acceptodds
Withdrawn from ICLR 2027

TailCast: Release-Order Tail Extrapolation for Frozen Time-Series Foundation Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Time-series foundation models forecast zero-shot, yet they expose their predictive distribution only on a finite grid of native quantiles—Chronos-Bolt and TimesFM2.5 stop at 0.90, Chronos-2 at 0.99—while tail decisions are defined above it. Extending the profile is hard because the frozen model supplies no local shape there and validating outcomes are rare by construction, and because a label may arrive only after later forecasts have gone out. Existing post-hoc layers estimate the tail from scarce exceedances, discarding the conditional shape the backbone already encodes. We introduce TailCast, which treats that conditional quantile profile as an issue-specific anchor and learns only a small residual correction to it, continuing the top native spacing in log-survival coordinates above the grid. The residual is estimated at two adaptation scales, whose thirteen candidates are combined by a learning-rate-free rule over released pinball losses, each token scored against the state that issued it; a weighted isotonic projection, provably nonincreasing for the reported score, returns the ladder. Across four frozen backbones, two benchmark groups holding 696 cases, and three scenarios, TailCast attains the lowest tail-weighted pinball loss of the thirteen methods compared, by 6.7% and 8.9% relative to the strongest of the other twelve on each group, issuing finite, noncrossing ladders under every delay and missingness condition tested. Because it reads only native quantiles, targets, and release timestamps, the layer drops onto any frozen backbone without retraining, bringing such tail levels within reach of operational alerting.

Primary area:
learning on time series and dynamical systems
Keywords:
Time-Series Foundation Models, Tail Extrapolation, Quantile Calibration, Delayed Feedback

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_VFZPOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes TAILCAST, a post-hoc layer model that is capable of predicting tail quantiles beyond the pre-trained quantile levels. TAILCAST can actively recalibrate its predictions when the true values of previous predictions are observed. Experiments with 12 baselines and four foundation models show that TailCast achieves the lowest tail-weighted pinball loss on both benchmark groups.

Critical strengths
  • Extensive experiments are performed, comparing against 12 baseline across different scenarios
  • Ablation study is provided to justify the model design
Critical weaknesses
  • It is very difficult to follow the presentation. The descriptions are unclear. Terms appear before they are defined, and some never are, among other issues.

  • I tried to follow and understand the paper, but it is likely I misunderstood the content due to the unclear presentation

  • In table 3, it seems that the gain is mostly on using Chronos-2 as the base model. The gains for other base models are modest. Could the authors comment on the reason?

  • Results reported in table 1 averaged over different experimental scenarios: no shift, level shift, and scale shift. To understand the model's performance better, it would be better to separate the scenarios to see where the gains mostly come from.

  • The experiments used on context length 128, which seems to be on the shorter end. However, the base models support a context length beyond that. Experimenting with longer context lengths could strengthen the paper. Right now, it is unclear whether the performance gain is because the base model did not observe enough context to make an accurate prediction, leaving more errors for TAILCAST to correct.

Questions for authors

See weakensses

Reviewer_NpS1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

TAILCAST extends the quantile outputs of frozen time-series forecasters beyond their native grid. It extrapolates the two highest native quantiles in log-survival coordinates to form a forecast-conditioned anchor, adds residual corrections from recent windows and online quantile trackers, combines these candidates using AdaHedge as delayed outcomes arrive, and projects the resulting quantiles to prevent crossing. The paper reports theoretical properties of the projection and delayed aggregation and evaluates the method on 58 series with four backbones and three scenarios. On its primary tail-weighted pinball-loss metric, TAILCAST has the lowest reported mean among thirteen methods in both benchmark groups.

