acceptodds
Under review as a conference paper at ICLR 2027

Reflected Fenchel-Young Conformal Calibration for Decision-Aware Contextual Robust Optimization

Abstract

Conformal prediction is commonly used to construct uncertainty sets for contextual robust optimization. However, residual-based calibration can enlarge these sets in response to prediction errors that do not affect the optimal decision, inducing unnecessary conservatism. This motivates calibrating the regret of the nominal decision directly. However, we show that, under regularity conditions, minimizing worst-case regret over the resulting regret set uniquely recovers the nominal decision, so robust optimization leaves the decision unchanged. We address this limitation using reflected Fenchel-Young (FY) conformal scores for convex optimization with linear objective uncertainty. The score provides a richer characterization of decision performance, admitting an exact decomposition into nominal regret, the suboptimality of the hindsight optimizer under the prediction, and a reflected stability gap. Split conformal calibration gives finite-sample marginal guarantees. We further characterize when the calibrated set improves worst-case regret and develop a master-pricing algorithm with optimality certificates. Experiments confirm the predicted self-centering phenomenon and show that reflected FY refinement improves decision quality over residual-based robust baselines in portfolio and real-data solar-grid allocation problems.

Primary area:
optimization
Keywords:
decision-aware, contextual robust optimization, fenchel-young gap, conformal prediction

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_F3ePOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper studies conformal calibration for contextual robust optimization with linear objective uncertainty. It first shows that minimizing worst-case regret over an uncertainty set defined by nominal regret recovers the original predict-then-optimize (PTO) decision under certain conditions. To enable decision refinement, the authors propose a reflected Fenchel–Young (FY) score, which decomposes into nominal regret, a reverse optimality gap, and a reflected stability gap. The paper establishes finite-sample marginal regret guarantees, characterizes when robust optimization changes the nominal decision, and develops a master–pricing algorithm. Experiments consider two-asset allocation, portfolio optimization, and solar-grid allocation.

Critical strengths

The self-centering result provides a useful characterization of exact-regret calibration. The two-asset allocation example clearly illustrates why minimizing worst-case regret over the nominal regret set cannot improve the PTO decision. The three-gap decomposition and the optimality conditions provide a mathematical explanation of when decision refinement is possible. The portfolio experiments also include global optimization certificates to assess the accuracy of the proposed algorithm.

Critical weaknesses

The motivation for decision refinement could be clearer. Theorem 1 establishes that exact-regret calibration retains PTO, but it is unclear why changing this decision is desirable in itself. A reduction in worst-case regret over the FY set does not necessarily imply an improvement over direct-regret calibration.

The tightness of the regret guarantee is not sufficiently examined. Conservatism can arise from three sources: conformal coverage exceeding 1−α1-\alpha; the nonnegative reverse and reflected gaps in SλS_\lambda; and the inequality R∗(U)≤ρU(x^)≤ηα,λR^*(U)\leq\rho_U(\hat x)\leq\eta_{\alpha,\lambda}. In the portfolio experiment, the calibrated threshold is larger the largest certified upper bound on worst-case regret. The contributions of these sources of slack remain unclear.

The practical interpretation of the guarantee could be better motivated. The guarantee controls regret relative to the hindsight optimum of the penalized objective, rather than portfolio losses or returns. It also does not imply improvement in mean regret or CVaR_95 relative to PTO, which are the main empirical performance metrics.

Questions for authors

The self-centering result shows that direct-regret calibration retains PTO, but this does not necessarily imply inferior decision performance. Could the authors provide stronger theoretical support for why reflected FY calibration should be preferred?

3 comments
AuthorsOpenReview ↗

Response to Reviewer F3eP (Part I)

We thank Reviewer F3eP for the positive assessment and constructive questions. We hope this response adequately addresses the reviewer's concerns, and we are more than happy to engage in further discussions.

W1: The motivation for decision refinement could be clearer. Theorem 1 establishes that exact-regret calibration retains PTO, but it is unclear why changing this decision is desirable in itself. A reduction in worst-case regret over the FY set does not necessarily imply an improvement over direct-regret calibration.

Q1: The self-centering result shows that direct-regret calibration retains PTO, but this does not necessarily imply inferior decision performance. Could the authors provide stronger theoretical support for why reflected FY calibration should be preferred?

Response: We agree that PTO is not necessarily inferior; indeed, it minimizes expected regret when the prediction equals the true conditional mean. However, our focus is robust decision-making under substantial uncertainty, where prediction bias and costly adverse outcomes motivate protection beyond nominal optimization, as extensively studied in the contextual robust optimization and decision risk assessment literature (Chenreddy et al., 2022; Sun et al., 2023; Zhou et al., 2026).

Theorem 1 identifies a structural limitation: direct-regret calibration necessarily retains PTO under the stated conditions, precluding further minimax-regret refinement. Our reflected FY construction addresses this limitation through decision-aware uncertainty calibration, enabling robust refinement that can improve decision quality when prediction bias undermines nominal decisions.

