976): 287-311.

[3] Xu, Han, and Christopher Yeh. "Robust Decisions via Generative Wasserstein Distributionally Robust Optimization." NeurIPS 2025 Workshop: Generative AI in Finance.


We would be happy to discuss, expand, or provide any further clarifications that the reviewer may consider beneficial.

AuthorsOpenReview ↗

A Gentle Reminder

Dear reviewer,

Did our rebuttal address your concerns? Please let us know. We are happy to add any further clarifications.

The end of the discussion period is quickly approaching.

Thank you.

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

The paper proposes distributionally robust decision-focused learning to protect decisions against misspecification of a learned conditional distribution. It constructs a conformally calibrated Wasserstein ambiguity set around the predicted distribution and trains through both calibration and robust optimization. The analysis provides marginal coverage guarantees and a corresponding bound on expected decision loss under stated assumptions. Experiments evaluate the framework on synthetic and real-data versions of a mean–CVaR portfolio allocation problem.

Critical strengths
  1. The paper addresses a meaningful limitation of stochastic DFL: modeling outcome uncertainty does not by itself protect decisions against errors in the learned distribution. Integrating distribution-level calibration with robust decision-focused training is a coherent response to this problem.
  2. Comparisons with stochastic DFL and robust predict-then-optimize using the same predictor family provide useful evidence for the combined framework. The coverage experiment and confidence-level analysis complement the decision-performance results, and the authors acknowledge that robustness can hurt when there is no distribution shift.
Critical weaknesses
  1. The experiments support the combined framework more clearly than the distinctive contribution of conformal calibration. Comparisons using the same predictor family reduce the concern that gains merely reflect a more expressive model, and the robust predict-then-optimize baseline supports the value of end-to-end training. However, neither isolates conformal radius selection from other ways of choosing the radius in the same robust model. This is particularly relevant given the paper's discussion of Gen-WDRO as a closely related approach with a task-learned radius. The synthetic coverage result is useful, but does not establish the comparative tradeoff among coverage, conservatism, and decision quality. The added empirical value of differentiating through calibration is also not isolated.
  2. The connection between the formal guarantees and the real-data experiment remains insufficiently clear. The paper does address temporal dependence and explicitly reserves held-out data for final calibration. My concern is narrower: the real-data construction treats returns measured at different times within a day as repeated outcomes for a given context, while the sampling correction relies on conditional independence assumptions. It is unclear whether the temporal extension covers this within-period dependence, and how the bounds required by the theory are justified in the implemented calibration. This leaves uncertainty about what coverage interpretation is warranted for the reported real-data results, rather than showing that the theoretical results themselves are invalid.

=> my main concern (W1) is addressed by the authors, so I am raising my score from 2 -> 3.

Questions for authors
  1. What evidence distinguishes conformal radius selection from validation-based or task-learned radius selection with the same predictor and robust optimization layer? A focused comparison reporting coverage, conservatism, and decision quality would help establish its specific contribution. Within that comparison, does allowing gradients through the calibrated radius provide an additional benefit over treating that radius as fixed during backpropagation?
  2. Which guarantee applies to the real-data construction when outcomes within a day are dependent? Please explain whether the temporal extension covers this dependence and how the theoretical sampling corrections and bounds are justified in the implemented calibration. If additional assumptions are needed, please clarify the resulting scope of the empirical coverage claims.
5 comments
AuthorsOpenReview ↗

Response to Reviewer GtQo (Part 1)

We thank the reviewer for the careful reading and the precise formulation of both concerns. We address each weakness together with the corresponding question.

Conformal radius vs. validation-based or task-learned radius (W1, Q1)

To our best knowledge, Gen-WDRO is the only published paper studying a similar problem and using task-learned radius selection. We implemented the framework of Gen-WDRO according to its paper. We compare our DRDFL against it in both synthetic and real-world experiments, which can be found in Table 1 (DRDFL outperforms Gen-WDRO).

Why Gen-WDRO fails?

  • In experiments, the learned radius of Gen-WDRO decreases to a very small value at the end of the end-to-end training, which comes from using the same data to train the CNF and the neural network for predicting the radius in parallel shown in Figure 1 of [4]. This makes Gen-WDRO's performance very close to stochastic DFL. When there is an obvious distribution shift between the training data and the test data, Gen-WDRO and stochastic DFL are unable to capture such uncertainty and shift and perform worse than DRDFL.

