acceptodds
Under review as a conference paper at ICLR 2027

ClockUp: Clock-Structured Additive Uncertainty Modeling for Non-stationary Probabilistic Time Series Forecasting

Abstract

Reliable probabilistic time series forecasting requires adapting to local distribution shifts while capturing recurring temporal structure. Although instance normalization mitigates variation in local level and scale, similar normalized histories can occur at different cycle positions even when their future distributions differ. Consequently, latent uncertainty models conditioned solely on normalized history may struggle to distinguish phase-dependent predictive distributions, a limitation we term . To address this limitation, we propose , a -structured framework that explicitly incorporates temporal phase into latent ncertainty modeling and robabilistic forecasting. ClockUp combines a history-conditioned Gaussian state component with a Fourier-phase-conditioned Gaussian clock component to form an additive latent prior, enabling direct reparameterized sampling with closed-form mean and variance. A push-forward decoder further combines lead-time-specific and phase-conditioned mappings to generate forecasts non-iteratively. We train ClockUp using empirical CRPS and additionally investigate P-CRPS, a phase-balanced training variant designed to reduce scale-driven loss imbalance across phases. Experiments on seven real-world benchmarks demonstrate improved probabilistic and point forecasting accuracy. Together with phase-wise analyses on selected datasets, these results support explicit phase conditioning as a useful inductive bias for probabilistic forecasting and highlight the value of phase-wise calibration diagnostics.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Clock-Structured uncertainty modeling, Non-stationary time series, Probabilistic forecasting

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

This paper proposes ClockUp, which combines history-conditioned and phase-conditioned Gaussian priors with a phase-aware decoder for non-stationary probabilistic time series forecasting. Results on seven datasets demonstrate its effectiveness, although the architectural novelty, latent-variable utilization, and training costs require further justification.

Critical strengths

Clear motivation. The paper addresses phase ambiguity in normalized histories and shows how aggregate calibration metrics can conceal phase-specific errors.

Consistent results. ClockUp improves CRPS across all seven datasets while retaining the inference efficiency of non-iterative generation.

Useful supporting analyses. The paper includes component ablations, comparisons of phase representations, and sensitivity analyses, while clearly acknowledging its theoretical limitations.

Critical weaknesses

Limited methodological novelty. The approach primarily combines conditional Gaussian priors, Fourier features, and additive decoding. The theoretical analysis supports the value of phase information but does not establish the distinctive benefit of the proposed additive state–clock structure.

Insufficient analysis of latent degeneration. The t-SNE visualizations do not establish that both prior branches are effectively utilized. The authors should examine whether branch variances approach zero or the decoder ignores stochastic latent inputs, using quantitative variance statistics and latent perturbation experiments.

Unclear training complexity. Inference efficiency is evaluated, but training time and memory costs are insufficiently characterized. Direct computation of the pairwise CRPS term requires O(K2)O(K^2) operations. The authors should clarify whether sorting-based acceleration or approximate estimators are used and report how training costs scale with the sample count.

Limited evaluation scope. The main experiments use a single lookback length and forecast horizon. Further evaluation across forecasting horizons and under drifting or unstable periodic patterns would better establish the method’s applicability.

Questions for authors

See weaknesses

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

The paper introduces ClockUP a probalistic forecasting model that makes use of the date time. The authors argue that a forecast should only depend on the observed history but should utilize phase (e.g. hour of day) as a future covariate. A latent normal distribution is computed by combining 2 seperatly computed gaussians. One is computed only based on the phase and one based on the history. The samples from the "combined" latent normal distribution are first concetanted with an embedding of the oberservations and then further processed by an MLP. ClockUP introduces an output layer which is adding the result of a vanilla layer with a linear layer which is paramterzed based on the phase. In their ablation studies the authors justify all their architectural specifications. However, many aspects have a really small effect and it is doubtful that it would survive a significance test.