Theory and experiments also support the value of refinement. For a fixed context and a population uncertainty set with conditional coverage at least 1−α1-\alpha, Eq. (21), proved in Appendix A.2, shows that FY improves the population regret-quantile guarantee when its optimization gain exceeds the additional certification cost (Grob>CλG_{\rm rob}>C_\lambda). Examples 1–2 in Appendix A.5 give conditions for strict mean and tail improvement. In our portfolio experiment, historical training and tuning data are shifted relative to the common calibration and test distribution, preserving calibration/test exchangeability. FY reduces mean regret by 8.93% and CVaR95 by 5.51% (Table 1). The accompanying portfolio diagnostic links the mean gain to bias correction net of nominal opportunity cost (Eq. (22); Table 12, Appendix B.3).

References

  • Chenreddy, A. R., Bandi, N., & Delage, E. (2022). Data-Driven Conditional Robust Optimization. Advances in Neural Information Processing Systems, 35, 9525–9537.

  • Sun, C., Liu, L., & Li, X. (2023). Predict-then-Calibrate: A New Perspective of Robust Contextual LP. Advances in Neural Information Processing Systems, 36, 17713–17741.

  • Zhou, W., Orfanoudaki, A., & Zhu, S. (2026). Conformalized Decision Risk Assessment. International Conference on Learning Representations, 2026, 10581–10610.

AuthorsOpenReview ↗

Response to Reviewer F3eP (Part II)

W2: The tightness of the regret guarantee is not sufficiently examined. Conservatism can arise from three sources: conformal coverage exceeding 1−α1-\alpha; the nonnegative reverse and reflected gaps in SλS_\lambda; and the inequality R∗(U)≤ρU(x^)≤ηα,λR^*(U)\le\rho_U(\hat{x})\le\eta_{\alpha,\lambda}. In the portfolio experiment, the calibrated threshold is larger the largest certified upper bound on worst-case regret. The contributions of these sources of slack remain unclear.

Response: The calibrated score threshold is a direct conformal bound, rather than the tightest regret certificate attainable from the FY uncertainty set. In fact, our manuscript already provides the sharper, decision-dependent certificate Reg⁡(xR;θ)≤R∗(U)≤ρU(x^)≤ηα,λ\operatorname{Reg}(x_R;\theta)\le R^*(U)\le\rho_U(\widehat x)\le\eta_{\alpha,\lambda} on the conformal coverage event (Corollary 2).

We further quantify the sources of slack in an added experiment (Table 3; Appendix B.3, Eq. (41)). In the main-text portfolio experiment (1,000 instances, κ=2\kappa=2, α=0.1\alpha=0.1), our diagnostic finds that the mean certified policy bound is 68.64% below the score threshold. The reverse and reflected gaps explain 97.14% of this tightening:

ComponentShare of tightening
Unused score budgetNegligible
Weighted reverse gap32.62%
Weighted reflected gap64.52%
Optimization gain (certified lower bound)2.86%

Empirical score coverage is 89.3% against the 90% target, while the tightened regret certificate covers 95.0% of realized outcomes (Table 13, Appendix B.3).

W3: The practical interpretation of the guarantee could be better motivated. The guarantee controls regret relative to the hindsight optimum of the penalized objective, rather than portfolio losses or returns. It also does not imply improvement in mean regret or CVaR95\mathrm{CVaR}_{95} relative to PTO, which are the main empirical performance metrics.

Response: Our guarantee bounds realized shortfall from the modeled objective’s hindsight optimum, with finite-sample marginal coverage (Corollary 2).

The portfolio objective combines realized return with a quadratic penalty. The common hindsight benchmark gives

Reg⁡PTO(θ)−Reg⁡FY(θ)=θ⊤(xFY−x^)−12(xFY⊤HxFY−x^⊤Hx^).\operatorname{Reg}_{\mathrm{PTO}}(\theta)-\operatorname{Reg}_{\mathrm{FY}}(\theta)=\theta^{\top}(x_{\mathrm{FY}}-\widehat{x})-\tfrac{1}{2}\bigl(x_{\mathrm{FY}}^{\top}Hx_{\mathrm{FY}}-\widehat{x}^{\top}H\widehat{x}\bigr).

Thus, mean regret reductions equal absolute increases in the mean penalized objective value.

The conformal guarantee and comparative performance address different questions. The certificate does not guarantee lower mean regret or CVaR95 than PTO. Under linear objective uncertainty, PTO minimizes expected regret when predictions equal the conditional means, precluding universal strict improvement in both metrics. However, we focus on robust decision-making under prediction bias. Gains depend on prediction errors, decision structure, and the outcome distribution. Refinement improves expected regret when the benefit of correcting bias exceeds nominal opportunity cost (Eq. (22)); tail gains require analysis of the distribution (W1/Q1).

