acceptodds
Under review as a conference paper at ICLR 2027

SHIFT: Progressive Feedback for Time Series Forecasting

Abstract

A point forecast is a compromise: when several futures are consistent with the observed history, the conditional mean matches none of them. This compromise is unnecessary in staged deployment, where a long-horizon forecast is released block by block and each released block is measured before the next one is due— early observations discriminate between futures, and the forecaster should change its mind rather than merely its output. We present SHIFT, which maintains a bank of complete future hypotheses together with a belief over them. A historyconditioned generator writes hypotheses as low-rank residual fields around a base forecast; as blocks are revealed, a causal filter updates the belief through a discounted log-linear rule that combines a likelihood energy with a learned, calibrated evidence term, while a recurrent assimilator rewrites the not-yet-released portion of every hypothesis in response to sample-specific innovations. Supervising such a filter is the central difficulty: the correct belief is only unambiguous in hindsight. SHIFT therefore trains the causal student against a hindsight teacher— a posterior computed with the full target—distilling privileged information into every intermediate belief, and augments this with coverage, diversity, balance, and entropy-contraction objectives that shape the bank and the filter.

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

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

This paper introduces Progressive Feedback Forecasting, a new forecasting paradigm that moves beyond conventional one-shot prediction by progressively incorporating newly revealed observations to revise the still-unseen future. Under this paradigm, the authors propose SHIFT, a framework that maintains multiple future hypotheses and dynamically updates them through causal belief filtering and recurrent trajectory assimilation. Extensive experiments on multiple real-world datasets demonstrate substantial performance improvements over conventional forecasting and online adaptation baselines.

Critical strengths
  1. The paper introduces Progressive Feedback Forecasting, a novel forecasting paradigm that enables progressive forecast revision using newly revealed observations, offering a meaningful alternative to conventional one-shot forecasting.

  2. SHIFT presents a well-designed framework that integrates multiple future hypotheses, causal belief filtering, and trajectory assimilation. Hindsight-teacher distillation further enables effective supervision while preserving causal inference.

  3. The framework provides an interpretable feedback-driven forecasting process and supports efficient inference without online parameter updates.

Critical weaknesses
  1. The paper does not explore how SHIFT handles noisy or missing feedback, which is quite common in real-world applications.

  2. It is not very clear how SHIFT generates different future hypotheses from the same initial forecast and keeps them sufficiently diverse.

Questions for authors
  1. Feedback may not always arrive on time. How would SHIFT perform when observations are revealed with some delay?

  2. Could the authors explain how the feedback block size is chosen in practice? Would different applications need different settings?

  3. Online forecasting and continual learning methods also use newly observed data to improve predictions. Could the authors explain more clearly how SHIFT differs from these approaches, especially in how feedback is used and what gets updated?

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

In this paper, the authors address the problem of long-horizon forecasting under a setting they term Progressive Feedback Forecasting (PFF). In PFF, the horizon is divided into blocks, and once a block has been committed, its ground truth becomes available and can be used to revise the forecasts for the blocks that follow. The authors present PFF as a new forecasting setting and propose SHIFT as the new solution. Starting from a base forecast, SHIFT maintains a set of future hypotheses that are reweighted by a causal belief filter as the blocks are revealed, and a GRU-based assimilator then revises the portions that have not yet been revealed. Hindsight-teacher distillation is used to supervise the intermediate beliefs, drawing on the full future only during training. On eleven benchmark datasets, SHIFT lowers MSE and MAE by 25.2% and 15.1% on average relative to the best baseline for each dataset. These results are supported by component ablations, sensitivity analyses, a direct comparison with rolling-window forecasters, and additional experiments on different backbones.

Critical strengths
  1. The Progressive Feedback Forecasting setting is practically relevant, and the SHIFT design respects causality. Systems in operation tend to use predictions progressively as new observations become available, and this paper formalizes that practice through a clear commit-then-reveal protocol and an objective defined over the committed blocks. Causality is built into the model's construction. Teachers are used only as stop-gradient training targets, assimilation corrections apply only to steps after kS, and the decoder is zero-initialized so that the revision starts at the identity. The history is encoded once, and as stated in Appendix B, the per-stage cost does not depend on L. The method also performs no gradient computation or parameter update at inference, which makes each stage cheaper than rolling re-forecasting or online adaptation. I consider this an important point, since it means the framework can serve as a credible component regardless of the size of the claimed gains.