  • In Gen-WDRO the radius is merely trained to minimize the task loss rather than to contain the true distribution. For the problem studied in [4], there can be multiple combinations of the learned distribution's mean and the learned radius giving the same objective value, which is also mentioned in [4] by the authors. The ambiguity set with the learned radius in Gen-WDRO cannot be viewed as the true uncertainty.

Why conformal radius? Any other framework training a distribution predictor and a radius predictor in parallel using the same data and the task loss, the 2 problems of Gen-WDRO mentioned above will occur. To resolve these 2 problems, different data from the one for learning the distribution has to be used to quantify the uncertainty of the predicted distribution in each training iteration. Conformal calibration is the simplistic realization of this and can be differentiated. See Appendix O Comparison with Gen-WDRO for more details.

To our understanding both the validation-based radius selection and the conformal calibration use held-out data different from the data for training the distribution predictor. Validation-based selection picks the radius with the best average validation loss. Such a radius says nothing about how likely the set is to contain the true distribution, Besides, when validation data resemble training data, the learned radius shrinks toward zero.

Coverage and Conservatism

The Sinkhorn divergences with 2 correction terms, namely an entropic correction (coming from the mismatch between Sinkhorn and Wasserstein) and a plug-in term (coming from using empirical measures based on finite samples) is an upper bound of the Wasserstein distances averaged over all feasible features. Using it to approximate the true Wasserstein distances between 2 distributions and as the nonconformity score in the conformal calibration makes sure the ambiguity set can cover the true distribution with a high probability. With a distribution shift between the training data and the test data, such an ambiguity set guarantees a good decision quality of the DRO.

We emphasize that using the Sinkhorn divergences does not over estimate the true Wasserstein distances too much. See Appendix M.1 for more details.

Fixed Radius in Backpropagation

Differentiating the calibrated radius, which representing the uncertainty of the predicted distribution in backpropagation is to train the parameter of the distribution predictor in the uncertainy-aware way. The calibrated radius' contribution to the gradient is the sensitivity of the task loss to the uncertainty of the predicted distribution. Setting the radius as a fixed value and not involved in backpropagation simply results in a too large or too small ambiguity set. A large fixed radius means a high confidence level, which makes the regularization coming from the ambiguity set decrease the performance as shown in Figures 4 and 5. In contrast, a small fixed radius cannot capture the uncertainty and the distribution shift.

AuthorsOpenReview ↗

Response to Reviewer GtQo (Part 2)

Guarantees for the real-data construction (W2, Q2)

We agree with the reviewer's diagnosis. Two kinds of dependence are involved, and they enter the analysis at different steps.

  • Across days. The temporal extension (Corollary 3.1, Appendix H) covers dependence between data points.
  • Within a day. Dependence among the SS intraday returns of one data point.

Following reviewer's comment, we have added two new results in Appendix I and a remark to address the Within day dependence.

  • Markov outcomes (Corollary I.1). The first result extends our marginal coverage guarantees to within-sample outcomes that, conditional on xi\boldsymbol{x}_i, form a possibly time-inhomogeneous Markov chain. The extension would serve as a step toward closing the gap between the theory and the empirical results.

  • Arbitrary dependence (Corollary I.2). The second result, shows that when inner-samples are arbitrarily dependent, our DRO formulation guarantees an empirical true-distribution coverage which is aligned with our experimental observations. Exchangeability (Proposition 2.1) requires only i.i.d. tuples, so it also holds for dependent intraday outcomes (Appendix C).

Please note that we distinguish the applicability of the DRDFL optimization framework from the scope of its statistical guarantees. DRDFL is a general framework containing portfolio allocation as one application. DRDFL's empirical bilevel objective, nonconformity scores, and score-quantile radius remain well defined for dependent outcome samples, subject to the stated optimization conditions. Within-day independence is used in our original analysis to establish finite-sample population coverage; it is not required merely to formulate and implement the financial decision problem within DRDFL. This separation between theory and empirical applicability has precedents in closely related literature:

  • [1]: Distributionally robust portfolios with an i.i.d. approximation within estimation windows.

