acceptodds
Under review as a conference paper at ICLR 2027

When Does Coarse-to-Fine Forecasting Help? An Exact Error Decomposition and a Checkable Criterion

Abstract

Coarse-grained future structure is often easier to predict than fine-grained dynamics, motivating a broad class of coarse-to-fine and multiscale forecasting methods. Yet predictability does not necessarily imply predictive value: under the same forecasting protocol, conditioning on a coarse-grained future can substantially improve a matched backbone, have little effect, or even degrade the final prediction. This raises a more fundamental question than whether a particular hierarchy works: when can predictable coarse-grained information actually be converted into better fine-grained forecasts? We study this question by separating three quantities that are often conflated by standard forecasting metrics: the predictability of future coarse-grained structure, the value of the resulting coarse anchor relative to a matched unconditional forecaster, and the additional error that can be recovered through fine-grained refinement. This decomposition yields a simple necessary condition that can be evaluated before training the refinement model. We partition the final error into a component determined by coarse prediction and a component that refinement can potentially recover. If the former already exceeds the error of the target direct forecaster, then no refinement can make the hierarchical predictor outperform that baseline. Across 28 dataset–horizon settings from seven multivariate forecasting benchmarks and four prediction horizons, this condition identifies eight such cases, and the hierarchical method loses in all eight. The remaining cases reveal qualitatively different success and failure regimes. In some settings, the coarse anchor itself accounts for nearly all of the improvement; in others, substantial refinement is required. More importantly, a harmful anchor does not necessarily imply failure, nor does improving an anchor guarantee a net gain. On Electricity, refinement fully compensates for an initially harmful anchor and ultimately surpasses the baseline. On long-horizon Traffic, refinement improves the anchor but remains insufficient to outperform the direct forecaster. Targeted interventions further separate information availability from model realization. On ETTm1, coarse future structure remains predictable at long horizons even when the learned hierarchy fails to exploit it. Improving only coarse extraction or only refinement provides partial recovery, whereas improving both produces a strongly coupled rescue effect. On Traffic, increasing coarse resolution reduces unresolved within-block structure and anchor error, turning H = 336 from a net loss into a net gain and substantially narrowing the deficit at H = 720. These results show that coarse-grained predictability and realized predictive value are distinct properties. The success of coarse-to-fine forecasting depends not only on whether coarse future structure exists and can be predicted, but also on whether it can be converted into a useful anchor relative to the target baseline, and whether refinement can recover the information that remains unresolved after anchoring.

Primary area:
learning on time series and dynamical systems
Keywords:
time series forecasting, long-horizon forecasting, coarse-to-fine forecasting, multiscale forecasting, hierarchical forecasting, error decomposition, coarse-grained prediction, refinement

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

This paper studies when coarse-to-fine forecasting improves time-series predictions.

The authors propose an exact decomposition of the prediction error. They separate the error of the coarse forecast from the error of the refinement step.

They then derive a necessary condition for a coarse-to-fine model to beat a direct forecaster.

The experiments cover classical datasets such as ett and classical long horizon. The authors compare several model variants and study the effects of coarse prediction, refinement, and coarse resolution.

The hierarchical model performs worse than the direct baseline. Overall, the hierarchical model beats DLinear in only 4 of the 28 settings.

Critical strengths

The error decomposition is exact and easy to understand. It separates coarse prediction errors from refinement errors. This helps explain why a hierarchical model succeeds or fails.

Authors are intellectually honest since they also report negative results. An important finding is that a large improvement over an underlying model does not always mean better performance than a strong baseline. This seems useful for evaluating hierarchical forecasting methods.

Critical weaknesses

The main error decomposition follows from a standard orthogonal decomposition of squared error. It also requires the refinement step to have zero mean within each block.

This is a strong restriction. A more general refiner could correct errors in the coarse forecast. The proposed necessary condition would then no longer hold.

This limits the scope of the main theoretical contribution.

The authors identify 8 settings where the necessary condition fails. However, these cases are identified using test-set errors. Their failure then follows directly from the mathematical result.

This does not show whether the criterion can predict failures on new data. The paper also focuses on controlled architectures rather than established coarse-to-fine models.

These points make it difficult to assess the practical value and generality of the proposed criterion.