Critical strengths
  1. The problem is interesting and practical: a frozen forecaster may not output the very high quantiles needed for an alert.
  2. The release-order treatment of delayed feedback is a useful design choice, and the paper describes what is stored when each prediction is issued (Sections 3–4).
  3. The empirical study includes several backbones, scenarios, and baselines. The reported tail-weighted pinball loss is better than the strongest comparator on both benchmark groups (Table 1).
Critical weaknesses
  1. The paper motivates TAILCAST partly through operational alerts at very high quantile levels, yet the reported coverage at the most extreme level remains substantially below the nominal target. At p = 0.999, Tables 4 and 5 (p. 18) show mean coverage gaps of −0.52 and −0.69 percentage points on Benchmarks A and B, corresponding to mean exceedance frequencies of approximately 0.62% and 0.79% rather than the nominal 0.10%. Figure 3 (p. 22) also reports a −0.20-point gap in the natural scenario. Although TAILCAST achieves the lowest tail-weighted pinball loss among the compared methods, that score alone does not establish reliable 99.9% thresholds for alerting. The paper should report forecast and exceedance counts, coverage uncertainty by source and scenario, and the practical consequences of these deviations before making a strong operational claim.

  2. The evidence for generalization across independent data sources is weaker than the number of evaluation cases may suggest. Section 5.1 (p. 7) reports 276 cases in Benchmark A and 420 in Benchmark B, formed by evaluating 23 and 35 series across four backbones and three scenarios, while Figure 4 (p. 23) states that the groups contain only six and three data sources, respectively. The paper uses paired source-cluster bootstrap intervals, and the Figure 4 caption acknowledges the limitation for Benchmark B. Resampling only three sources cannot create additional independent evidence. Source-level paired TP differences and leave-one-source-out results are needed to establish whether the overall Benchmark B advantage is consistent across sources rather than dominated by one of them.

Questions for authors
  1. At p = 0.999, how many evaluated forecasts and observed exceedances are there in each data source and scenario? Please report source-level coverage estimates and uncertainty, and clarify what level of miscoverage would still support the proposed operational-alert use.

  2. For Benchmark B, do the paired TP improvements over Rolling hold within each of its three data sources? Please provide the source-level differences and leave-one-source-out results, as these would show whether the aggregate advantage depends on a single source.

Reviewer_u2hSOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes TailCast a method to extend the forecast quantiles given by a time series foundation model (TSFM) further into the tail of the distribution. For example, Chronos-Bolt gives quantiles up to 0.9, TailCast aims to post-hoc add further quantiles above 0.9. TailCast does this by a) in log-survival coordinates, fitting an affine function to the two largest quantiles given by the TSFM and extending it into the tail, to give a base quantile prediction and b) adjusting the prediction using online feedback, TailCast does this in two ways 1) using the empirical residual distribution and 2) using a score tracking update rule. Both these methods are then combined to provide the overall adjustment by using the adaHedge method. The paper tests the capabilities of TailCast using two curated benchmarks. The results find that TailCast performs well in terms of a weighted pinball loss, the loss TailCast is fitted with. The paper also reports many ablations, two of which perform better than TailCast (Extended ladder and Unweighted projection).

Critical strengths
  • I think modeling tail risk is an important and underexplored topic in TSFMs.

  • The paper performs a wide range of ablations to test the importance of components of TailCast.

Critical weaknesses
  • The paper is very hard to read. Many terms are used without definition. Additionally, non-standard terms are often used. The formulas are also hard to read, where again many of the symbols and variables used are not defined before use. TailCast is quite simple, therefore it should be possible to rewrite the paper to make it understandable. I encourage the authors to do so, and without substation revision to improve clarity, I do not see myself improving my score.

  • The paper describes the method but there seems to be a lack of discussion of why design decision were taken. This makes the paper read like the authors put lots of separate ideas together without a cohesive story. This makes it hard to learn generalizable ideas from the paper.

  • The paper uses its own datasets not standard benchmarks. No discussion of why this is done is given and makes it hard to evaluate TailCast. Additionally, details of the datasets used are not given in enough detail for it to be reproducible nor is the dataset given as supplementary material.

  • No description of how the many hyperparameters of TailCast is fitted given, nor their sensitivity.

  • The first method to estimate residuals is very similar to conformal prediction, yet not comparison is given in the paper. This comment can be made more generally in the sense that no comparison to the many methods created to model residuals online is given in the paper.

  • Proofs are lacking many details, many things are stated as facts instead of being derived.

Questions for authors
  • Can you explain how you performed your cluster bootstrap and at what significance level.

  • What are the two controls Extended ladder and Unweighted projection? These outperform TailCast, which should be made clear in the main text, yet they are never described.

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

Related papers

Discussion (0)

Sign in to comment.