The portfolio experiment shows advantages in both mean and tail regret over PTO and the reported robust baselines (Table 1), while the real-data solar experiment demonstrates improved tail performance relative to PTO (Table 4).

I thank the authors for their response. My concerns have been addressed, and I will maintain my positive score.

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

This paper considers optimization problems of the form max⁡x⟨θ,x⟩−Ψ(x)\max_x \langle \theta, x \rangle - \Psi(x), where θ\theta is an uncertain parameter. For a fixed θ\theta, this optimization problem is the convex conjugate. The paper uses conformal prediction to consider the worst performance around an uncertainty set for the values of θ\theta. Each θ\theta leads to a decision xθx_\theta, and the authors consider the regret of a decision xθ^x_{\hat \theta} based on an estimate θ^\hat \theta, when deployed on the realized θ\theta. Theorem 1 shows that this does not affect the min-max decision, which follows from the linearity of the objective in θ\theta. As a result, the authors next consider a weighted symmetrization of regret, which admits a decomposition in Theorem 2. Theorem 3 provides a conformal prediction guarantee based on this decomposition. Theorem 4 uses subdifferential calculus to discuss the resulting min-max optimization from using this symmetrized variant as a score function. Theorem 5 is a Jensen-inequality-like characterization of the optimal value of the min-max optimization, which draws from distributionally-robust optimization. The numerical experiments consider a budget allocation problem, a portfolio optimization problem, and a solar grid installation problem.

Critical strengths
  • The paper has a good idea at its core, combining classical results in convex analysis with conformal prediction. The consequences of classical Fenchel conjugate results are well-explored and thought out in this draft. The theorems are a meaningful combination of the ideas in new ways, finding new applications for established proof techniques in convex analysis, regret minimization, conformal prediction, and distributionally-robust optimization.
Critical weaknesses
  • This paper introduces many expressions throughout, some of which do not seem necessary to me (e.g., introducing AA for Ψ∗\Psi^*). It does hamper readability to need to learn LΨL_\Psi, the gap in the Fenchel inequality, ρU\rho_U, R∗(U)R^*(U), WU(x)W_U(x), and such. It is just a suggestion, but perhaps writing out these expressions rather than introducing new functions may improve readability, as it is hard to keep all of these variables in mind at once. Additionally, the suppression of the dependence on zz actually hampered readability for me, and keeping it explicit throughout the paper would make it more clear whether the results hold on average or conditionally.
  • The results seem to rely very strongly on the properties of the convex conjugate; it would be nice to discuss which parts of the results will generalize to problems other than calculating the convex conjugate. For example, Theorem 1 is a direct consequence of the form of the objective function, where θ\theta only enters the objective linearly, and, from that perspective, is not too surprising.
  • The numerical results show that the min-max optimization performs nicely, but it would be good to see more empirical results that validate the theory of the paper as well.
Questions for authors
  • Can you clarify more details regarding the reflection pattern in Equation (7)? It is unclear to me why the weighting mentioned should be used.
  • Can you provide proof sketches for the results? From my reading, many of the proofs follow from standard methods in conformal prediction, or convex analysis; I want to make sure that I am not missing something critical. For example, how critical is it that the objective is linear in θ\theta? The proof itself does not seem to make it easily clear how essential some assumptions are. For example, in Equation (5), it seems the fact that the supremum over BηB_\eta equals η\eta seems to be directly by definition of BηB_\eta. Can it be clarified why this is a surprising result?
5 comments
AuthorsOpenReview ↗

Response to Reviewer EntJ (Part I)

We appreciate Reviewer EntJ's positive comments on the strength of this work and the constructive review of the paper. We hope our response addresses the questions raised. We welcome further discussion.

W1: This paper introduces many expressions throughout, some of which do not seem necessary to me (e.g., introducing AA for Ψ∗\Psi^{\ast}). It does hamper readability to need to learn LΨL_{\Psi}, the gap in the Fenchel inequality, ρU\rho_U, R∗(U)R^{\ast}(U), WU(x)W_U(x), and such. It is just a suggestion, but perhaps writing out these expressions rather than introducing new functions may improve readability, as it is hard to keep all of these variables in mind at once. Additionally, the suppression of the dependence on zz actually hampered readability for me, and keeping it explicit throughout the paper would make it more clear whether the results hold on average or conditionally.

Response: We appreciate the suggestion that the notation can be simplified. We have removed redundant aliases such as AA and now use Ψz∗\Psi_z^{\ast} directly. We also write out expressions used only locally. For recurring quantities, we retain concise notation with brief reminders of their meaning at key uses. Context dependence is explicit in definitions and theorem statements, with abbreviated notation confined to arguments that explicitly fix zz. The deterministic structural results apply at a fixed context, whereas Theorem 3 and Corollary 2 provide marginal guarantees over calibration and test pairs.

We have also reorganized Appendix A to improve readability: A.1–A.2 collect the main proofs in manuscript order, while A.3–A.5 present supplementary results and examples.

