acceptodds
Under review as a conference paper at ICLR 2027

Beyond One-Shot Forecasting: Predict-then-Reveal Time Series Prediction

Abstract

Long-term time-series forecasting typically predicts the entire future horizon in one shot from a fixed historical context. However, as the forecast unfolds, earlier observations within the prediction horizon become available and can provide valuable evidence for refining the still-unseen future. We introduce Predict-then-Reveal Forecasting, a forecasting paradigm in which predictions are progressively committed, revealed, and subsequently used to improve later forecasts. Under this paradigm, we propose PRISM, a progressive forecasting framework that maintains multiple candidate future trajectories and continuously updates them using newly revealed observations. At each stage, PRISM jointly adjusts the relative beliefs over candidate futures according to their consistency with observed feedback and revises their unrevealed trajectories based on the resulting residual patterns. To learn effective intermediate revisions, PRISM further exploits complete future outcomes as privileged supervision during training, while inference relies exclusively on observations that have already been revealed. This design enables forecasts to evolve without test-time parameter optimization or access to future targets. Extensive experiments across diverse real-world time-series benchmarks and prediction horizons demonstrate that PRISM consistently outperforms strong standard forecasting and online adaptation methods.

Primary area:
learning on time series and dynamical systems
Keywords:
Time series forecasting

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_zbcqOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper starts from the observation that most time series models forecast the entire horizon in one shot, even though there are clear benefits to exploiting revealed outcomes. The authors propose PRISM, a new paradigm where the model predicts one block, observes the true outcomes, and corrects the remaining forecast based on past errors. The method consists of three main components: a shared backbone that enlarges and updates the context window as new predictions are revealed to produce a "fresh forecast," a GRU that tracks errors accumulated across the forecasting horizon, and a fusion gate that weights the fresh and previous forecasts. Experiments across several datasets demonstrate the method's improvements, and thorough ablations validate the contribution of each component.

Critical strengths
  1. The idea of exploiting revealed predictions is sound. 2) The method presents several intuitive steps for leveraging revealed information. 3) The experiments are convincing and well-conducted, and the paper is clearly written and easy to follow. Great work by the authors in including comparisons with other online approaches.
Critical weaknesses
  1. Inference-time compute is not discussed, even though the computational cost of this approach could be a bottleneck for practical deployment. 2) Multivariate time-series dynamics are not fully or properly tackled.
Questions for authors

1- What is the computation cost of the method (inference time compute) compared to competitors ?

2- How to handle efficiently multivariate and covariate ?

Reviewer_9qbcOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies blockwise forecasting with sequential feedback: predictions are committed before the corresponding observations become available, and revealed outcomes inform subsequent forecasts. PRISM combines rolling-context re-encoding, recurrent memory of forecast residuals, and memory-conditioned fusion and affine correction of fresh and previous predictions. Training supervises both initial and intermediate forecasts, without test-time parameter updates. Evaluation covers 12 datasets and four horizons, supplemented by component ablations, feedback-frequency analysis, and matched-feedback comparisons on Weather.

Critical strengths

The revision mechanism is well motivated. Forecast residuals provide information about recent model errors, and recurrent memory allows this information to accumulate across updates. Conditioning forecast fusion and correction on this memory provides a practical way to revise predictions without test-time optimization.

The controlled experiments support some benefit beyond refreshing the input. PRISM reduces MSE over naive re-encoding by 2.6%, 3.8%, and 6.4% on the three ablation datasets. Under matched feedback on Weather, it also improves over the strongest competing MSE by approximately 11.5% and 10.6% at horizons 96 and 720, respectively.

Critical weaknesses

The headline comparisons do not sufficiently isolate the methodological contribution. Table 1 does not establish comparable feedback access across methods, while explicitly matched comparisons cover only Weather at two horizons. With a fixed block size (B), the scored predictions have lead times of at most (B), even as the overall horizon (H) increases. Stable error across longer horizons therefore does not, by itself, establish better long-range forecasting. This limitation is particularly consequential for the large improvement reported on Exchange, where a matched-feedback comparison is missing.

The novelty claims exceed what the formulation establishes. The protocol closely resembles blockwise rolling-origin forecasting, and TAFAS already uses partial observations to revise the unobserved portion of a forecast. Learned residual memory without test-time optimization is a useful distinction, but the argument for a new forecasting paradigm remains weak. Similarly, Eq. (8) describes supervision of intermediate predictions using their targets; a distinct privileged-information mechanism is not apparent. The contribution would be more convincing if framed around the revision architecture and supported by stronger evidence of its incremental value.

Questions for authors

What block size (B) is used for each horizon in Table 1? What setting changes explain PRISM’s Weather MSE of 0.162 at (H=720) in Table 1 versus 0.2078 in Table 3?

Do the gains persist beyond Weather, particularly on Exchange, when methods receive identical feedback? Does naive re-encoding use the same progressive training objective in the ablation? Please report variability across runs to help assess the incremental gains.

What distinguishes the proposed privileged supervision from ordinary supervision of intermediate forecasts? How do the abstract’s candidate beliefs correspond to the fresh/previous forecast gate in Eq. (6)?

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

The paper studies forecasting with progressively revealed observations: a model commits predictions for the next block, observes the corresponding outcomes, and uses this feedback to revise the remaining target horizon. The proposed method repeatedly re-encodes an updated rolling context and maintains a per-variable GRU state containing information about observations and prediction residuals. This state controls the fusion of fresh forecasts with the unissued portion of previous forecasts, followed by horizon-dependent affine modulation. Training supervises the initial forecast and intermediate revisions, while inference requires no test-time parameter optimization. The evaluation covers 12 multivariate time-series benchmarks, 11 competing methods, and multiple prediction horizons. Additional experiments include component ablations on three datasets, analyses of feedback frequency and revision behavior, and a comparison on Weather in which all methods receive the same progressively revealed observations.

