acceptodds
Under review as a conference paper at ICLR 2027

FACET: Geometry and Guarantees at the Conformal Prediction-Box Interface

Abstract

Multivariate conformal methods are calibrated and compared as prediction regions, while many downstream interfaces expose only coordinatewise lower and upper bounds: the axis-aligned box of that region. These are different prediction tasks. We characterize the resulting interface. For any nested box family, exact box membership depends only on its coordinate-envelope path, so two arbitrarily different native regions with the same envelopes become the identical box predictor after exact recalibration; for homothetic families this reduces the geometry to its axial support vector, and for Mahalanobis ellipsoids to . Containing conversions preserve coverage but can be asymptotically vacuous: the box of a calibrated Gaussian ellipsoid has coverage tending to one and is wider than a valid centered box. Finite non-containing conversions admit no distribution-free transfer and can instead lose nearly all coverage. We test these predictions in a locked, external-only benchmark on high-dimensional forecasting panels. At the native threshold, containing conversions add - percentage points of coverage, whereas a rank-truncated non-containing conversion loses points in median. The box-task winner differs from the native-region winner in of eligible cells. Exact online box calibration reduces median absolute coverage error from to percentage points at a median width ratio of , and equal envelope paths yield equal boxes to . The strongest external geometry remains wider in median than a validation-selected coordinatewise-support box. On four untouched panels, time-varying coordinate supports reduce width by over static supports, while pooled support models win of comparisons. The result is a task-faithful evaluation principle and a simple prescription: predict coordinate supports, then calibrate exact membership in the box that is actually delivered.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Conformal prediction, uncertainty quantification, multivariate forecasting, prediction sets, online calibration, high-dimensional prediction

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_A96jOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2(+1)
Read review
Paper summary

This paper investigates the challenges associated with converting general high-dimensional multivariate prediction regions into box-shaped prediction sets. The authors point out that such a conversion may lead to excessive coverage, resulting in overly conservative prediction sets, or may invalidate the original coverage guarantees. To address these issues, the authors propose directly calibrating nested box-shaped prediction sets to obtain smaller boxes while maintaining coverage guarantees. Furthermore, they extend this approach to the online setting by incorporating existing adaptive conformal calibration methods and further improve prediction efficiency by estimating coordinate-wise scales.

Critical strengths

The paper clearly highlights the distinction between the original prediction regions and the resulting box-shaped prediction sets, and theoretically characterizes the potential issues arising from this conversion.

Critical weaknesses

(1) Limited practical relevance of the problem setting.\textbf{Limited practical relevance of the problem setting.} The problem considered in this paper appears somewhat artificial. If the goal is to construct box-shaped prediction sets, a more natural approach would be to calibrate such sets directly, rather than first constructing general prediction regions and then converting them into boxes. Although the paper provides several theoretical results, the practical significance of addressing this conversion problem remains unclear.

(2) Limited novelty of the online extension.\textbf{Limited novelty of the online extension.} The extension to the online setting appears largely incremental. It mainly adopts an existing adaptive conformal calibration approach similar to that of Gibbs et al., without introducing substantial methodological innovations. Therefore, I do not consider this extension a significant contribution of the paper.

Questions for authors
  1. The Introduction does not clearly explain why box-shaped prediction sets are needed or why prediction regions of other shapes need to be converted into boxes. Could the authors provide more concrete examples or practical motivations for considering such a conversion?

  2. The presentation of the paper could be improved. The authors introduce too much mathematical notation, making the paper difficult to follow. Moreover, several theoretical results, particularly Theorem 2, are not stated clearly. I suggest simplifying the notation and presenting the main theorems in a more concise and accessible manner.

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

In this paper, the author investigates what happens to a multivariate conformal guarantee when the downstream system only accepts per-coordinate lower and upper bounds, that is, the axis-aligned box of the calibrated region. On the theory side, the authors show that coverage transfers distribution-free only under containment. Moreover, they prove that exact box membership depends only on the coordinate-envelope path (Theorem 2), so for an ellipsoid only diagΣ\sqrt{\mathrm{diag}\Sigma} survives. In high dimension, containing conversions become very conservative (Theorem 4), while finite summaries of rank-deficient regions inherit no guarantee (Corollary 3). The proposed method predicts per-coordinate scales from a small EWMA/seasonal library, scores with the max-normalized residual, and tracks the threshold online. The evaluation separates four stages (native region, its box at the same threshold, split recalibration, online recalibration) across 9 forecasting panels and 4 held-out ones. It reports large conversion distortion, a different winner in 69.3% of cells, online coverage error of 0.10 points. The strongest external geometry remains 28% wider than the selected coordinate-support box.

Critical strengths

In my opinion, the main upside of this paper is the framing and the four-stage evaluation. Comparing the native region and its box at the same threshold does show what the conversion itself does, without blaming published methods for failing a task they weren’t built for. Table 1 shows the three regimes clearly. I think this protocol is the most reusable part of the paper. The authors are also very honest in reporting the criteria and failures. Even the fact that independent re-implementation caught a bug in the reversal count is included in the paper.

As far as I am concerned, the central theory also correct, and the experiments successfully examined its predictions. I checked Lemma 1, Theorem 2, Corollary 3 and Theorem 4 under the stated assumptions and found no noticeable problems. The Theorem 4 ratio of about 11.4 at N = 2048 is reproducible. I would say Theorem 2 is not that deep, but its consequence is clean: fitted off-diagonal covariance is invisible at a box interface. The matched-support experiment gives a convincing validation of this.