  2. The paper's own diagnostic analyses show clearly which parts of SHIFT produce the improvement. The update breakdown in Table 3, the sample-level improvement rates in Table 4, and the qualitative analysis in Appendix C.7 go into greater depth than most forecasting papers. Assimilation improves about 87% to 88% of Weather test windows on every target block and about 86% of CzeLan windows on the later blocks (Table 4), which shows that the gain is not due to just a few favorable cases. Together with Table 6, which shows that SHIFT improves two different backbones, DLinear and iTransformer, in 15 of 16 configurations, these analyses give a clear picture of which parts of the design transfer. This kind of transparency matters a great deal to my evaluation because it shows which factors drive performance and makes it possible to verify the concerns raised below.

Critical weaknesses
  1. The main comparison does not separate the method from the extra information it receives. At S = 96 and H = 720, SHIFT reaches an MSE of 0.393 on ETTh1 and 0.183 on ETTm2, close to what iTransformer already achieves at H = 96 in Table 1 (0.386 and 0.180). The 25.2% average gain and the 75.6% gain on Exchange therefore reflect the evaluation setting as much as the model. A fair baseline is also missing. The only matched-feedback comparisons are Table 5, which covers three datasets at one horizon with external forecasters trained for H = 720 and then cut to each block, and Figure 8, which covers two datasets with a rolling version of the same backbone. A more natural baseline would train the same backbone for horizon S, re-run it block by block with the same input window, and evaluate it on all datasets and horizons.

  2. The multi-hypothesis design is the main conceptual contribution, yet it explains little of the result. In Table 3, the MSE reduction from filtering on ETTh1 and CzeLan is three to five orders of magnitude smaller than that from assimilation, which accounts for virtually the entire gain. Belief filtering over several hypotheses is the central idea of SHIFT, but it barely changes the final forecasts. The hypotheses also collapse. In the four-hypothesis examples of Appendix C.7, the belief weights stay nearly uniform and the trajectories largely overlap, which is expected because the base loss supervises only the mixture forecast and no loss term explicitly encourages diversity. The motivating case of clearly distinct futures, such as a sharp spike versus a flat continuation, appears only as a schematic in Figure 1 and is never demonstrated on real data.

Questions for authors

The following are clarification requests, and the answers would affect my assessment.

  1. Why is a GRU used for assimilation rather than attention? A recurrent state is a reasonable way of combining past and newly revealed information, yet the GRU is given only four summary statistics for each revealed block (Eq. 7). An attention module over the revealed blocks, or over the history together with the revealed blocks, could make direct use of the raw observations and long-range dependencies. Was such a variant tried, and how does it compare in accuracy and cost?

  2. What does it cost to train SHIFT? The method introduces several loss terms, and Appendix B states that gradients flow through every feedback stage and that each stage oracle adds a pass over the hypothesis bank. Figures 5 and 6 report only inference-time latency and memory. Please report the training time per epoch and the peak training memory, with the cost of the loss computation and backpropagation included, compared against the baselines.

  3. Do the hypotheses actually differ from one another, and does greater diversity help? Appendix C.7 notes that the four candidate trajectories largely overlap in its examples, and Figure 6 shows only limited and inconsistent gains as the bank grows beyond a few hypotheses. Please report a diversity measure across the bank, such as the average pairwise distance between hypothesis trajectories at each stage, and state whether a diversity-encouraging loss term was tried and how it affected accuracy.

  4. What sets SHIFT apart from prior methods of a similar structure, and what is its concrete advantage over them? The belief update looks much like particle filtering, the assimilator looks much like an ensemble Kalman style correction, and the hypothesis bank is close to multiple choice learning and mixture density forecasting.

  5. The component ablation in Table 2 shows only small differences between Assimilation only, No teacher, Single corrector and Full SHIFT. Could you report the mean and standard deviation over several seeds to show whether these differences are meaningful?

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

The paper studies Progressive Feedback Forecasting (PFF), where predictions are committed block by block, and observations from previously committed blocks become available for revising subsequent predictions. SHIFT implements this protocol using a bank of candidate trajectories, a causal belief filter, and a recurrent assimilator that revises the unobserved portions of those trajectories. Training additionally uses global and stage-specific hindsight teachers to supervise the belief updates. Experiments cover eleven datasets, with component ablations and supplementary comparisons against rolling forecasting and feedback-based correction methods.