Critical strengths

The controlled experiments provide evidence of additional predictive value beyond updating the input observations. Appendix A.4 is the strongest evidence supporting the submission. All methods use the same input length, block size, reveal schedule, forecasting origins, and evaluation interval. Under these conditions, the proposed method achieves Weather MSE of 0.1027 versus 0.1161 for the strongest competing baseline at H=96H=96, and 0.2078 versus 0.2325 at H=720H=720, corresponding to approximately 11.5% and 10.6% reductions. Figure 2 also consistently favors the complete model over naive re-encoding on three datasets. These results support the possibility that the learned revision mechanism is useful. Their scope is limited, and the comparison against other architectures does not independently isolate the contribution of the revision modules, but they provide a substantive positive result that should be credited in assessing the paper.

Critical weaknesses
  • The conceptual novelty is overstated, and the significance of the remaining architectural contribution is insufficiently established. Advancing the forecasting origin and generating updated predictions when observations arrive is established rolling-origin forecasting. This can be done by updating the input context without refitting model parameters, as discussed explicitly in Section 4.1.2 of [1]. Furthermore, TAFAS [2], which the submission itself cites, already uses partially observed targets to revise the still-unobserved portion of a forecast while retaining predictions for elapsed timestamps. Thus, revising predictions within the current target horizon is also insufficient to establish a new forecasting paradigm. The defensible methodological contribution therefore lies primarily in the particular combination of recurrent residual memory, forecast fusion, and affine modulation. This is a coherent design, but the paper does not sufficiently demonstrate that this combination constitutes a substantial advance beyond a straightforward rolling forecaster. Because introducing a new paradigm is presented as a central contribution, the gap between that framing and the narrower architectural contribution materially affects my assessment of the paper’s significance

  • The headline evaluation does not adequately establish the contribution, reliability, or practical value of the additional revision machinery. Section 3 evaluates the concatenation of successively committed blocks. Each scored prediction is therefore issued at most BB steps before its target, whereas fixed-origin forecasting evaluates predictions extending up to HH steps without intervening observations. Substantial improvements are unsurprising when forecasts receive newer information and are evaluated at shorter lead times. Hence, the reported 87.9% MSE reduction on Exchange does not establish an improvement of that magnitude attributable to PRISM’s architecture under matched forecasting conditions. Figure 2 illustrates the distinction. On ETTh1, ETTm1, and Weather, naive re-encoding achieves approximately 89–92% of the absolute MSE reduction from the fixed-origin backbone to complete PRISM. The additional relative reductions from the complete model are 2.6%, 3.8%, and 6.4%, respectively, as Section 5.3 acknowledges. These figures apply to the three ablation datasets and do not directly decompose the Exchange result. Nevertheless, they show why the incremental architectural contribution requires more careful validation than the headline results provide. Appendix A.4 partially addresses this concern, but its matched-protocol evaluation covers only one dataset and two horizons. It is also unclear whether naive re-encoding receives the same progressive training objective and tuning budget. No direct comparison quantifies the additional inference cost of the memory, fusion, and modulation components relative to naive re-encoding.

  • The presentation and statistical reporting of the main results further impede assessment. Table 1 is extremely difficult to parse because of its very small text and dense arrangement of datasets, horizons, models, and metrics. Although the results are averaged over five independent runs, the table provides no standard deviations or confidence intervals. The component ablations likewise lack uncertainty estimates, making it difficult to assess the reliability of the relatively modest incremental improvements. The unspecified main-experiment block size and the unexplained difference between the main and appendix Weather results introduce additional uncertainty about the experimental conditions.

Taken together, the evidence does not yet establish that the added architectural complexity delivers a sufficiently reliable and general benefit to support the paper’s claims.

References

[1] Hansika Hewamalage, Klaus Ackermann, and Christoph Bergmeir. “Forecast evaluation for data scientists: common pitfalls and best practices.” Data Mining and Knowledge Discovery, 37, 788–832, 2023. DOI: 10.1007/s10618-022-00894-5.

[2] HyunGi Kim, Siwon Kim, Jisoo Mok, and Sungroh Yoon. “Battling the Non-stationarity in Time Series Forecasting via Test-time Adaptation.” Proceedings of the AAAI Conference on Artificial Intelligence, 39(17), 17868–17876, 2025. DOI: 10.1609/aaai.v39i17.33965.

Questions for authors
  • What block length (B) and reveal schedule are used for each setting in Table 1 and Figure 2? Appendix A.4 specifies (B=H/4), but the corresponding main-experiment setting is unclear. Why does PRISM’s Weather result at (H=720) change from 0.162 MSE / 0.211 MAE in Table 1 to 0.2078 / 0.2510 in Table 3? Please identify which experimental conditions differ.

  • Is naive re-encoding trained with the same backbone, progressive supervision, tuning budget, and reveal schedule as full PRISM? Under these controls, how consistently does the complete method improve over naive re-encoding and strong rolling baselines beyond the reported settings, particularly on Exchange at (H=720), which supplies the headline improvement? Please report standard deviations or suitable confidence intervals for the main comparisons and ablations.

  • What additional parameter count, inference latency, and memory usage do the revision components introduce relative to naive re-encoding at matched block size and forecasting origins? These measurements would materially affect whether the reported incremental accuracy gains justify the additional machinery.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.