- 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) ?