W2: The results seem to rely very strongly on the properties of the convex conjugate; it would be nice to discuss which parts of the results will generalize to problems other than calculating the convex conjugate. For example, Theorem 1 is a direct consequence of the form of the objective function, where θ\theta only enters the objective linearly, and, from that perspective, is not too surprising.

Response: We agree that several proofs follow standard arguments in convex analysis and convex conjugacy. Theorem 1 relies on linear coupling; its significance is that, under the stated conditions, minimax-regret optimization over a direct-regret set uniquely returns PTO. This identifies a limitation of direct-regret calibration and motivates our reflected score. Our study focuses on convex decision problems with uncertainty entering linearly in the objective, and our contribution lies in using these established tools to design and analyze decision-aware uncertainty calibration methods supporting robust refinement.

The following arguments also apply to convex decision problems with a common feasible set, even when the loss depends nonlinearly on uncertainty:

  • Certification and coverage (Prop. 1; Thm. 3; Cor. 2): regret-quantile certification, score ranks under exchangeability, and coverage-to-regret transfer.
  • Decision convexity (Prop. 2): the supremum of convex regret functions remains convex.
  • Optimality and gain (Thm. 4; Prop. 3): active-gradient conditions extend under compactness and smoothness; uniform strong convexity gives analogous displacement and gain bounds.
  • Computation (Prop. 4): restricted-scenario lower bounds and global pricing upper bounds certify optimality error.

The reflected three-gap identity and parameter-barycenter formula (Thms. 2 and 5) use linear coupling and a shared Ψ\Psi. Nonlinear transformations of raw uncertainty are also accommodated when they enter as objective coefficients.

AuthorsOpenReview ↗

Response to Reviewer EntJ (Part II)

W3: The numerical results show that the min-max optimization performs nicely, but it would be good to see more empirical results that validate the theory of the paper as well.

Response: We agree that experiments validating theoretical predictions are valuable. The submitted manuscript already includes the following checks:

  • Self-centering and switching (Theorem 1; Corollary 4): On 5,000 two-asset LP contexts at α=0.05\alpha=0.05, direct-regret calibration retains PTO. FY changes 423 decisions when the absolute predicted return gap falls below the switching threshold, reducing worst-case regret over the same FY set by 0.272 on average among those decisions (Figure 3).

  • Optimizer-region geometry (Theorem 4): The exact two-asset QP diagnostic at α=0.1\alpha=0.1 retains PTO for all 43 single-region FY sets; 919 of 957 region-crossing instances have robust gain (Appendix B.3, Table 8 in the revised manuscript).

The revised manuscript adds the following ablations and diagnostics to examine the theoretical mechanisms:

  • Score ablation and boundary release (Proposition 5): With each score separately calibrated at λ=2\lambda=2 and α=0.05\alpha=0.05, two-gap calibration releases 0 of 840 binding nominal bounds in the QP, while full FY releases 838 (99.76%). The LP ablation likewise shows decision movement only with full FY (Tables 5 and 9, Appendices B.2–B.3).

  • Refinement under prediction bias: In a controlled two-asset QP at α=0.05\alpha=0.05, full FY moves the allocation toward the known conditional-mean optimum. At the same empirical coverage of 95.06%, it reduces mean regret and CVaR95 by 29.40% and 18.89% relative to PTO and two-gap calibration (Table 10, Appendix B.3).

  • Three-gap decomposition and regret certificates (Theorem 2; Corollary 2): Across all 1,000 portfolio instances (κ=2\kappa=2, α=0.1\alpha=0.1), the mean FY certificate is 68.64% below the score threshold; reverse and reflected gaps explain 97.14% of this tightening. Empirical set and regret coverage are 89.30% and 95.00%; every score-covered outcome satisfies the tighter bound (main-text Table 3; Appendix B.3, Eq. (41) and Table 13).

We have included these results in the revised version.

Q1: Can you clarify more details regarding the reflection pattern in Equation (7)? It is unclear to me why the weighting mentioned should be used.

Response: Thank you for pointing this out. We agree that the current presentation may obscure the order of our construction. The reflection weights follow from our FY construction. For a chosen relative weight λ>1\lambda>1, we seek a single scaled FY representation of the two directional losses: LΨ(θ;x^)+λLΨ(θ^;xθ)=⟨λθ^−θ,x^−xθ⟩−(λ−1)[Ψ(x^)−Ψ(xθ)]=(λ−1)[LΨ(rλ;xθ)−LΨ(rλ;x^)].L_\Psi(\theta;\widehat x)+\lambda L_\Psi(\widehat\theta;x_\theta)=\langle\lambda\widehat\theta-\theta,\widehat x-x_\theta\rangle-(\lambda-1)[\Psi(\widehat x)-\Psi(x_\theta)]=(\lambda-1)[L_\Psi(r_\lambda;x_\theta)-L_\Psi(r_\lambda;\widehat x)].