  • [2]: Explicitly distinguish their single-period Wasserstein portfolio theory from its heuristic rolling-horizon implementation in a dynamic market, reporting favorable empirical performance.

  • [3]: End-to-end conformal optimization, with improved decision performance reported under temporal distribution shift while acknowledging that the exchangeability-based coverage guarantee no longer applies.

We made this distinction explicit in Section 5.2.

Bounds in the implemented calibration

The support diameters and spectral-gap constants required by Theorems 3.1–3.2 and Corollary I.1 are not available for real returns, and we do not claim to verify them. The real-data calibration therefore uses rS,M(p)=0r_{S,M}^{(p)}=0 (Appendix M.4), i.e., ρθ=s(k)+ϵ2log⁡(SM)\rho_\theta=s_{(k)}+\frac{\epsilon}{2}\log(SM) with k≥k0k\ge k_0. Corollary I.2 covers this radius without within-day independence or support bounds.

Scope of the empirical coverage claims

  • Section 5.1 (Figures 2–3). The outcomes are conditionally i.i.d. and the support or transport–entropy constants are known, so the figures validate the population coverage of Theorems 3.1 and 3.2.
  • Real data (Figure 5, Table 1). The warranted interpretation is coverage of the test day's realized intraday return distribution, together with the resulting bound on the realized average task loss (Corollary I.2). This holds at level 1−τ1-\tau for exchangeable days, or 1−τ−Δmix1-\tau-\Delta_{\mathrm{mix}} under Corollary 3.1. Section 5.2 (first paragraph) and Appendix M.4 now state this scope.

References

[1] Erick Delage, Yinyu Ye, (2010) Distributionally Robust Optimization Under Moment Uncertainty with Application to Data-Driven Problems. Operations Research 58(3):595-612.

[2] Blanchet, J., Chen, L., and Zhou, X. Y. (2022). Distributionally Robust Mean-Variance Portfolio Selection with Wasserstein Distances. Management Science, 68(9), 6382–6410.

[3] Christopher Yeh, Nicolas Christianson, Alan Wu, Adam Wierman, and Yisong Yue. End-to-end conformal calibration for optimization under uncertainty. Transactions on Machine Learning Research, 2025. ISSN 2835-8856.

[4] Xu, Han, and Christopher Yeh. "Robust Decisions via Generative Wasserstein Distributionally Robust Optimization." NeurIPS 2025 Workshop: Generative AI in Finance.


We would be happy to discuss, expand, or provide any further clarifications that the reviewer may consider beneficial.

AuthorsOpenReview ↗

A Gentle Reminder

Dear reviewer,

Did our rebuttal address your concerns? Please let us know. We are happy to add any further clarifications.

The end of the discussion period is quickly approaching.

Thank you.

Reviewer_cgC1OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
3(+1)
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes a decision-focused learning framework that combines conditional distribution prediction, conformal calibration, and Wasserstein distribution-based robust optimization, which using the Sinkhorn divergence to construct a nonconformity score and calibrates the radius of the fuzzy set through error correction, providing a finite-sample marginal coverage guarantee and an upper bound on the decision loss under certain assumptions.

Critical strengths

This paper introduces conformal calibration into distribution-robust decision-oriented learning. Specifically, it uses the Sinkhorn divergence as a measure of non-consistency to calibrate the Wasserstein fuzzy set radius and provides two guarantees for the finite-sample coverage of fuzzy sets, thereby giving the radius a clear statistical interpretation.

The paper derives gradients tgradient propagation through conformal calibration and the lower-level optimization layer to enables end-to-end training, and achieves improvements over predict-then-optimize methods and stochastic decision-focused learning baselines on synthetic and real-world portfolio allocation datasets.

Critical weaknesses

The theoretical assumptions do not align with the actual experimental data. The coverage guarantee relies on the assumption that target samples are independent of one another given a specific feature; however, in the actual experiment, the samples represent returns at the same time point relative to the previous trading day, and the time windows of adjacent samples overlap significantly, potentially resulting in strong dependence. The time-dependence extension described in Appendix H primarily addresses cross-day dependence and does not correct for the impact of intraday sample correlation on the sampling error bounds.

