acceptodds
Under review as a conference paper at ICLR 2027

Rethinking Long-Term Time Series Forecasting as Patch-to-Horizon Contribution Learning

Abstract

Long-term time series forecasting (LTSF) is commonly formulated as learning global representations from historical observations, where the contributions of individual historical patches to different forecasting horizons remain implicit. This representation-centric paradigm not only limits the ability to exploit horizon-dependent historical evidence but also obscures how past observations influence future predictions. To address this issue, we reformulate long-term forecasting as Patch-to-Horizon Contribution Learning (PHCL), a contribution-oriented formulation that explicitly learns the contribution mapping from historical patches to future horizons. Based on this paradigm, we propose Patch-Horizon Lifting (PHL), a contribution-aware forecasting framework that constructs a multi-resolution historical evidence space and adaptively routes resolution-patch candidates to horizon-specific contribution estimators. Extensive experiments on standard long-term forecasting benchmarks demonstrate consistent improvements over representative forecasting methods, validating the effectiveness of horizon-specific contribution learning.

Primary area:
learning on time series and dynamical systems
Keywords:
Long-Term Time Series Forecasting, Patch-to-Horizon Contribution Learning, Multi-Resolution Temporal Modeling, Horizon-Specific Routing, Interpretable Forecasting

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_s52HOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper reformulates long-term time-series forecasting as Patch-to-Horizon Contribution Learning (PHCL). Its instantiation, Patch-Horizon Lifting (PHL), constructs raw, fixed-wavelet, and learnable-lifting patch candidates; uses horizon queries and a recency prior to route a sparse subset of candidates; maps each selected patch to a signed scalar effect; and fuses their weighted sum with a persistence anchor. Experiments on eight standard benchmarks report the best average MSE among the listed methods, while routing diagnostics and visualizations are presented as evidence that different horizons use different historical patches.

Critical strengths
  1. The modeling motivation is intuitive. Different future steps can depend on different parts and resolutions of the history. Combining multi-resolution candidates with horizon-conditioned sparse routing is a reasonable inductive bias for this setting.

  2. The reported forecasting accuracy is strong. PHL improves average MSE on all eight listed datasets, including high-dimensional Electricity, Traffic, and Solar. The gains on several ETT datasets are substantial relative to the reported recent baselines.

  3. The paper makes a useful attempt to inspect internal routing behavior. The router ablations, sparsity analysis, and patch-resolution maps go beyond reporting only aggregate forecasting errors.

Critical weaknesses
  1. The learned quantities are not established as predictive contributions. The softmax weights α\alpha are routing weights, whereas the output terms are αv\alpha v, with vv jointly learned by patch-, position-, resolution-, and horizon-specific experts. Multiple router/expert parameterizations can produce the same forecast but very different maps. The final output is further altered by the anchor and gate. Without perturbation, deletion/insertion, retraining, or synthetic ground-truth tests, the maps are not shown to be faithful explanations of model predictions. Calling them the patches "responsible" for a forecast is therefore stronger than the evidence supports.

  2. The central horizon-specific claim is not isolated by the ablations. The uniform, prior-only, and dense-router variants test different routing and sparsity choices, but there is no parameter-matched horizon-agnostic router or shared-query control. Likewise, the paper does not compare against an otherwise identical model that uses the same multi-resolution candidates, experts, and anchor but a shared historical representation. The results establish that the complete PHL architecture works, not that PHCL is a distinct or necessary forecasting paradigm.

  3. Extra output-head capacity is a major confounder. Equation (7) assigns separate weights Wh,f,p,lW_{h,f,p,l} and biases bh,f,pb_{h,f,p} to every horizon, resolution, and patch position. This is already a direct horizon-specific linear mapping from the full look-back window, in addition to adaptive routing and anchor fusion. The paper does not compare against a dense linear head or a static mixture with the same position-specific parameters, so the gain cannot be attributed specifically to contribution learning.

  4. The scalability cost is omitted despite an unfavorable dependence on horizon and channel count. With the stated L=512L=512, patch length 16, stride 8, and a two-level decomposition, there are 64 patches and 7 resolution components. At H=720H=720, Equation (7) alone implies about 5.16M weights, while router scores scale as BCHFPBCHFP; for Traffic this is about 278M scores per sample before top-kk selection. No parameter count, FLOPs, memory, training time, or inference latency is reported.

