acceptodds
Under review as a conference paper at ICLR 2027

Online Forecast Aggregation with Quantile-Anchored Tail Protection

Abstract

A source forecast can improve expected continuous ranked probability score (CRPS) while substantially distorting the probability assigned to extreme events. We propose a quantile-anchored source transfer method (QAST), which lets source models reshape a baseline forecast only within cells defined by its tail quantiles, while preserving the baseline's exceedance probability and protected quantiles. Any convex combination of these forecasts keeps the combined-ratio supremum distance (CRSD) at a fixed threshold within a user-chosen radius of the baseline's CRSD, for every outcome sequence and every choice of weights. Within the protected class, we establish sublinear and constant regret for the aggregate forecast's finite-node cell-conditional square score against fixed transferred experts. An explicit iid construction isolates within-cell tail-shape misspecification and shows improvement in expected CRPS and limiting CRSD. On EUPPBench, QAST reduces annual proper-score losses relative to unprotected mixtures using the same weights under two baselines.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Probabilistic forecasting, Online learning, Forecast aggregation, Tail calibration, Distributional transfer

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_j7rwOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
3
Read review
Paper summary

Suppose you have an existing forecaster that you trust to make good predictions about extreme events, but may be sub-optimal elsewhere. You also have access to other models that contain useful information, but might make errors on the extreme events. How should you combine them? Averaging the forecasts can destroy the predictive power of the original model on the extreme events. The solution in this paper is to fix quantiles of the baseline predictor, and allow the new models to update the forecast distribution between those quantiles. This is broadly sensible if you have reason to believe the selected baseline predictor quantiles.

For example given a collection of quantiles, this partitions the unit interval into a set of intervals of fixed probability. We freeze this distribution over the intervals, but then set the distribution within each interval to be the conditional distribution of a source model within that interval.

The paper uses a diagnostic measure called CRSD, and the main theorem controls how much updating the base predictor using the proposed algorithm changes this metric. It does not establish that the base predictor (or the updated predictor) is good in any absolute sense --- the premise is that the goal is to avoid substantially changing this metric. The proposed algorithm is just to optimize conditional squared error (Brier score) within each quantile interval, conditional on lying in that interval.

Critical strengths

The paper gives a well defined problem and the results seem to be correct.

Critical weaknesses

This is a very poorly written paper that is extremely difficult to understand. I suspect it is because very little care has been put into writing up the result. On a hunch given the poor quality of writing, I ran both the abstract and the introduction through pangram, and unsurprisingly it came back as 100% AI generated. This goes well beyond the copy-editing the authors suggest was used for the exposition in the AI disclosure. But I do not mean to AI-shame --- I do not think this is the main concern. Ultimately I do not think the authors have taken the care to make their result readable, and as a result the paper as written has very little value. I cannot recommend acceptance.

There are other strange expositional choices that make me question whether the paper was written with any care. For example, the body of the paper foregrounds an algorithm event-doubling QAST-E which obtains a sub-optimal regret bound, whereas its own Theorem B.1 gives a different algorithm with a substantially better regret bound. It is unclear if there is any reason to prefer the algorithm in the body --- the paper seems to give none.

The paper also gives essentially no evidence that the proposed algorithm improves over simply retraining the baseline. In the main experiment, the differences from the baseline seem to be essentially zero, with confidence intervals overlapping zero.

Questions for authors

N/A

1 comment

Thank you for your suggestions on exposition, algorithm presentation, and the choice of comparisons. We have reorganized the narrative around preserving specified baseline probabilities while using the conditional shapes of source forecasts, and added direct comparisons of fixed-rate and event-doubling weights. The revision makes the connection between the construction, theoretical guarantees, and practical gains more explicit.

