acceptodds
Under review as a conference paper at ICLR 2027

Same Forecast, Different Capacity: Acquisition Paths in Time-Series Forecasting

Abstract

How much structure does a forecasting model need to retain its prediction accuracy? The answer depends on how the smaller model is obtained. We study capacity as the first structural budget that reaches a fixed accuracy target, comparing models learned at that budget with restrictions of a learned full model. Across six public forecasting model families, the two routes can require substantially different capacities. More surprisingly, the complete forecasting function need not change for its restricted capacity to change: transferring functionally invisible components between trained linear representations changes the required budget in both directions, consistently across three seeds on four external model–task pairs. We identify two distinct sources of path dependence. A restriction can expose components that cancel in the full function; even when this source is eliminated, retained and discarded features remain coupled through the training covariance, producing different restricted and refitted solutions. Exact linear controls separate these sources from incomplete optimization. An exact risk response connects function-preserving interventions to capacity thresholds. Residual alignment explains why closer approximation of the full model need not improve task accuracy. Capacity is therefore an operational property of a forecasting task and its acquisition path, not a consequence of the full predictor alone.

Primary area:
learning on time series and dynamical systems
Keywords:
time series forecasting, model capacity, model compression, function equivalence, representation analysis

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

The paper studies how the structural budget needed to retain forecasting accuracy differs between directly trained compact models and restrictions of trained full models. Using function-preserving interventions and exact linear controls, it separates representation exposure from training-covariance coupling as sources of this path dependence, with experiments across six forecasting families.

Critical strengths
  • Careful separation of mechanisms. The paper goes beyond observing differences between training and restriction. Function-preserving interventions, kernel-preserving negative controls, and exact ridge solutions help distinguish representation exposure from training-covariance coupling and incomplete optimization.
  • Substantial robustness analysis. The evaluation spans six forecasting families and includes checks across seeds, metrics, tolerances, and alternative capacity readouts. Reporting complete risk profiles alongside first-hit capacities helps qualify the findings and exposes cases where a single capacity value conceals different behaviors.
Critical weaknesses
  • Practical implications remain insufficiently established. The main comparison considers restriction without subsequent refitting, whereas practical compression commonly includes post-pruning fine-tuning. Moreover, the structural budget KK does not directly measure parameter count or computational cost. A small fine-tuning control or cost-matched comparison would help establish whether the observed capacity gaps materially affect practical compression decisions.
  • Presentation could be more concise and space-efficient. Some main-text figures occupy substantial space relative to the additional information they convey, while their captions and surrounding discussion repeat similar explanations. The authors should reconsider panel sizes and placement, consolidate overlapping illustrations, and tighten the writing to improve information density and reading flow.
Questions for authors

please see weakness above

Reviewer_MX6pOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper provides a deep dive into the workings of NLinear and DLinear: time-series forecasting models. Specifically, the structure of the models is interrogated in a systematic and structured set of experiments.

Critical strengths

The paper provides an analysis of time-series forecasting models. It is a nice paper, which certainly merits consideration for publication, but perhaps in a more specialty journal. Certainly the work is of interest, but quite specialty. And even more special as it is not clear that DLlinear and Linear are the models for sole consideration.

Critical weaknesses

I have found the paper a bit peculiar in the sense that from the beginning, they operate on some time-series models that they have not justified why they considered NLinear and DLinear. It seems very niche and highly restricted. Also, the fundamental question: "How much structure does a forecasting model need to retain its prediction accuracy?" seems much too broad compared what is actually addressed in the manuscript. I do find some fundamental failings of the paper in this regards.

Questions for authors

Can you please justify your choice of time-series models? How much does the analysis actually hold for other time-series models.

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

This paper addresses a key question in time series forecasting: the relationship between model complexity and forecasting accuracy. The authors propose acquisition-conditioned capacity, defined as the first structural budget at which a specified acquisition path recovers the accuracy of a selected full forecaster. They find that direct learning and restriction yield different capacities despite sharing the same full endpoint. This leads to the key finding that a fixed complete forecasting function can have different restricted capacities. Their analysis is based on model restriction.

Critical strengths
  1. This paper addresses an important and emerging problem. Studies such as MixLinear, TimeBases, and SparseTSF have demonstrated that simple models perform well in time series forecasting. There is therefore a need for research that explains why this is the case and identifies directions for future work.

  2. The analysis is sound. In the experiments, factors such as model structure and capacity are controlled to isolate the effect of a single component on forecasting accuracy.

  3. This paper presents many findings that are useful for future research. For example, the finding that “acquisition changes capacity across forecasting structures” helps explain the efficiency of simple models.

Critical weaknesses
  1. The paper appears somewhat too focused on engineering aspects. For example, the large volume of numerical results reported in the experimental sections of the main text tends to obscure the key findings.

  2. The paper does not report some of the underlying numerical results, such as the actual MSE and MAE values.

  3. The paper lacks comparisons with some baselines, such as SparseTSF and TimeBases.

  4. There are some well-known problems with the datasets used in this paper, such as excessive periodicity and insufficient data cleaning. Therefore, the robustness of the findings is questionable.

  5. The experimental section and some of the figures give a strong impression of having been generated by an LLM. Although this does not affect the substance of the paper, it makes the presentation somewhat harder to follow.

Questions for authors
  1. The current experiments mainly compare compact models trained from scratch with pruned models whose retained weights are not updated after pruning. How would the capacity gap change if the authors introduced a third acquisition path involving limited post-pruning fine-tuning using ground-truth labels?

  2. Were the models adequately tuned? The paper currently reports limited results across different hyperparameter configurations. Can the authors demonstrate that these findings hold across a broader range of baseline models and datasets and are statistically significant?

  3. These issues affect the paper’s scientific rigor to some extent, but I find its research question and conclusions valuable. I have therefore assigned a score of 3. I would be happy to raise my score if the authors can adequately address the concerns raised above.

Comments on the paper

1 comment
Area_Chair_bc3FOpenReview ↗

please engage

Dear reviewers,

Please join the discussion with the authors.

Best, AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.