Questions for authors
  • Can you compare to very simple and competitive baseline such as Samformer (results are in the paper you can find on NeurIPS or arxiv) ?
Reviewer_XFEoOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper asks when coarse-to-fine forecasting actually beats direct forecasting. It defines three test errors: N (a matched unconditional backbone), A (the upsampled coarse anchor) and F (the anchor plus refinement). It then defines relative gains G_A, G_R and G, related by the identity G = G_A + G_R − G_A·G_R.

Critical strengths

This paper dose give a useful diagnostic lens. The split into ε²_C and σ²_W does give an interpretable account of why settings differ. Reporting these quantities is a reasonable practice recommendation.

Critical weaknesses

The decomposition is the standard orthogonality between a block-mean projection and its complement. The practical lesson, for example, comparing against a strong direct baseline, not only your own ablated backbone, is already widely accepted in this field.

On channel-normalized data, NOCOND reaches MSE 3.54 (ETTh2-H96), 3.57–4.15 (ETTm2-H720) and 1.30 (ETTh1-H96). DLinear gets 0.29, 0.39 and 0.38 on the same settings. A model this much worse than a linear map, and in several cases worse than predicting the training mean, is not a meaningful reference.

Questions for authors

Why is NOCOND's MSE above 3 on ETTh2/ETTm2? How does it compare with trivial baselines (training mean, last value, seasonal naive)? If the backbone were tuned to DLinear-level error, would the regime structure survive? (This could change my soundness assessment.)

What would the criterion and grid look like with stronger anchors, e.g., block-averaged DLinear or PatchTST forecasts, or a Ridge model tuned on validation? Would any of the eight "structural" settings remain structural?

Reviewer_82oiOpenReview ↗
Rating
1
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper explores cases where coarse to fine forecasting leads to an enhancement in fine-grained predictions. An exact decomposition is obtained to dissect the coarse prediction error and recoverable error within blocks. Necessary condition is obtained for improvement over direct forecaster.

Through experiments over seven datasets for four horizons, eight cases are found to violate the condition. More interventions are done to explore coarse extraction, residual refinement, coarse resolution.

Critical strengths
  1. Good decomposition, informative in terms of diagnosis.

The exact decomposition enables to see what part is error carried forward from coarse predictor, what can be recovered through refinement. The fact that there are cases where dramatic improvement is possible compared to matched backbone while being inferior to direct forecaster is insightful.

  1. Thoughtful experiments on mechanism.

The interaction between coarse extraction and residual refinement is shown via ETTm1 factorial experiment. Additional evidence of inadequate coarse resolution leading to failure in forecasting is presented through the traffic resolution intervention. Moreover, negative results and limitations of their forecasting architecture are admitted.

Critical weaknesses
  1. Weak independent validation on the proposed necessary condition.

This criterion is assessed with test set coarse prediction error and reference error. And since the final error is algebraically lower bounded by coarse prediction error. Then all eight found settings should fail the reference given model structure.

Therefore, the 8/8 result only validates the math identity, rather than validates the potential effectiveness of the criterion in finding failures.

Much stronger validation is needed to measure the criterion's usefulness by testing-only mode with independent test results.

  1. Narrow theoretical coverage and weak empirical generality.

The decomposition assumes block mean preserving residual refinement, and uniform squared-error measurement. The necessary condition doesn't hold in general for refiners that may adjust coarse anchor.

Moreover, the math finding is very similar to the regular orthogonal error decomposition. The theoretical novelty of the math is somewhat small.

One controlled transformer architecture is mostly used for experiments. Meanwhile the other backbone is a customized patch tokenized model. Also, the new hierarchical predictor fails DLinear in 24/28 settings. Which makes it relatively uncompetitive in practical applications.

While mainly diagnostic the work could have more importance by providing more architecture validation and an even stronger indication of utility of the criterion in model selection.

Overall, the paper offers a clean mathematical decomposition and insightful controlled experiments. The distinction between hierarchical improvement and real predictive power is interesting.

But the core necessary condition follows from imposed decomposition and its empirical verification does not show its value in prospective decision making. In addition the architecture is very restricted and there is little theoretical novelty.

Finally, the written of the paper need to be largely improved. it is hard to read.

Questions for authors

Q1: Is it possible to have the necessary condition estimated (training or validation data) and to test its predictive power for failure on an independent test data?

Q2: How does the criterion react if refiner is allowed to tinker with means of the coarse blocks? Is the criterion applicable to anything else than the current residual architecture that preserves the mean?

Q3: Can you show that the same decomposition is helpful on other coarse to fine architectures/coarse resolutions (esp. in comparison to better direct forecasting models?)

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.