  1. Readability and AI-use disclosure

The revised abstract and introduction first explain the roles of the baseline and sources, then introduce the results and learning rules. Section 3.1 describes within-cell shape transfer; Section 3.2 states the anchor-count and CRSD guarantees and points to Appendix A.4 for the proof, while Section 3.3 explains the within-cell PIT displacement bound; and Section 4.2 introduces the fixed-rate aggregate-loss guarantee earlier and explains the role of event-doubling in source-weight analysis. Sections 3.3 and 4.3 now also contain short, blue-highlighted summaries of the anchor-selection and online-learning results, referring to Appendix D.2.5, Table 16, and Appendix D.2.4, Table 15, respectively. Appendix A.3 provides the algorithm, and the new Appendix D.2 connects the additional experiments to the review questions.

Regarding AI use, we used AI to polish the final wording, which may explain the AI-like phrasing you noted.

  1. The different theoretical roles of fixed-rate and event-doubling weights

The revision introduces the constant-regret fixed-rate result earlier in Section 4.2, with the full proof in Appendix B.4, Theorem B.1. Under the full-feedback protocol, a fixed η∈(0,1/2] and no weight resets, exp-concavity of the square loss controls the actual aggregate forecast's cumulative loss relative to a fixed transferred expert.

Event-doubling connects the source-weight analysis to the distributional-improvement theorem. Its bound on the linear weighted expert loss is used in Corollary B.1 and subsequently in Theorem 4.2, explaining how an advantageous source receives greater weight and yields improvements in the full forecast distribution. The actual aggregate square loss equals the weighted expert loss minus a nonnegative diversity term, so these two loss controls serve different proof tasks.

We have added the direct weather replay in Appendix D.2.1, Tables 11–12, and the controlled simulation in Appendix D.2.4, Table 15. Section 4.3 now summarizes the simulated learning gains. With strongly differentiated sources and weights frozen after 800 historical events, both rules have positive pointwise Monte Carlo CRSD improvement intervals relative to static QAST: the mean improvement is 0.02293 for event-doubling and 0.05377 for fixed-rate EW (η=0.5). The weather replay uses η=0.25 as its primary fixed rate; the online increments there are small, with the main empirical gains arising from protection.

QAST therefore supplies the same protected forecast class for both rules. Fixed-rate EW has the stronger aggregate-loss regret order, while event-doubling supports the source-weight and usefulness arguments developed in the paper. Users can select the rule according to their learning objective while retaining the same pathwise protection.

  1. QAST's role and evidence relative to further baseline training

QAST addresses a specific application need: multiple predictive distributions are available, and certain baseline probabilities have a meaning that should be preserved. It transfers source shapes at the distribution level, with explicit preservation constraints and an ex ante perturbation bound. This allows it to serve as a forecast-combination step on top of an existing baseline.

Appendix D.1.3 adds a further-training DRN comparator using physical-scale CRPS plus a tail-calibration penalty, with validation selection of the penalty strength. Table 1 and Appendix Tables 4 and 6 report selection and paired comparisons. At p95, the mean CRPS difference, QAST-E minus the fine-tuned DRN, is only +0.000172 mm, with a pointwise paired 95% interval of [−0.001374, +0.001477] mm; QAST also preserves the baseline probability structure specified by its construction.

The matched unprotected comparison further demonstrates the practical value of protection. At p95 with J=5 and κ=0.75, twCRPS decreases by 27.5% under EMOS and 27.2% under DRN, with paired intervals supporting both reductions. The static configurations in Section 5.3 also illustrate shape-adjustment space within the protected class: retaining two internal tail-quantile anchors gives p90/p95 CRSD point reductions of 9.68%/9.47% relative to the corresponding configuration-specific baseline. Together with the pathwise protection theorem, these results establish the value of QAST as a distributional forecast-combination method.