Critical weaknesses

My first concern is the native-stage baselines, which feed the 69.3% winner-change and 36.2% reversal claims. When ⌊α(n+1)⌋<N\lfloor \alpha(n + 1)\rfloor < N, DCP-MT falls back to the coordinatewise sample maximum (116 of 306 records). That box has no joint guarantee. Specifically, with 99 calibration points and 20 independent coordinates, joint coverage is about 0.82 at nominal 0.9. Its stage-A coverage of 0.816, against 0.895 for MaxNorm on the same data, is consistent with this. Bonferroni’s order statistic is undefined under exactly the same condition, and the paper does not say what it emits there. Native coverage also ranges from 0.689 to 0.909 while every method sits at 0.900 at stage D. If native efficiency is ranked without accounting for coverage, some “reversals” just reflect undercoverage rather than the interface. The −86.3 point MultiDimSPCI result also needs a precise source for its coordinate summary.

My second concern is that the efficiency result mostly measures temporal volatility modeling. By the paper’s own Theorem 2, most baselines become max-normalized boxes with static, design-fitted scales at stage D. Online ExactBox, by contrast, gets validation-selected EWMA scales. Thus, the 28% gap looks like the same static-versus-time-varying gain (30.2%) that Section 5.4 reports for the method against itself. MaxNorm uses exactly the same score, so with the same supports the two would coincide. What remains is EWMA volatility, a max score and an ACI-style update, all fairly standard. Since the theory is short, the paper’s case rests mainly on the framing.

It seems like the writing of this manuscript is heavily influenced by LLMs, with some jargons ("provenance", "interface", etc.) that may cause readers some trouble understanding. I understand that the use of AI is allowed, but the authors should really consider spending some time writing this paper themselves. The paper would greatly benefit from human narration and explanation of its finding, both theoretically and empirically.

Smaller issues: Proposition 11’s conclusion is backwards. Once the clipped branch is vacuous, the region is within the emitted box, and its coverage cannot be higher. Appendix G describes a different box. In Appendix E, the “at most 59.2%” binomial bound is already exceeded by the observed 79.9%, so the failed criterion should simply be reported as failed. Equation (13) needs WpW_p defined, and A.6 should read −log⁡(1−α)/N-\log(1 - \alpha)/N rather than α/N\alpha/N. Table 1 is also missing the unbounded-output fractions promised in A.8.

Questions for authors

1 In the 116 flagged records, what do DCP-MT and Bonferroni emit? How do the ranking statistics change if those records are excluded or given infinite bounds?

2 How is native efficiency ranked when coverage differs across methods? What makes a cell “eligible”?

3 What scale does MaxNorm use? Is any part of the 28% gap due to geometry rather than to static versus time-varying scales?

Reviewer_x5Z6OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the problem of providing coordinatewise lower and upper bound for multi-variate prediction. It identifies 4 prediction tasks - the native region, the axis-aligned envelope, calibrated envelope using box membership and online box calibration.

The main theorem states that exact calibration at this interface can see a native region only through its path of coordinate envelopes. For homothetic regions, only axial supports survive; for Mahalanobis ellipsoids, these are the square roots of the covariance diagonal. The main mathematical claims appear to be correct and the empirical study has good coverage. The primary concern witht he empirical study is whether the gains highlighted emerge from FACET or from temporal adaptation and tuning.

Critical strengths
  • The paper correctly emphasizes that coverage belongs to a specific set. A native multivariate region and its displayed coordinate box are different prediction tasks. The four-stage decomposition makes this clear.
  • The formal results Lemma 1, Theorem 2, Corollary 3 and Theorem 4 appear to be correct with understandable proofs
  • The expertimanetal design is good - the use of chronological blocks to separate design, calibration and test is clean. The study includes two backbones, three coverage levels, multiple origins, nine primary panels and four untouched confirmation panels.
  • The paper is well-written - qualifications are generally placed near claims, and the appendices contain the implementation and source details needed to understand the benchmark. The paer documents negative results clearly e.g. when containing conversions become vacuous, when the authors' original structured construction fails since this help with understand the arguments and results.
Critical weaknesses
  • The paper claims that “the strongest published geometry remains 28% wider” than its reference and that coordinatewise support prediction supplies the dominant efficiency gain. However it is unclear if the gains are from temporal tuning. For example, online ExactBox selects on validation among Static, EWMA at three half-lives, and seasonal EWMA at two half-lives—support candidates with time-varying per-coordinate scales. Comparing static covariance diagonals against adaptive coordinatewise scales cannot determine whether off-diagonal geometry is useless once marginal dynamics are matched.
  • The novelty in the theory is more in the synthesis of existing results rather than. in discovering new geometry. “Main result” and “complete invariant” are correct, but the paper should position these as a task/interface theorem and synthesis rather than implying that difficult geometric structure was uncovered. A fuller discussion of prior simultaneous conformal bands and standardized-max scores would sharpen the novelty boundary.
Questions for authors

provide ablation study that compare, using identical centers, candidate budgets and calibration:

  1. adaptive diagonal supports;
  2. adaptive marginal scales plus static correlation;
  3. adaptive marginal scales plus time-varying factor/correlation structure;
  4. the same marginal support path with and without native interior geometry.

The theorem guarantees equivalence only when envelope paths agree; it does not say rich native models cannot predict better envelope paths. The current experiments support “adaptive coordinate supports win in this benchmark,” but this does not imply dependence models do not help at a box interface?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.