The experiments are insufficient to support the core claim. The coverage experiments do not explain how to calculate the distance between the true distribution and the predicted distribution, nor do they explain how the constant required for radius correction is obtained. The fact that most actual coverage rates in Figure 2 are below the nominal levels also requires further explanation. Furthermore, while the paper acknowledges that Gen-WDRO is the closest comparable method, it neither compares its results with Gen-WDRO nor conducts key ablation experiments (with fixed or learnable radii), making it impossible to distinguish between the benefits derived from the calibration mechanism itself and those from general robust regularization.

The consistency between the derivation of the gradient and its practical implementation requires further clarification. Appendix J.1 uses the average of the particle-by-particle path derivatives to compute the gradient; however, for piecewise maximum loss, this method may overlook the gradient contributions arising from branch switches. Although Eq. (58) addresses a similar issue, the paper has not yet explained how the actual gradient calculation handles this situation.

Questions for authors

In real-world data, daily return samples show a degree of overlap and strong correlation. How can we satisfy the assumption of independent and identically distributed (i.i.d.) data? What is the basis for the coverage guarantee with such data? Aside from Corollary 3.1, are additional adjustments for internal dependence required?

The number of test data points listed in Section 5.1 and the caption for Figure 2 do not match. Which one should be considered the correct value? How was the coverage in Figure 2 calculated? Compared to Gen-WDRO or the learnable radius scheme (the existing methods most closely related to the proposed method) what is the incremental benefit of the conformal calibration radius?

How is the gradient actually used in Appendix J.1 handled with respect to the switching term of the piecewise loss? Can a numerical comparison with finite differences be provided?

3 comments
AuthorsOpenReview ↗

Response to Reviewer cgC1 (Part 1)

We thank the reviewer for the time, the thoughtful comments, and the helpful suggestions for improving our paper. We address each weakness together with the corresponding question.

Conditional Independence Assumption (W1,Q1)

We distinguish the applicability of the DRDFL optimization framework from the scope of its statistical guarantees. DRDFL is a general framework containing portfolio allocation as one application. DRDFL's empirical bilevel objective, nonconformity scores, and score-quantile radius remain well defined for dependent outcome samples, subject to the stated optimization conditions. Within-day independence is used in our original analysis to establish finite-sample population coverage; it is not required merely to formulate and implement the financial decision problem within DRDFL. This separation between theory and empirical applicability has precedents in closely related literature:

  • [1]: Distributionally robust portfolios with an i.i.d. approximation within estimation windows.

  • [2]: Explicitly distinguish their single-period Wasserstein portfolio theory from its heuristic rolling-horizon implementation in a dynamic market, reporting favorable empirical performance.

  • [3]: End-to-end conformal optimization, with improved decision performance reported under temporal distribution shift while acknowledging that the exchangeability-based coverage guarantee no longer applies.

We made this distinction explicit in Section 5.2.

New results for the dependent case

Following reviewer's suggestion, we have added two new results in Appendix I and a remark.

  • Markov outcomes (Corollary I.1). The first result extends our marginal coverage guarantees to within-sample outcomes that, conditional on xi\boldsymbol{x}_i, form a possibly time-inhomogeneous Markov chain. The extension would serve as a step toward closing the gap between the theory and the empirical results.

  • Arbitrary dependence (Corollary I.2). The second result, shows that when inner-samples are arbitrarily dependent, our DRO formulation guarantees an empirical true-distribution coverage which is aligned with our experimental observations. Exchangeability (Proposition 2.1) requires only i.i.d. tuples, so it also holds for dependent intraday outcomes (Appendix C).

Clarifications of Coverage Validation Experiments (W2,Q2)

Distance Calculation & Constants Selection

In the coverage validation experiments, the closed forms of the 2 distributions are available (one distribution is shifted from the other). The Wasserstein-1 and the Wasserstein-2 distances can be computed using the exact formula.