Reviewer_9XLEOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies online aggregation of probabilistic forecasts when information transferred from source models may improve average predictive scores but distort tail probabilities. The authors propose quantile-anchored source transfer (QAST), which fixes the baseline forecast's exceedance probability and quantile-cell masses while allowing source models to modify the conditional distribution within each cell. A positive baseline component is retained to preserve the crossing events at protected quantiles. Theorem 3.1 establishes a pathwise bound of (1−κ)h(1-\kappa)h on the change in combined-ratio supremum distance (CRSD), uniformly over outcome sequences and aggregation weights. The paper then combines the protected forecasts through exponential weighting. Theorem 4.1 establishes event-scale regret for a finite-node cell-conditional square score, while Theorem 4.2 demonstrates improvements in expected CRPS and limiting CRSD under a specific iid construction. Experiments on EUPPBench and a supplementary WeatherBench 2/METAR panel evaluate the protection mechanism and online aggregation. Below are several specific comments.

Critical strengths
  • The problem is well motivated. Improvements in an overall proper scoring rule do not necessarily imply improvements in extreme-event probabilities. Separating the preservation of selected baseline tail probabilities from the transfer of within-cell conditional shapes addresses a meaningful issue in probabilistic forecast combination.

  • The protected forecast class has a clear and useful structural property. In particular, Theorem 3.1 gives a pathwise CRSD bound for every pre-outcome weight sequence, rather than only for a specific learning algorithm or a particular stochastic data-generating model. The positive baseline-retention device also addresses the possible flat regions of source CDFs at protected quantiles.

  • The paper distinguishes the distributional protection guarantee from the online learning guarantee. It provides both event-doubling and fixed-rate exponential-weighting results, as well as an explicit iid example connecting within-cell prediction improvements to CRPS and CRSD. The empirical study includes matched-weight unprotected mixtures and fixed QAST, which are useful comparisons for identifying the role of the protection step.

Critical weaknesses
  • The incremental methodological and theoretical novelty of the online learning component appears limited. The weight updates use standard exponential weighting under full feedback, and the regret analysis concerns a bounded finite-node square score. The distinctive contribution lies mainly in the baseline-anchored distribution construction; the manuscript could more clearly distinguish the technical novelty of this construction from the established expert-advice machinery.

  • The empirical evidence for the adaptive weighting component is weak. In Table 1, QAST-E, Fixed QAST, and the baseline have nearly identical proper scores at the reported precision. Table 5 and Figure 3 further show very small QAST-E versus Fixed QAST differences, with the reported paired intervals including zero. The substantial gains relative to unprotected mixtures therefore primarily demonstrate the benefit of protection against negative transfer, rather than a meaningful additional benefit from online weight learning.

  • The scope of the theoretical guarantees is narrower than the principal empirical claims. Theorem 4.1 bounds regret for the finite-node cell-conditional square score, whereas the experiments emphasize CRPS, threshold-weighted CRPS, and CRSD. Theorem 4.2 establishes improvement in these forecast-level quantities only under a specially constructed iid model. The paper does not provide a general guarantee translating small cell-score regret into improvements in the reported evaluation metrics.

Questions for authors
  • In Section 5.2, Table 1 shows essentially identical CRPS and twCRPS for QAST-E, Fixed QAST, and the corresponding baselines at three significant figures. More precise results in Table 5 show that the QAST-E--Fixed QAST differences are extremely small and their paired confidence intervals contain zero; Figure 3 gives the same general picture across thresholds and partitions. The authors are encouraged to report the practical magnitude of the online-learning gain more prominently and to investigate settings with longer exceedance histories or more informative differences among the transferred experts. This would help establish when the adaptive updates are preferable to fixed protected averaging.

  • Section 4.1 trains the weights with a finite set of cell-conditional Brier-score nodes, while Section 5 evaluates complete predictive distributions using CRPS, twCRPS, and CRSD. Theorem 4.1 guarantees regret only for the former score, and Theorem 4.2 links it to the latter quantities in a particular iid construction. The authors should discuss conditions under which the finite-node objective is informative about global or tail-weighted proper scores, and preferably include a sensitivity study over the number and placement of score nodes.