Factoring out λ−1\lambda-1 gives rλ=(λθ^−θ)/(λ−1)r_\lambda=(\lambda\widehat\theta-\theta)/(\lambda-1). Since the subtracted gap is nonnegative, we calibrate Sλ=(λ−1)LΨ(rλ;xθ)S_\lambda=(\lambda-1)L_\Psi(r_\lambda;x_\theta), which upper-bounds the weighted losses.

The reflection also has a direct decision interpretation: θ^=θ/λ+(λ−1)rλ/λ\widehat\theta=\theta/\lambda+(\lambda-1)r_\lambda/\lambda lies between the two parameters. The nominal and reflected gaps assess the nominal decision’s suboptimality on both sides, while the reverse gap measures the opportunity cost of adopting the hindsight decision under the prediction. Moreover, at λ=2\lambda=2, reflection is symmetric and self-inverse, and the score recovers the SPO+ loss of Elmachtoub and Grigas (2022) in LP cases (Corollary 1).

References

  • Elmachtoub, A. N., & Grigas, P. (2022). Smart “Predict, then Optimize”. Management Science, 68(1), 9–26.
AuthorsOpenReview ↗

Response to Reviewer EntJ (Part III)

Q2: Can you provide proof sketches for the results? From my reading, many of the proofs follow from standard methods in conformal prediction, or convex analysis; I want to make sure that I am not missing something critical. For example, how critical is it that the objective is linear in θ\theta? The proof itself does not seem to make it easily clear how essential some assumptions are. For example, in Equation (5), it seems the fact that the supremum over BηB_\eta equals η\eta seems to be directly by definition of BηB_\eta. Can it be clarified why this is a surprising result?

Response: Linearity in θ\theta defines the problem setting we study: it enables the FY representation of regret and underpins the self-centering and reflection results. For Eq. (5), yes, the definition gives an upper bound of η\eta, and the stated assumptions yield equality; the key conclusion is that PTO uniquely minimizes worst-case regret.

We give proof sketches in manuscript order below and have added sketches for key results to the main text.

  • Theorem 1 (Self-centering). In case (i), for any x≠x^x\ne\widehat x, minimize ⟨θ,x−x^⟩\langle\theta,x-\widehat x\rangle over BηB_\eta. Compactness and interiority yield an adverse boundary point where continuity gives nominal regret η\eta; linear coupling and nominal optimality make the alternative decision’s regret strictly larger. Positive homogeneity gives the analogous argument for LPs. Thus PTO is the unique minimax-regret optimizer.

  • Proposition 1 (Nominal certification). Any set with coverage at least 1−α1-\alpha must have worst-case nominal regret at least the corresponding regret quantile; the regret sublevel attains this lower bound. Nondecreasing transformations also yield regret sublevels.

  • Theorem 2 (Three-gap decomposition). Expand the score using Fenchel equality at xθ,x^x_\theta,\widehat x and (λ−1)rλ=λθ^−θ(\lambda-1)r_\lambda=\lambda\widehat\theta-\theta. Linear coupling and the shared Ψ\Psi give the three nonnegative gaps.

  • Corollary 1 (SPO+). With Ψ=IX\Psi=I_{\mathcal X} and λ=2\lambda=2, rewriting return maximization as cost minimization transforms the support-function expression into SPO+ for the same selected hindsight optimizer.

  • Theorem 3 (Joint conformal guarantee). Fix the score before calibration. Exchangeability gives a uniform test rank after tie-breaking, so the corrected quantile gives coverage. The nonnegative gaps inherit simultaneous bounds; strong convexity converts the reverse-gap bound into optimizer localization.

  • Proposition 2 (Convexity). The FY gap is convex in each argument separately. A supremum preserves decision convexity; Jensen bounds the gap at a parameter mixture by the largest component gap. Thus convexifying UU preserves the robust value.

  • Corollary 2 (Marginal regret guarantee). On the coverage event, realized regret is at most worst-case regret. Nominal feasibility and the score bound give R∗(U)≤ρU(x^)≤ηR^*(U)\le\rho_U(\widehat x)\le\eta. A certified ε\varepsilon-optimal policy inherits the bound η+ε\eta+\varepsilon.

  • Theorem 4 (Robust optimality). Compactness of UU and continuity of Ψ∗\Psi^* give attained worst cases. The maximum and sum rules yield ∂ρU(x)=∂Ψ(x)−conv⁡WU(x)\partial\rho_U(x)=\partial\Psi(x)-\operatorname{conv}W_U(x). Optimality, 0∈∂ρU(x)0\in\partial\rho_U(x), gives the active-hull condition.

  • Corollary 3 (Finite screening and descent). Caratheodory gives a finite active-scenario certificate. Local finiteness of Ψ\Psi makes the directional derivative a subgradient support function; minimizing it over the unit ball gives rate −M-\mathfrak M.