Critical strengths
  1. The operational problem is meaningful. Incorporating observations that arrive while a forecasting episode unfolds is practically relevant. The explicit commit-then-reveal protocol makes the intended causal information flow clear, and the method supports forecast revision without test-time parameter optimization.
  2. The paper goes beyond comparisons with one-shot forecasting. Appendix C.4 compares against rolling and residual-correction variants of the same backbone, and Appendix C.5 evaluates several methods under matched feedback availability. Taken at face value, Table 5 reports MSE improvements over the strongest evaluated rolling baseline of approximately 14.9%, 9.9%, and 8.1% on Weather, ETTm2, and Exchange, respectively. Together with the ablations, these results suggest that learned trajectory correction can add value over naive re-forecasting and additive residual correction.
Critical weaknesses
  1. The main evaluation does not adequately separate the benefit of receiving new observations from the contribution of SHIFT. Appendix C.5 explicitly states that the standard baselines in Table 1 use one-shot forecasts without within-horizon observations. Under the objective in Section 3, SHIFT is evaluated on successive committed blocks, each with a prediction lead of at most (S), whereas a one-shot baseline must predict up to (H) steps from the original forecast origin. Consequently, the headline improvements combine a change in available information and effective prediction lead with an algorithmic improvement.

I appreciate the matched-feedback comparisons in Appendices C.4–C.5; the concern is not that these comparisons are absent. However, Table 5 covers only three datasets at (H=720), and its rolling iTransformer and KUP-BI models are trained to forecast 720 steps while only their first at most 96 predictions are committed at each stage. This leaves an important training-objective mismatch unresolved. A forecaster trained for the actual committed block length is a natural control for the stated evaluation objective. The current results therefore do not yet establish how much of the claimed advantage survives a comparison aligned in both feedback availability and prediction target.

  1. The incremental contribution of the multi-hypothesis filtering and distillation machinery is insufficiently established. Table 2 already contains the particularly relevant Single corrector baseline. Full SHIFT improves its MSE by only approximately 1.31% on ETTh1, 0.61% on ETTm2, 1.44% on Weather, and 1.57% on CzeLan; on ETTm2, the single corrector has slightly better MAE. These comparisons are restricted to (H=96), and variability across random seeds is not reported.

Table 3 further shows that assimilation accounts for nearly all of the measured update benefit, with very small direct contributions from filtering on ETTh1 and CzeLan. Appendix C.7 also acknowledges overlapping trajectories and nearly uniform beliefs in its small-bank examples. These observations do not prove that filtering or distillation is useless, but they leave the central methodological claim under-supported: the evidence is considerably stronger for learned trajectory correction than for the necessity of maintaining and distilling beliefs over multiple futures. Establishing a reproducible improvement over the single corrector, together with its computational cost, is important to justify the proposed complexity.

Questions for authors
  1. Protocol clarification and a targeted control: What observations and update opportunities are available to each online baseline in Table 1? For the matched-feedback experiment in Appendix C.5, does SHIFT retain its advantage over a rolling forecaster trained directly for the committed block length (S=96), rather than (H=720)? A representative comparison using the same backbone and validation procedure would help isolate the algorithmic contribution.

  2. Evidence for the additional components: Does the improvement over Single corrector in Table 2 persist across random seeds and at a longer episode horizon? Please report the variability of this difference and the corresponding inference cost. This would materially affect whether I view the hypothesis bank, filtering, and teacher supervision as a justified contribution beyond recurrent correction.

  3. Numerical clarification: In Table 5, KUP-BI + Rolling on Exchange has MSE (0.141) and MAE (0.401). With the same errors, normalization, and averaging weights, these should satisfy (\mathrm{MAE}^2 \leq \mathrm{MSE}), but (0.401^2 \approx 0.161 > 0.141). Also, the four SEMixer results on Exchange in Table 1 average to approximately (0.4165/0.4363) for MSE/MAE, rather than the reported (0.200/0.308). Could the authors reconcile these entries and clarify whether any aggregate improvement claims change?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.