  • Theorem 3.1 controls ∣C^A(n)−C^0(n)∣|\widehat C_A(n)-\widehat C_0(n)|, and QAST preserves the baseline exceedance probabilities exactly. Consequently, if the baseline is systematically miscalibrated in the occurrence probability or at the protected quantiles, this error is inherited by every protected aggregate. The paper acknowledges this issue in Section 6, but the empirical evaluation would be strengthened by a controlled experiment varying the magnitude and type of baseline tail misspecification. Such an experiment would clarify the regimes in which exact protection is beneficial and those in which it prevents useful correction.

  • In Section 5.1 and Figure 2, the three choices (J,κ)=(1,0.95),(3,0.85),(5,0.75)(J,\kappa)=(1,0.95),(3,0.85),(5,0.75) keep (1−κ)/J=0.05(1-\kappa)/J=0.05 fixed but vary JJ and κ\kappa simultaneously. Although this is a useful comparison at a common theoretical radius, it does not isolate the effect of either design parameter. A two-dimensional sensitivity analysis varying JJ at fixed κ\kappa and κ\kappa at fixed JJ would clarify the protection--transfer tradeoff, preferably alongside the corresponding empirical CRSD changes.

  • Section 3.1 relates the construction to Jeffrey's rule, while Section 4.2 uses standard exponential weighting. The authors are encouraged to identify more precisely which properties cannot be obtained directly from existing partition-reweighting or forecast-combination procedures, and to explain which part of the proof of Theorem 3.1 requires genuinely new analysis. A concise comparison with a simpler anchored or fixed-weight alternative would make the methodological contribution clearer.

3 comments

Thank you for the specific suggestions on weighting rules, learning nodes, and protection constraints. We have added the corresponding experiments in Appendix D.2.1–D.2.6. Main-text Sections 3.3 and 4.3 now summarize the favorable results on anchor selection and online learning, respectively, with references to the relevant appendices. These additions clarify the method's value: the protected construction delivers significant gains against matched unprotected controls in weather data; informative sources and sufficient feedback enable further learning within the class; and the anchor ablations provide evidence for choosing the protection scope.

  1. The value of online updating and when it is useful

We have added a common comparison of static, fixed-rate, and event-doubling weights in Appendix D.2.1 and Tables 11–12. The controlled simulation varying source information and historical event counts appears in Appendix D.2.4 and Table 15; its principal learning gains are now summarized and cross-referenced in main-text Section 4.3.

The controlled simulation directly demonstrates a setting in which online learning is useful. With strongly differentiated sources and weights frozen after 800 historical exceedances, mean CRSD is 0.11938 for static QAST, 0.09645 for event-doubling, and 0.06562 for fixed-rate EW (η=0.5). The two online rules improve over static QAST by 0.02293 and 0.05377, corresponding to reductions of approximately 19.2% and 45.0% in mean CRSD. Their paired, pointwise normal Monte Carlo 95% intervals are [0.02181, 0.02406] and [0.05031, 0.05722], both strictly positive. This comparison uses J=2 and κ=0.5, preserving the correct occurrence probability and conditional median. It comprises 200 independent repetitions, each evaluated on 4,000 independent future observations; all raw and transferred experts have shape error. Table 15 reports all three source-information regimes and six history lengths. In both the weak and strong regimes, both online rules have positive pointwise improvement intervals at every positive history length examined, and their mean improvements increase with the number of historical events.

In the weather data, the principal gains come from protection: at p95, twCRPS falls by 27.5%/27.2% under EMOS/DRN relative to the unprotected mixture with the same weights. The three-rule replay holds the baseline, 40 sources, J=5, κ=0.75, initial weights, and feedback timing fixed, covering 42,502 observations at each threshold. Fixed-rate EW with η=0.25 gives twCRPS point reductions of 0.00092%–0.00132% relative to static QAST, so the online increments are small. Table 11 reports absolute scores and Table 12 the primary paired comparisons; sensitivity conclusions for η=0.1 and 0.5 and simultaneous inference are described in Appendix D.2.1. Appendix D.2.3 and Table 14 additionally compare feedback accumulation starting in January, April, or July on the same 21,433 evaluation observations from July–December 2018.