  • Corollary 4 (Reverse-gap screening). The margin Γ\Gamma makes positive nominal regret imply Sλ>λΓS_\lambda>\lambda\Gamma. Thus η≤λΓ\eta\le\lambda\Gamma gives zero nominal worst-case regret.

  • Proposition 3 (Value of robustification). Theorem 4 gives positive gain exactly when M>0\mathfrak M>0. Strong convexity at the robust optimum gives the lower bound μ∥xrob−x^∥2/2\mu\|x^{\rm rob}-\widehat x\|^2/2; at the nominal decision, completing the square gives the upper bound M2/(2μ)\mathfrak M^2/(2\mu).

  • Theorem 5 (Adversarial barycenter). Fenchel inequality after averaging scenario regrets gives the Jensen-gap lower bound. At an attained robust optimum, Theorem 4 supplies a finite mixture of active parameters with mean in ∂Ψ(xrob)\partial\Psi(x^{\rm rob}); Fenchel equality attains the bound.

  • Proposition 4 (Scenario-generation certificate). Restricted scenarios give a lower bound; full-set worst-case evaluation at a feasible decision gives an upper bound. Their gap certifies optimality error; a global pricing upper bound suffices.

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

The paper studies decision-aware conformal calibration in the context of convex optimizations and proposes a reflected Fenchel-Young (FY) score that incorporates nominal regret, a reverse optimality gap, and a reflected stability gap. The calibrated sets produced by this approach provide marginal guarantees and enable minimax-regret decision refinement. Additionally, the paper presents structural characterizations and a scenario-generation algorithm, with validation on simplex allocation, synthetic portfolio QPs, and a real-data solar allocation task.

Critical strengths

This paper investigates an interesting problem and addresses the mismatch between prediction error and decision loss. Constructing calibrated uncertainty sets that better serve downstream optimization is a relevant research direction with practical value.

Critical weaknesses
  1. The central design of the reflected FY score in (8) lacks sufficient justification. The paper does not adequately explain the design rationale and the meaning of rλr_\lambda in (7), or why this specific reflection is suitable for calibration. Although coefficient matching clarifies how the expression is derived, it does not justify the choice of this particular construction. Similarly, by excluding the nonnegative term in (7), an upper bound is established, but there is no explanation for why this bound should be used as the score or why all three components in (9) are necessary, especially the additional reflected stability gap.

  2. The advantages of the FY score over the residual score or the unique benefits of reflection are not yet adequately established, which is the motivation of this paper. The experiments alter both the calibration score and the downstream optimization objective, making it challenging to isolate the score's impact. Additionally, there is a lack of a direct ablation that evaluates the added stability gap from reflection. These limitations leave the practical value of the proposed score insufficiently supported.

Questions for authors
  1. In the reflection construction, what is the design rationale and special significance of rλ(θ,θ^)r_{\lambda(\theta,\hat\theta)} in (7) beyond coefficient matching? Why is this specific reflection essential for comparing decisions and constructing calibrated sets?

  2. For constructing the FY score, why use the upper bound and all three score components in (8), instead of directly calibrating Reg⁡(x^;θ)+λΔθ^(xθ)\operatorname{Reg}(\hat x;\theta)+\lambda\Delta_{\hat\theta}(x_\theta)? What is the rationale behind penalizing the instability of the nominal decision under the reflected parameter? Additionally, in what situations does this penalty lead to an actual improvement in decision quality, rather than merely producing a set that permits the nominal decision to move?

  3. Are there theoretical advantages to residual calibration that might be enhanced by reflection? At the same level of coverage, which results support that FY offers an advantage over residual scores?

  4. Comparative results between FY and residual sets should be provided in experiments using the same minimax-regret objective.

5 comments
AuthorsOpenReview ↗

Response to Reviewer fmoT (Part I)

We genuinely appreciate Reviewer fmoT for recognizing the relevance and practical value of decision-aware calibration. We are very happy to clarify your concerns and engage in further discussions. We address W1, Q1, Q2, and W2 first, followed by Q3 and Q4, to group related comments.

W1: The central design of the reflected FY score in (8) lacks sufficient justification. The paper does not adequately explain the design rationale and the meaning of rλr_\lambda in (7), or why this specific reflection is suitable for calibration. Although coefficient matching clarifies how the expression is derived, it does not justify the choice of this particular construction. Similarly, by excluding the nonnegative term in (7), an upper bound is established, but there is no explanation for why this bound should be used as the score or why all three components in (9) are necessary, especially the additional reflected stability gap.

