acceptodds
Withdrawn from ICLR 2027

CADRE: Distributional Refinement for Probabilistic Time Series Forecasting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Existing probabilistic time series forecasters often map temporal observations to predictive distributions through predefined parametric heads, such as Student- and Gaussian. However, these distributional assumptions can limit their ability to capture asymmetric and irregular conditional distributions in real time series. To address this challenge, we introduce libration via istributional finement (CADRE), a framework that uses a base forecast to guide a more flexible predictive distribution. CADRE constructs input-adaptive bins from the forecaster’s predicted quantiles and uses a lightweight refinement head to learn each bin’s probability mass from shared temporal features. Experiments on six real world datasets demonstrate consistent improvements in probabilistic forecasting and gains in point forecast accuracy. Further experiments show that CADRE improves forecasting performance when transferring models across datasets. A channel-level analysis on Weather reveals an association between greater marginal skewness and larger probabilistic gains, suggesting that CADRE is particularly beneficial for forecasting asymmetric time series.

Primary area:
learning on time series and dynamical systems
Keywords:
probabilistic time series forecasting, distributional refinement, calibration, post-training adaptation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_BynFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

CADRE improves probabilistic time-series forecasting by using a Student‑t distribution’s quantiles to construct adaptive quantile bins and then learning an adaptive probability mass over bins. Training uses a dual objective: (1) a base loss to fit the base predictive distribution, and (2) a CBCE loss that supervises the refined cumulative probabilities so the resulting discretized CDF matches the ground-truth ordering across bins. Experiments on six multivariate datasets show that, compared to Gaussian and Student‑t baselines, CADRE yields consistent improvements in CRPS and mean wQL. Transfer experiments further demonstrate CADRE’s generalization across datasets.

Critical strengths
  1. CADRE is motivated by how tail probability mass and uncertainty calibration are handled. The method is formulated coherently around a base predictive distribution plus a refinement stage. The dual-object training loss function matches its motivation.
  2. CADRE integrates forecast-conditional bins with a neural refinement model. The adaptive binning is tied to the base predictive distribution, and the neural network learns the prediction head g\mathbf{g} to allocate probability mass across bins of representation z\mathbf{z}.
  3. CADRE introduces a novel refinement mechanism: it uses quantile-derived bins from three Base predictive heads (Gaussian / Student t / Quantile regression) and includes explicit tail handling, enabling the refinement to correct distributional probability allocation.
  4. Experiments across multiple multivariate datasets show consistent improvements. Additional transfer experiments further demonstrate generalization across datasets, supporting that the benefits are not dataset-specific.
  5. The paper is well organized, with clear statements and logical explanations. The motivation, methodology, and experimental results are easy to follow, making the proposed framework accessible to readers.
Critical weaknesses
  1. The comparison to other advanced methods is lacking; current results only compare to Gaussian / Student t / Quantile regression. To validate the effectiveness of CADRE, the experimental results need to be compared with advanced methods in related work.
  2. The method’s core idea is close to prior “distributional refinement” and “discretized probabilistic forecasting” lines. To investigate the mechanism of CADRE’s effectiveness, an ablation study needs to be conducted on key modules, such as removing adaptive bins, removing the tail-bin construction, and replacing CBCE with a standard likelihood.
  3. Figure 3 is intended to highlight how CADRE adapts to the ground-truth probabilistic distribution; however, the base forecaster setup is not explicitly specified. There are three Base forecasters based on Gaussian / Student t / Quantile regression; Figure 3 appears to illustrate results only for the Gaussian head. In appendix 4.1, the author provides full results but does not convince CADRE do better. Thus, a metric (e.g., Expected Calibration Error, ECE) needs to be calculated to support CARDRE’s outstanding. Further, this comparison needs to include advanced methods as in W.1.
Questions for authors

see weakness

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

The paper proposes to learn correction of probability mass of predicted quantiles and the quantiles themselves in a calibration phase using an extra refinement head. Results show improvements in both probabilistic forecasting and point forecasting and the refinement transfers to other datasets.

Critical strengths
  • The paper has good clarity, I enjoyed reading the paper.
  • The choice of CBCE makes a lot of sense to me.
  • Consistent improvements on both point and probabilistic forecasting performance.
Critical weaknesses

I might missed something, but I feel two important questions are not answered by the paper.

  • In the calibration phase, what if one just keeps training the original base forecaster without the refinement? How much improvement the base forecaster gets by the same amount of extra training (no LCBCEL_{CBCE})?
  • Following this, if LCBCEL_{CBCE} is a better loss for probabilistic forecasting, what will happen if one uses LCBCEL_{CBCE} loss with the existing head (gaussian, student-T and quantiles) already in the pre-training phase so the calibration phase is not needed anymore?
Questions for authors
  • Is the same training set used for pre-training the forecasters and the calibration refinement or some extra data is reserved for calibration?
  • The correlation between skewness and CADRE makes sense, especially for gaussian and student t heads because they are symmetric. In Figure 4, what head is used? How about the quantile based head?
  • In A.2, the lambda is 0.01, which is very low. Did the author try other values? How sensitive the results are to the choice of lambda?
Reviewer_4NxQOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This work introduces CADRE (CAlibration via Distributional REfinement). While the paper focuses on probalistic time series forecasting the method should be applicable to any univariate probabilistic regression task. The method works as follows: divide a density function into bins of equal mass and then re-order these bins to better approximate the ground truth distribution. For that the authors use the CBCE loss. In their experiments the authors show that this leads to better predictions (measured in CRPS and wQL) compared to simple Gaussian and Student-t distributions. This is expected as the re-ordered bins enable the model to model distributions more flexible compared to directly using parametrized and symetric distributions. Interestingly, the method also helps to improve on the predictions of a model trained on quantile regression; hence, a model which in principle already has a lot of flexibility.

Critical strengths
  • re-ordering equal bins is an interesting idea for distributional refinement

  • The authors show that their method improves upon on simpler distributions and interestingly also on the distributions inferred by a quantile regression model.

Critical weaknesses

My main issue with the paper is that CADRE is not compared with the state-of-the-art in in probalisitc forecasting of multivariate time series. While the idea of defining a distribution by shuffling the bins of another distribution is interesting in principle, its usefulness depends on whether the resulting model has an advantage over other approaches. The Normalizing Flow literature is not only ignored in the experiments, but also in the theoretical part of the paper. However, I think the two approaches are closely related as both methods model a density by manipulating the density function of another one.

Furthermore, the idea of re-ordering bins of a density function should not be restricted to the time series domain, and the papers evaluation section would greatly benefit from more principled experiments on syntethic and/or tabular data.

Questions for authors
  • Q1: How did you compute the emperical distribution from figure 2 ?

  • Q2: Can you give some more details about your quantile regression model (e.g. did you take care of quantile crossing ?)

  • Q3: Why did you choose time series as the main scope of this paper?

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

Related papers

Discussion (0)

Sign in to comment.