These experiments support two practical uses of QAST. Static QAST already provides the full protection-and-transfer construction; when sources are distinguishable and event feedback is sufficient, online learning can further exploit the adjustment space within the class. Both uses share the pathwise guarantee of Theorem 3.1.

  1. Finite learning nodes and external scores

Two forecasts can agree at every training node and still have different CDFs between nodes. The finite-node Brier regret bounds (Theorems 4.1 and B.1) therefore do not, in general, imply improvement in CRPS or twCRPS. Section 4.3 obtains a stronger result in its iid construction because the source weight scales the error of the entire CDF: Fₜᴬ − Gₚ = αₜ(F₀ − Gₚ), with αₜ = 1 − (1−ζ)w₁ₜ. Positive source weight thus reduces excess expected CRPS to αₜ² times the baseline excess risk. Theorem 4.2 uses source-weight learning in this construction to establish improvements in expected CRPS and limiting CRSD. The pathwise protection bound in Theorem 3.1 holds for any predictable convex weights.

Following your suggestion, we have added the learning-node sensitivity experiment in Appendix D.2.2 and Table 13. It compares 8, 16, and 32 midpoint nodes and 16 shifted nodes, using a separate common set of 64 nodes to evaluate conditional Brier scores over the full p90/p95 panels. Table 13 reports all direct node comparisons under the same learning rule, with CRSD intervals, allowing the effect of training-node choices to be assessed separately from the forecast class's protection property.

  1. Finite learning nodes and external scores (continued)

Appendix D.1.6 and Table 10 additionally compare 128-node and 256-node integration with an independent adaptive integrator for physical-scale twCRPS. Both selected station-day comparisons retain a positive twCRPS reduction for QAST-E relative to the unprotected mixture with the same weights. The learning-node check, independent score integration, and distribution-level theory address the learning objective, numerical evaluation, and methodological guarantee, respectively.

  1. Choosing protection when baseline probabilities are misspecified

QAST makes the scope of protection an explicit design choice: users can preserve trusted occurrence probabilities and quantile anchors while releasing constraints that require correction. We have added ten anchor-perturbation settings in Appendix D.2.5 and Table 16. Main-text Section 3.3 now summarizes the favorable results from releasing erroneous anchors and references the detailed experiment.

The main results are as follows. At conditional position 0.25 with perturbation +0.10, mean CRSD is 0.10543 for the baseline, 0.10507 when all internal anchors are retained, and 0.08802 when the erroneous anchor is released. The mean improvement over retaining all anchors is 0.01705, with a paired, pointwise normal Monte Carlo 95% interval of [0.01576, 0.01835]. The pointwise improvement intervals are positive in all eight nonzero-perturbation settings; at zero perturbation, the CRSD difference is zero to numerical precision. The perturbation is defined as the true conditional CDF minus the baseline CDF at the baseline cut. Intervals use 200 independent repetitions, each with 4,000 independent future observations.

These results support selective anchor protection: preserving trusted constraints while leaving room for source information to correct shape errors. Section 3.3 and Appendix C.2 provide the corresponding theoretical interpretation: exact probability constraints determine which features remain fixed, while transfer adapts shapes inside the cells. J=1 is available when only occurrence needs protection. If occurrence itself needs correction, the baseline can first be recalibrated and QAST then constructed around the updated baseline.

  1. Varying J and κ separately

We have added comparisons that hold the same DRN baseline and source bank fixed while varying J and κ separately in Appendix D.2.6 and Table 17. This design directly complements the original common-protection-radius comparison.

For static QAST at p95, the shared baseline CRSD is 0.09034702. With κ=0.5 fixed, the CRSD point estimates for J=1, 3, and 5 are 0.08635, 0.08822, and 0.08987. With J=5 fixed, the values for κ=0.25, 0.5, 0.75, and 0.9 are 0.08940, 0.08987, 0.09006, and 0.09045. Table 17 provides pointwise intervals for these configurations and simultaneous bands covering the full 40-contrast parameter family.

