acceptodds
Withdrawn from ICLR 2027

StateComp: Shared Transfer of Seasonal States for Compact Forecasting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Long-term time-series forecasting requires preserving fine-grained temporal structure over many future steps, yet learning a fully flexible history-to-future transformation can be parameter-intensive. We explore a complementary hypothesis: time-series shapes may be complex, while the rules governing their evolution can be comparatively simple and shared. Based on this idea, we propose , a compact forecaster that separates the representation of temporal states from their learned transfer. StateComp constructs multi-width seasonal states directly from each input window and propagates them using a small set of coefficients shared across phases and variables. A low-rank history correction complements these prescribed states by capturing residual dependence that seasonal summaries miss. This design allows rich, input-dependent temporal structure to be retained without assigning learned parameters to every phase or variable. We further analyze the interaction between the two components, showing that shared seasonal transfer can represent multiple predictive directions and that correction rank controls the remaining error relative to the coverage of the state dictionary. Across five long-term forecasting benchmarks, StateComp achieves competitive accuracy with only 61–389 learned parameters and obtains lower MSE than SparseTSF and MixLinear at every evaluated horizon on Electricity and Traffic.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Time Series, Parameter Efficiency

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_wQCGOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes StateComp, a compact long-term time-series forecasting framework that separates input-derived temporal state representation from the learned temporal transfer. The model constructs multi-width seasonal states and longer-cycle contrasts directly from each input window, transfers these states using coefficients shared across phases and variables, and adds a low-rank temporal correction to account for residual historical dependence. The central hypothesis is that temporal shapes can be highly detailed while their evolution can be governed by a small number of shared coefficients.

Critical strengths

s1: The experimental evaluation is comprehensive. The authors attempt to demonstrate the effectiveness of the proposed model from multiple perspectives.

s2: StateComp achieves competitive performance with an extremely small number of learnable parameters, which is an impressive aspect of the proposed method.

Critical weaknesses

w1: The abstract states that “time-series shapes can be highly complex, while the dynamics governing their evolution may be relatively simple and consistent.” However, the paper does not provide sufficient justification for this assumption. In real-world time series, temporal patterns can be highly complex and dynamic, and they may be influenced by a variety of factors. Therefore, it is not necessarily true that the underlying evolution dynamics are simple or consistent. The authors should provide stronger theoretical or empirical evidence to support this assumption and clarify under what conditions it is expected to hold.

w2: From my understanding, StateComp appears to largely recombine and structurally redesign ideas that have already been explored in existing methods, particularly CycleNet. Therefore, the novelty of the proposed method is not sufficiently clear. The authors should explicitly clarify what the key innovations of StateComp are and how it fundamentally differs from CycleNet and other existing compact time-series forecasting methods. In particular, the paper should clearly identify which components or modeling principles are genuinely novel.

w3: The Related Work section does not sufficiently discuss several representative lightweight/simple time-series forecasting methods, such as SparseTSF, CycleNet, TimeBase, and MixLinear. Since the main motivation of StateComp is to achieve competitive forecasting performance with an extremely small number of learnable parameters, these methods are highly relevant to the positioning and novelty of the proposed approach. The authors should provide a more comprehensive discussion and comparison with these works.

w4: The experimental evaluation is mainly conducted on relatively small benchmark datasets. To better demonstrate the scalability and practical effectiveness of StateComp, I recommend evaluating the proposed method on larger-scale time-series datasets. Experiments on datasets with substantially more variables and/or longer time series would provide stronger evidence that the proposed extremely low-parameter design remains effective beyond the current benchmark settings.

Questions for authors

See the Weaknesses.

Reviewer_ew4uOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The authors propose a parameter-efficient method with only 61–389 parameters. Their assumption is that you can manually do simple feature engineering upfront and just learn the transfer with very few parameters to come up with more complex features. This is done by taking each input window and computing seasonal averages over several widths, and by doing a history correction.

Critical strengths

1. Very small model. It is really cool how tiny a model can be and still reasonably solve the task. All coefficients are shared across variables, so even Traffic with 862 variables is forecast with fewer than 100 learned parameters.

2. Interpretable by design. The method is inherently readable and interpretable: each learned coefficient scales a whole seasonal state computed from the input window (for example, a daily profile averaged over the last w days), so the learned gains can be read off directly.