ClockUp is trained on the empercial CRPS and the authors introduce a special CRPS version which is normalized by the phase. This is motivated by the fact that certain phases have a higher impact on the CRPS (due to higher variance) and the phase-normalized CRPS is supposed to prevent the model from over-focusing on such phases.

In their main experiments ClockUP consistently outperforms its baselines on CRPS and QICE.

Critical strengths

S1

The idea of utlizing the datetime / phase is well-motivated and Figure 1 provides a nice Visualization of this motivation

S2

ClockUP establishes a new state-of-the-art on the evaluation protocol from the PPM paper.

S3

The paper is clearly written and well structured.

Critical weaknesses

W1

The authors claim the articulation of "phase blindness" as a contribution. However, the fact that the date-time or phase is usful future covariate is well-known and already used previous work.

W2

The ClockUP architecture contains many details which seem a little bit odd at first glance. While the author partly justify these decisions by their ablation experiments, the lack of e.g. signifance tests, leave some doubts on wether all these design choices are actually useful. Interestingly, ClockUP would already be a very good model without using the phase at all (see the w/o τ\tau ablation; it appears to beat the best baseline PPM consistently). Hence, a significant part of the lift might stem from using a CPRS-trained latent varialbe model.

W3

The authors did not include K2VAE, a very recent and well-published model, in the list of baselines.

W4

It is unclear whether the baseline models are also getting the information about the phase. Any model which can use future covariates should in principle be able to make use of information about the date-time/phase.

W5

The structured prior is based on the assumption that the latent states of phase and state should be indipenend (eq. 12). While the ablation studies shows that the structured prior performs slightly better (again, we cannot make any statements about significance) is not clear why it is better than encoding history and phase together in order to parameterize the latent prior.

Questions for authors
  • Q1: Could you add K2VAE to the baselines ?

  • Q2: Could you provide the standard deviations or sigificance tests especially for the structured prior ?

  • Q3: What do the colors in the tables mean ?

  • Q4: Does any of the baselines use the phase too ?

  • Q5: You mention training your model with NLL as an ablation. How does that work ?

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

ClockUp combines history-conditioned and phase-conditioned Gaussian components with a phase-conditioned decoder. It tests whether this structure improves uncertainty estimates beyond simply supplying phase features. The evaluation includes objective and architecture controls, phase misalignment and deseasonalization experiments, and phase-wise calibration diagnostics that can reveal errors hidden by aggregate metrics.

Critical strengths
  • The architecture expresses a clear, testable hypothesis about phase conditioning. Appendix A is appropriately careful about identifiability, architectural superiority, and the limits of marginal prediction, keeping the theoretical claims within their scope.
  • The controls directly address the proposed mechanism through matched objectives, phase features, misalignment, and deseasonalization. The phase-wise diagnostics add useful information beyond pooled calibration, making this more informative than a performance comparison alone (Table 3; Figures 3, 7, 11).
Critical weaknesses
  • The architectural contribution is less clear than the overall improvement. Changing the objective explains some of the gain, and the remaining advantage over CRPS-trained PPM and phase concatenation needs paired seed results and clearer capacity matching. Removing clock variance also improves Weather CRPS while worsening QICE, qualifying the claim that it is the dominant driver (Section 5.3; Table 3).
  • Calibration values in the tables and figures differ beyond rounding. Different runs or subsets could explain this, but the aggregation details need clarification because phase-wise calibration is a central part of the evidence (Table 1; Figure 11).
Questions for authors
  1. Clarification and evidence request: In the CRPS-trained PPM* and concatenation comparisons, which model sizes and optimization settings are matched? How consistent are the architectural gains across paired runs?
  2. Clarification: In the existing variance-removal experiments, when does clock variance improve both proper scoring and phase-wise calibration, and when does it trade one against the other?
  3. Clarification: Do the main tables and phase-wise figures use the same runs and forecast samples? What aggregation and QICE conventions explain the differences?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.