These point estimates make the protection–adjustment tradeoff concrete. Under uniform partitions, increasing J or κ tightens the perturbation radius (1−κ)/J while restricting shape adjustment. The paper therefore provides complementary comparisons: different partitions at a common protection radius, and separate changes in J and κ under a fixed baseline. Users can choose the protection level according to the probability features they wish to preserve.

  1. What QAST adds to Jeffrey's rule, simple anchoring, and fixed weights

QAST brings source conditional-shape transfer, tail-probability protection, and a pathwise diagnostic guarantee into one learnable forecast class. It jointly provides preservation of the baseline exceedance probability and selected tail-quantile structure, a CRSD perturbation bound for arbitrary predictable convex weights, and a common framework for static combinations and online learning within the class.

Positive baseline retention plays a specific role. The main-text example “Anchor values and PIT events” shows that equal CDF values at an anchor and equal quantile-crossing events are distinct requirements when a source CDF has flat regions. Positive baseline retention restores strict crossing and proportionally tightens the perturbation radius; Lemma A.3 then connects within-cell PIT displacement to CRSD stability. This explains how the components of the construction jointly deliver the guarantee.

The experiments directly support this complete protection transformation. Holding the source bank, aggregation weights, and nominal baseline coefficient fixed, p95 twCRPS falls by approximately 27% relative to matched unprotected pooling. Static QAST already satisfies the same protection theorem and achieves most of the matched-control gains; the additional controlled simulation shows the value of learning source weights within this class. The revision thus connects the protected construction, learning within the class, and their respective empirical contributions.

Thanks to the authors for their hard work. Some issues have been supplemented and clarified. However, the overall quality of the paper has not improved much. I maintain my score.

Reviewer_3ArsOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper develops methods for improving a baseline probabilistic forecast while preserving its overall shape. This is motivated by the fact that common proper scoring rules (e.g., CRPS) can be small even when tail probabilities are estimated inaccurately. The main method proceeds by using the base forecast to partition the outcome space into bins and determine the total probability mass assigned to those bins. The conditional distribution within each bin is then learned by aggregating candidate forecasts using an exponential weights algorithm. The main theoretical results show that this procedure can improve forecast quality within individual bins while guaranteeing that the aggregate CRSD score remains approximately unchanged.

Critical strengths
  • It is well known in the probabilistic forecasting literature that common proper scoring rules can over or under emphasize certain aspects of the predicted distribution. In light of this, I believe the current paper develops a reasonable method for improving tail estimates that are typically underemphasized by CRPS. This method appears to be quite practical and simple to implement for practitioners.
Critical weaknesses
  • While I believe the methods presented in this paper are reasonable, they do not have significant technical novelty. After binning the outcome space, the application of the exponential weights algorithm is quite standard. The associated theory is similarly not significantly novel.
  • I am not convinced that the result presented in Theorem 3.1 is a strength of the proposed method. In particular, ideally we would like to produce a forecast that has both low CRPS score and accurate tail estimates. This theorem shows that the the constraints imposed by the proposed approach are so restrictive that the method cannot learn meaningful information about the bulk of the distribution. This seems quite undesirable to me.
Questions for authors
  • I found the result stated in Section 4.3 for i.i.d. data to be quite difficult to follow. My interpretation is that this result shows that given a candidate forecast that exactly matches the true distribution the exponential weights algorithm will eventually put all of its mass on this value. This result seems to be somewhat unsurprising and I would appreciate any clarification from the authors as to whether this interpretation is correct and if this result has more to offer.
2 comments

Thank you for recognizing the practical relevance of the problem and the implementability of the method. The revision sharpens our central contribution: using source forecasts to improve within-cell distributional shapes while preserving specified baseline event probabilities and tail-quantile structure, with an explicit pathwise bound on the resulting diagnostic change. The theory and experiments address a practical question: how can additional source information improve forecasting while retaining baseline probability features that the user values?