Critical weaknesses

1. Unclear motivation and missing context. I am really missing the core motivation of the paper: why is it important for a model to be so tiny that one could, in theory, write the computation on a piece of paper? Efficiency is an important concept, but the paper does not make the case for it. The abstract states that "learning a fully flexible history-to-future transformation can be parameter intensive", but is this really a problem, and in which situation? Existing compact forecasters such as DLinear, SparseTSF, and CycleNet are already cheap to train and run.

In my opinion, the related work section is also not optimal: it lists many architectures without actually bringing them into context in terms of efficiency. In addition, the paper does not acknowledge concept bottleneck approaches for forecasting (e.g., van Sprang et al., 2024), which seem relatively close in focus, since they also route the forecast through a small set of interpretable, prescribed quantities.

2. Unclear practical benefit. The accuracy could be better, or at least I do not see the benefit. In Table 7, StateComp has the lowest MSE (or ties for it) in only 2 of the 20 settings: CycleNet/Linear is better at every Electricity horizon, ...

I also do not see the practical value of moving from ~900 parameters to a few dozen. I am really missing a comparison of FLOPs, runtime, and memory, and an accuracy–cost trade-off. Moreover, the parameter count excludes structural choices and fixed constants.

Finally, I think the method is unnecessarily complex, especially for so few parameters, and I fail to see how the model can actually utilize all of these tricks. For example, the history correction changes the MSE by at most 0.002 on Traffic (Tables 9 and 11), this does not sound significant to me.

Questions for authors

Clarification requests

  1. I am not sure about the ablations and the role of the individual components. Could the authors explain Table 1? It is only referenced in Appendix F.1. The caption states: "The shared map uses 8, 13, 14, and 28 parameters by horizon; state-only uses 27 and the full model uses 61 to 87." This seems like an unfair comparison. Would it not make more sense to keep the number of parameters per component identical to see the actual contribution, or at least to match state-only and the shared weekly map? With so few parameters, this might be more important than in the ablations of other papers.
  2. ETTh1/2, ETTm2, Electricity, and Traffic all have a clear daily (and weekly) period. Could the authors show what happens on datasets with weak or no periodicity, such as Exchange Rate, and on other standard benchmarks such as Weather?

Suggestions for presentation

  • Figure 2 could focus on a few examples, with the rest moved to the appendix. It is barely readable, and the different scales make it hard to decipher anything.
  • Section 3.2 states that "the history branch answers the third question", or at least i cannot find it in the rest of the paper.
Reviewer_kaf5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes StateComp, a compact time-series forecaster that separates input-derived seasonal states from their learned transfer. Multiwidth base-cycle profiles and longer-cycle contrasts are combined through coefficients shared across phases and variables, with a low-rank history correction capturing residual dependencies. Theoretical results explain the capacity of prescribed seasonal routing under stylized assumptions. Experiments on five benchmarks report competitive accuracy with 61–389 learned parameters, supported by seasonal controls, branch ablations, rank sweeps, and initialization studies.

Critical strengths
  1. The separation of input-derived seasonal detail from shared transfer coefficients yields a useful accuracy–parameter trade-off. With 61–389 parameters, StateComp remains competitive and consistently improves on SparseTSF and MixLinear in MSE on Electricity and Traffic.

  2. The analysis and ablations support the architectural rationale. The theory illustrates how prescribed routing preserves multiple predictive directions, while seasonal controls and rank sweeps help distinguish the contributions of the state bank and history correction.

Critical weaknesses
  1. Table 7 combines published results and logs without a consolidated account of baseline input lengths and tuning budgets; Appendix H.1 also does not enumerate the complete ETTm2 search space and selected configurations.

  2. Practical efficiency remains unquantified beyond trainable parameter counts. No training-time, inference-latency, or memory comparisons are reported, despite multiwidth state construction and 1,000 alternating initialization cycles. Table 16 suggests that much cheaper initialization may suffice.

  3. The curve in Figure 1 is a bit elegant and not very aesthetically pleasing, resembling lines generated by AI, and needs to be modified.

  4. The article is suspected to have extensively used AI polishing or generation, such as the use of the marker word "therefore" (totally 7 times) in sentences.

Questions for authors

Weaknesses and the following:

  1. Could you provide the complete per-horizon ETTm2 configurations and search space, and clarify whether the closest compact baselines use matching look-back windows and comparable tuning? Do the small MSE advantages persist under a matched evaluation?

  2. How do end-to-end training cost, inference latency, and peak memory compare with SparseTSF and MixLinear on the same hardware? Given Table 16, would one-cycle initialization preserve the reported accuracy while reducing setup cost?

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

Related papers

Discussion (0)

Sign in to comment.