Response: We agree that the current presentation obscures the order of our construction. Our starting point is the FY optimality gap LΨ(θ;x)L_{\Psi}(\theta;x), which measures suboptimality of xx under θ\theta. Nominal regret is exactly LΨ(θ;x^)L_\Psi(\theta;\widehat x), whose conformal sets retain PTO decisions under Theorem 1's conditions. To retain information about alternative decisions, we also consider LΨ(θ^;xθ)L_\Psi(\widehat\theta;x_\theta), the hindsight optimizer's suboptimality under the prediction. We therefore seek one scaled FY gap whose comparison of the two optimizers captures both directional losses, for λ>1\lambda>1: LΨ(θ;x^)+λLΨ(θ^;xθ)=⟨λθ^−θ,x^−xθ⟩−(λ−1)[Ψ(x^)−Ψ(xθ)]=(λ−1)[LΨ(rλ;xθ)−LΨ(rλ;x^)].L_\Psi(\theta;\widehat x)+\lambda L_\Psi(\widehat\theta;x_\theta)=\langle\lambda\widehat\theta-\theta,\widehat x-x_\theta\rangle-(\lambda-1)\bigl[\Psi(\widehat x)-\Psi(x_\theta)\bigr]=(\lambda-1)\bigl[L_\Psi(r_\lambda;x_\theta)-L_\Psi(r_\lambda;\widehat x)\bigr].

Factoring out λ−1\lambda-1 gives rλ=(λθ^−θ)/(λ−1)r_\lambda=(\lambda\widehat\theta-\theta)/(\lambda-1) in the effective parameter space. Thus we calibrate Sλ=(λ−1)LΨ(rλ;xθ)S_\lambda=(\lambda-1)L_\Psi(r_\lambda;x_\theta), which upper-bounds the weighted sum of both directional losses. The three-gap decomposition follows from this construction, so the reflected stability term is derived rather than introduced as a separate penalty. We have revised Section 2.3 to make this design order more explicit.

Q1: In the reflection construction, what is the design rationale and special significance of rλ(θ,θ^)r_\lambda(\theta,\hat\theta) in (7) beyond coefficient matching? Why is this specific reflection essential for comparing decisions and constructing calibrated sets?

Response: The reflected stability term follows from our FY construction rather than being introduced as a separate penalty. The reflection places θ\theta and rλr_\lambda on opposite sides of the prediction, with θ^=1λθ+λ−1λrλ\widehat\theta=\frac{1}{\lambda}\theta+\frac{\lambda-1}{\lambda}r_\lambda.

This gives a direct interpretation of Sλ=LΨ(θ;x^)+(λ−1)LΨ(rλ;x^)+λLΨ(θ^;xθ).S_\lambda=L_\Psi(\theta;\widehat x)+(\lambda-1)L_\Psi(r_\lambda;\widehat x)+\lambda L_\Psi(\widehat\theta;x_\theta). The first two terms measure the nominal decision’s suboptimality on both sides; the reverse gap measures the opportunity cost of adopting the hindsight decision under the prediction.

Reflection is not required for conformal validity. Its role here is to control decision-relevant parameter directions that the two-gap score can leave unrestricted, as shown by Proposition 5.

At λ=2\lambda=2, reflection is symmetric and self-inverse, and the LP score recovers the SPO+ loss of Elmachtoub and Grigas (2022) for decision-focused learning (Corollary 1).

References

  • Elmachtoub, A. N., & Grigas, P. (2022). Smart “Predict, then Optimize”. Management Science, 68(1), 9–26.
AuthorsOpenReview ↗

Response to Reviewer fmoT (Part II)

Q2: For constructing the FY score, why use the upper bound and all three score components in (8), instead of directly calibrating Reg⁡(x^;θ)+λΔθ^(xθ)\operatorname{Reg}(\hat{x};\theta)+\lambda\Delta_{\hat\theta}(x_\theta)? What is the rationale behind penalizing the instability of the nominal decision under the reflected parameter? Additionally, in what situations does this penalty lead to an actual improvement in decision quality, rather than merely producing a set that permits the nominal decision to move?

Response: As clarified in our response to W1, our construction seeks a single FY optimality gap encoding both directional comparisons. Let Tλ=LΨ(θ;x^)+λLΨ(θ^;xθ)T_\lambda=L_\Psi(\theta;\widehat x)+\lambda L_\Psi(\widehat\theta;x_\theta) denote the two-gap score. The reflected term keeps our score in the form of a single FY gap:

=T_\lambda+(\lambda-1)L_\Psi(r_\lambda;\widehat x).$$ Another reason to retain it is the decision geometry it supplies. Although $T_\lambda$ can also be conformally calibrated, it can leave parameter directions uncontrolled and thereby obstruct decision refinement. **Proposition 5 (Appendix A.3).** Let $\Psi=\phi+I_{\mathcal X}$, where $\mathcal X$ is a compact polytope and $\phi$ is continuous and convex. If the nominal optimizer $\widehat x$ is a vertex of $\mathcal X$, then for every $\lambda>1$ and finite $\eta\ge0$, $$\arg\min_{x\in\mathcal X}\sup_{\theta:T_\lambda(\theta,\widehat\theta)\le\eta}L_\Psi(\theta;x)=\{\widehat x\},$$ over the full effective parameter space. This includes LPs with a unique nominal optimizer. Under these conditions, two-gap calibration cannot adjust the nominal decision. The reflected term truncates the unbounded normal-cone rays responsible for this obstruction, allowing alternative decisions to be selected when they improve the robust objective. The controlled prediction-bias experiment illustrates when refinement improves actual decision quality. Nominal-regret, two-gap, and FY calibration all have empirical coverage of 95.06%. FY reduces mean regret by 29.40% and CVaR95 by 18.89% relative to PTO and two-gap calibration, moving the allocation toward the known conditional-mean optimum (Table 10, Appendix B.3). > **W2:** The advantages of the FY score over the residual score or the unique benefits of reflection are not yet adequately established, which is the motivation of this paper. The experiments alter both the calibration score and the downstream optimization objective, making it challenging to isolate the score's impact. Additionally, there is a lack of a direct ablation that evaluates the added stability gap from reflection. These limitations leave the practical value of the proposed score insufficiently supported. **Response:** We added both a direct ablation and an objective-matched comparison. The two-asset QP ablation ($\lambda=2$, $\alpha=0.05$) separately calibrates nominal regret, $T_\lambda$, and full FY on the same observations and uses minimax regret for all three: $T_\lambda$ releases 0 of 840 binding nominal bounds, while full FY releases 838 (99.76%) (Table 9, Appendix B.3). This supports the structural result in Q2. Q4 reports the objective-matched performance comparison with residual calibration (Table 2, Section 4).
AuthorsOpenReview ↗

Response to Reviewer fmoT (Part III)

Q3: Are there theoretical advantages to residual calibration that might be enhanced by reflection? At the same level of coverage, which results support that FY offers an advantage over residual scores?

Response: The advantage over residual calibration depends on the problem and uncertainty distribution. Decision-aware calibration may reduce unnecessary conservatism when the magnitude of prediction errors poorly reflects their impact on decision quality. The following example uses the two-asset allocation model from our experiments to establish strictly lower regret at identical coverage under the stated conditions.

Example 2 (Appendix A.5). Consider a two-asset allocation LP with fixed predicted return difference h>0h > 0. Let calibration and test observations be iid copies of a continuous return difference DD, with D>2hD > 2h almost surely and E[D]<∞\mathbb{E}[D] < \infty. In the effective return-difference coordinate, separately calibrate FY (λ=2\lambda=2) and residual ℓ2\ell_2 (score ∣D−h∣|D-h|) on the same nn calibration observations. Let qq be the kk-th calibration order statistic of DD, with k=⌈(n+1)(1−α)⌉≤nk=\lceil(n+1)(1-\alpha)\rceil\le n.

The calibrated sets have the same coverage event Dtest≤q\\{D_{\mathrm{test}}\le q\\}, but minimizing worst-case regret selects different allocations:

pFY=qq+(q−4h)+>q2(q−h)=pℓ2.p_{\mathrm{FY}}=\frac{q}{q+(q-4h)_{+}} > \frac{q}{2(q-h)}=p_{\ell_2}.

where (a)+=max⁡a,0(a)_{+}=\max\\{a,0\\} and pp is the weight on asset 1. Since every true outcome has D>0D > 0,

Reg⁡FY(D)=(1−pFY)D<(1−pℓ2)D=Reg⁡ℓ2(D).\operatorname{Reg}_{\mathrm{FY}}(D)=(1-p_{\mathrm{FY}})D < (1-p_{\ell_2})D=\operatorname{Reg}_{\ell_2}(D).

Thus FY yields strictly lower mean regret and CVaR for the calibrated policies. The residual interval [2h−q,q][2h-q,q] extends farther into the region where asset 2 appears preferable than the FY interval [min⁡0,4h−q,q][\min\\{0,4h-q\\},q]. Residual minimax regret therefore puts more weight on the inferior asset. FY avoids this unnecessary hedging.

The portfolio experiments provide complementary evidence of relative performance advantages for FY sets under a common minimax-regret objective (Table 2, Section 4).

Q4: Comparative results between FY and residual sets should be provided in experiments using the same minimax-regret objective.

Response: We agree that a common downstream objective helps isolate the effect of the uncertainty-set construction. We have added a portfolio comparison using minimax regret for every calibrated uncertainty set, with the predictor, data, feasible set, and nominal 90% coverage target held fixed (Table 2).

MethodMean regret ×103\times10^3CVaR95 ×103\times10^3Coverage
PTO3.1727.307—
FY2.8886.90489.30%
Residual L23.2597.42090.50%
Residual box4.1798.69590.50%
Residual ellipsoid3.2637.42390.50%
KNN RO3.3747.60589.90%

Relative to L2, box, and ellipsoid, FY lowers mean regret by 11.37%, 30.88%, and 11.48%, and CVaR95 by 6.96%, 20.60%, and 6.99%.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.