  1. Technical contribution: a protected forecast class with a pathwise guarantee

QAST translates the choice of probability features to preserve into explicit distributional constraints and quantitatively connects those constraints to tail-diagnostic stability. Baseline tail quantiles define the cells and their probability masses; source models supply conditional shapes within the cells; and a positive baseline component maintains strict crossing at the protected probability levels. The resulting class is closed under convex aggregation and accommodates both static combinations and online weight learning.

Under the forecast regularity conditions in the paper, Theorem 3.1 gives, for every admissible outcome path and every finite prefix,

|CRSD_A(n) − CRSD_0(n)| ≤ (1−κ)h.

Users can therefore determine the maximum perturbation radius before observing future outcomes through the anchor spacing h and baseline-retention parameter κ. The guarantee applies to any convex weights selected before the outcome. For uniform grids with J≥2, Proposition A.2 further establishes worst-case sharpness over the admissible forecast class.

The example in Section 3.1 shows why strict crossing matters for preserving anchor counts. Section 3.3 gives the within-cell excess-PIT displacement bound. With the common exceedance-probability denominator, Lemma A.3 in Appendix A.4 turns this displacement bound into the scalar CRSD bound. Related Work and Section 3.1 explain the connection to Jeffrey's rule, while Section 4.2 identifies the established exponential-weighting framework. The methodological contribution centers on the protected class and its joint guarantees.

  1. Flexibility in the body and scope for improvement under protection

The general construction in Section 3.1 includes both body and tail cells. It preserves the total probability of the body cell while allowing its conditional shape to change with source information. Example 2.1 gives a direct illustration. As shown in Appendix C.1, transferring the source’s conditional body shape reduces excess expected CRPS from 5ε²/6 for the baseline to κ²·5ε²/6 for the transferred source, while its entire tail distribution remains correct.

The weather implementation freezes the body CDF to focus the evaluation on tail transfer; the learning objective in Section 4 likewise targets tail cells.

Protection leaves substantial room for improvement. Theorem 4.2 gives an iid construction in which every expert has shape error, yet QAST-E improves both expected CRPS and limiting CRSD while preserving the baseline exceedance probability and internal anchors. This establishes that protection and useful transfer can coexist.

The empirical comparisons also support the value of protection. At p95 with J=5 and κ=0.75, QAST-E reduces twCRPS relative to the unprotected mixture with the same weights by 27.5% under EMOS and 27.2% under DRN, with paired 95% intervals supporting both reductions (Table 1 and Appendix Table 6). Static QAST already achieves most of these gains, demonstrating the practical role of the protected construction itself.

Section 5.3 and Table 2 further illustrate the adjustment space within the protected class. For the J=3, κ=0.50 configuration, two internal tail-quantile anchors remain fixed, while the p90/p95 CRSD point reductions relative to the configuration-specific DRN baseline are 9.68%/9.47%.

The parameters J and κ give users explicit choices: adding anchors or increasing baseline retention strengthens protection, while relaxing selected constraints expands the available shape adjustment. We now summarize the favorable anchor-perturbation results in main-text Section 3.3, with a reference to Appendix D.2.5 and Table 16, illustrating how trusted probability features can be protected selectively.

  1. The experts in the construction of Section 4.3

Theorem 4.2 establishes improvement when all available experts are misspecified. Using the notation of Section 4.3, let the true distribution be Gₚ and the baseline be F₀≠Gₚ. The raw and transferred sources are

P₁ = νF₀ + (1−ν)Gₚ,
F₁* = ζF₀ + (1−ζ)Gₚ,

where ζ=κ+(1−κ)ν and 0<ν,ζ<1. Thus both the raw and transferred sources differ from Gₚ. The revised Section 4.3 states this explicitly.

The construction also explains the respective roles of transfer and learning: positive source weight improves risk, and event-doubling makes the time-averaged source weight converge to one in probability, connecting learning to improvements in the full forecast distribution. Section 4.3 now also summarizes a separate controlled simulation and refers to Appendix D.2.4 and Table 15: with two shape-misspecified source experts, source information and historical feedback jointly generate online learning gains.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.