For bounded support (uniform ball), the support diameter of each distribution (DD and D^\widehat{D}) is available and the diameter of the 2 supports' union (DuniD_{\mathrm{uni}}) can be computed. For unbounded support (Gaussian), the standard deviations of the Gaussian distributions give the transport–entropy constants (νP\nu_P and νθ\nu_\theta). Together with the sample numbers (SS and MM) and 2 user-defined parameters (confidence level 1−τ1-\tau and failure level of sampling η\eta), the radius correction terms can be computed. Details of Coverage Validation is presented in Appendix M.1.

Figures 2 & 3 revised

We thank the reviewer for pointing out the typo. We have revised Figures 2 and 3. The number of calibration points per replicate is now 500 and the averaged empirical coverage is no longer below the user-defined confidence level.

The new figures validate the marginal coverage guarantees in Theorems 3.1 and 3.2 for both distributions with bounded supports and unbounded supports and isolates the impact of different terms in the bounds. Details of experiments and how the constants were canlculated can be found in Section 5.1 and Appendix M.1.

A test point is covered when the exact W1W_1 or W2W_2 distance (computed using the closed forms of the 2 distributions) between the true and the predicted distributions is at most the calibrated radius. For each radius, we compute the probability of this event over an independent test point, and then average those probabilities across calibration sets. This is the empirical coverage in Figures 2 and 3, which estimates the marginal coverage in Theorem 3.1 and 3.2.

The results in the revised Figures 2 and 3 show the empirical coverage is at least the user-defined confidence level in average (marginal coverage guarantees), which supports Theorems 3.1 and 3.2 for bounded supports and their extensions for unbounded supports.

AuthorsOpenReview ↗

Response to Reviewer cgC1 (Part 2)

Comparisons with Gen-WDRO and the impact of conformal radius (W2,Q2)

We implemented the framework of Gen-WDRO according to its paper. We compare our DRDFL against it in both synthetic and real-world experiments, which can be found in Table 1 (DRDFL outperforms Gen-WDRO).

Why Gen-WDRO fails?

  • In experiments, the learned radius of Gen-WDRO decreases to a very small value at the end of the end-to-end training, which comes from using the same data to train the CNF and the neural network for predicting the radius in parallel shown in Figure 1 of [4]. This makes Gen-WDRO's performance very close to stochastic DFL. When there is an obvious distribution shift between the training data and the test data, Gen-WDRO and stochastic DFL are unable to capture such uncertainty and shift and perform worse than DRDFL.

  • In Gen-WDRO the radius is merely trained to minimize the task loss rather than to contain the true distribution. For the problem studied in [4], there can be multiple combinations of the learned distribution's mean and the learned radius giving the same objective value, which is also mentioned in [4] by the authors. The ambiguity set with the learned radius in Gen-WDRO cannot be viewed as the true uncertainty.

Why conformal radius? To our best knowledge, Gen-WDRO is the only published paper studying a similar problem and using a learnable radius. Any other framework training a distribution predictor and a radius in parallel using the same data, the 2 problems of Gen-WDRO mentioned above will also happen. To resolve these 2 problems, different data from the one for learning the distribution has to be used to quantify the uncertainty of the predicted distribution in each training iteration. Conformal calibration is the simplistic realization of this and can be differentiated. See Appendix O Comparison with Gen-WDRO for more details.

Also, a fixed radius can simply result in very conservative performance. If a user sets a too large fixed radius, whose effect is similar to a too high confidence level, the regularization coming from the ambiguity set will decrease the performance very much as shown in Figure 4 and 5. In contrast, a too small fixed radius cannot capture the uncertainty and the distribution shift.

Gradient Derivation vs Implementation (W3,Q3)

We have revised Appendix K.1 (formerly J.1) to separate the population derivative from the implemented one:

  1. Population derivative. The updated Eq. (60) gives the mixed derivative including the switching term, T1=E[∇c∇zℓk(c~)(z∗,c~) J]+pΔ(0) E[∇zΔ α(z∗)⊤J ∣ Δ=0].T_1=\mathbb E\big[\nabla_c\nabla_z\ell_{k(\tilde c)}(z^\ast,\tilde c)\thinspace J\big]+p_\Delta(0)\thinspace\mathbb E\big[\nabla_z\Delta\thinspace\alpha(z^\ast) ^\top J\thinspace\big\vert\thinspace\Delta=0\big]. This is the first-order counterpart of the switching curvature in Eqs. (57)–(58) (formerly (58)–(59)). The second term accounts for probability mass crossing the switching boundary.

  2. Implemented gradient. The implementation does not average particle-wise path derivatives. It solves the lower-level DRO over the MM particles in epigraph form, Eq. (61), with a constraint um≥ℓk(z,c~m(θ))u_m\ge\ell_k(z,\tilde c^m(\theta)) for every particle mm and branch kk. It then differentiates this conic program with CVXPYLayers, through optimality conditions that include all branch constraints and their dual variables. A particle at which two branches tie therefore enters through two active constraints, so switching is handled by the active set. For the drawn particles, this is the gradient of the training loss as evaluated in the forward pass (Section 4, after Eq. (13)). Assumption K.1 states the conditions under which both derivatives exist.

  3. Finite differences. Given the time limits of the discussion period, we will add a finite-difference check to the final version. With the latent draws fixed, we will compare the backpropagated directional derivative with central differences along random directions, for W1W_1 and W22W_2^2, with ρθ\rho_\theta both differentiated and held fixed.

References

[1] Erick Delage, Yinyu Ye, (2010) Distributionally Robust Optimization Under Moment Uncertainty with Application to Data-Driven Problems. Operations Research 58(3):595-612.

[2] Blanchet, J., Chen, L., and Zhou, X. Y. (2022). Distributionally Robust Mean-Variance Portfolio Selection with Wasserstein Distances. Management Science, 68(9), 6382–6410.

[3] Christopher Yeh, Nicolas Christianson, Alan Wu, Adam Wierman, and Yisong Yue. End-to-end conformal calibration for optimization under uncertainty. Transactions on Machine Learning Research, 2025. ISSN 2835-8856.

[4] Xu, Han, and Christopher Yeh. "Robust Decisions via Generative Wasserstein Distributionally Robust Optimization." NeurIPS 2025 Workshop: Generative AI in Finance.


We would be happy to discuss, expand, or provide any further clarifications that the reviewer may consider beneficial.

AuthorsOpenReview ↗

A Gentle Reminder

Dear reviewer,

Did our rebuttal address your concerns? Please let us know. We are happy to add any further clarifications.

The end of the discussion period is quickly approaching.

Thank you.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the Discussions

Dear AC,

We summarize the changes we made and our responses to the reviewers' concerns below.

  1. Conditional Independence Assumption (Reviewer cgC1 and Reviewer GtQo). We extended our results to 2 dependent cases: samples are from time-inhomogeneous Markov chain and samples are arbitrarily dependent. See Corollary I.1 and I.2.

  2. Comparison with Gen-WDRO (Reviewers cgC1, GtQo and tjF3). We compared our work with Gen-WDRO and showed that our framework outperforms Gen-WDRO in both synthetic and real-world experiments. See Table 1 and Appendix O.

  3. Empirical validation of Theorems 3.1, 3.2 (Reviewers cgC1 and GtQo). We gave the technical details and the revised results successfully validated the theorems on coverage guarantees we propose. See Appendix M.1.

  4. Conservatism of Sinkhorn Divergences (Reviewers GtQo and tjF3). We empirically showed that Sinkhorn divergences are close to Wasserstein distances and the overestimation is negligible. See Appendix M.1.

  5. Fixed Radius vs. Calibrated Radius (Reviewers cgC1, GtQo and tjF3). We showed that a fixed radius can easily make the model's performance too conservative (a too large radius) or fail to capture the distribution shift (a too small radius).

  6. Gradient Computation and Conditions (Reviewers cgC1 and GtQo). We revised Section 4 and Appendix J. The current gradient computation and conditions are correct.

  7. Comparison with Contextual Stochastic Optimization (Reviewer tjF3). We theoretically proved a threshold of the distribution shift between training data and test data for ours to outperform contextual stochastic optimization. See Appendix P.

  8. Conservatism of Marginal Coverage (Reviewer tjF3). We showed the impossibility of using a radius with a conditional coverage in the end-to-end training and propose a new conformal calibration algorithm with a conditional coverage, which can be used after the end-to-end training. See Appendix N.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.