  5. The empirical protocol is insufficiently specified for the strength of the accuracy claim. The appendix gives wavelet choices for only ETTh1, ETTm1, and Weather, and omits key optimization details, tuning ranges, seeds, and uncertainty. It is also unclear whether all baseline results were rerun with the same input length, data pipeline, tuning budget, and number of trials. Without variance estimates, the reliability of smaller differences cannot be assessed.

  6. Several diagnostics partly reflect mechanisms imposed by construction. A top-kk contribution mass of 1.00 is expected after masking and renormalization, and high recent-patch mass is encouraged explicitly by Equation (5). Comparisons across H=96H=96 and H=720H=720 also appear to involve separately trained forecast-length settings and do not directly quantify how evidence changes across output steps within one model. These observations do not independently validate horizon-specific predictive responsibility.

  7. There are unresolved reporting and specification issues. The appendix states that PHL has the best average MSE on seven of eight datasets, whereas Tables 1 and A2 show it as best on all eight. The ordering and meaning of the normalized component index ufu_f in the hand-designed prior are not defined clearly, even though raw, fixed-wavelet, and learnable-lifting components do not naturally form one ordinal scale. Complete dataset-specific configurations are also missing.

Questions for authors
  1. Contribution faithfulness: How do the reported maps change under deletion or replacement of high-contribution patches, and do they outperform attention weights, gradients, or occlusion scores in predicting the resulting output change? Can this be tested on synthetic data with known horizon-dependent dependencies?

  2. Central-claim control: Can the authors compare full PHL with a capacity-matched shared-query or horizon-agnostic router, and with a conventional decoder using exactly the same multi-resolution encoder, patch experts, anchor, and parameter budget?

  3. Capacity attribution: How much of the improvement remains if Wh,f,p,lW_{h,f,p,l} is shared across patch positions or replaced by an equally sized dense linear forecasting head? Please report component-wise parameter counts.

  4. Efficiency: What are the measured peak memory, training throughput, and inference latency on Electricity and Traffic at H=720H=720, including the cost of constructing and scoring all candidates before top-kk?

  5. Reproducibility: Were all baselines rerun under exactly the same input length and tuning budget? Please provide complete configurations for every dataset, results over multiple seeds, and uncertainty or paired significance analysis.

  6. Horizon specificity: Can the authors quantify within-model differences among QhQ_h and the resulting selected patches across output steps, rather than mainly comparing models trained for different forecast lengths?

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

This paper proposes the Patch-Horizon Lifting (PHL) method for long-term time series forecasting, utilizing a mechanism that learns contributions from data "patches" toward specific "forecast horizons." Raw, fixed-wavelet, and learnable lifting patches are routed to horizon-specific estimators and then the signed patch contributions are aggregated and fused with an anchor forecast. Experiments across eight datasets demonstrate highly competitive forecasting performance; furthermore, ablation studies validate the effectiveness of the adaptive routing mechanism and the "recency prior" assumption.

Critical strengths
  1. The paper is well-written and generally easy to follow.
  2. In primary comparisons, PHL achieves the highest performance across forecast horizons on all eight datasets.
  3. Ablation studies conducted on the ETTh1 and ETTm1 datasets confirm the value of the adaptive routing mechanism and the recency prior assumption.
Critical weaknesses
  1. The contributions of individual data patches have not been verified through patch-removal tests, nor are they clearly distinguished from a horizon-query-based attention mechanism.
  2. The absence of "anchor-only" ablation studies and results based on multiple random seeds makes it difficult to ascertain the actual extent of improvement contributed by the "contribution branch."
  3. Lack of comparison with forecasting models based on LLM, such as Chronos [1], SCALER [2], and TimeLLM [3].

[1] Ansari, A. F., et al. Chronos: Learning the Language of Time Series. Transactions on Machine Learning Research, 2024.

[2] Le, X.-M., et al. Efficient Test-Time Scaling for LLM-based Time Series Forecasting (SCALER). KDD, 2026.

[3] Jin, M., et al. Time-LLM: Time Series Forecasting by Reprogramming Large Language Models. ICLR, 2024.

Questions for authors

Please address the weaknesses mentioned above.

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

This paper proposes Patch-to-Horizon Contribution Learning (PHCL), a formulation of long-term time series forecasting that explicitly models the contributions of historical patches to different prediction horizons. The proposed Patch-Horizon Lifting (PHL) framework combines multi-resolution decomposition, horizon-specific patch routing, signed contribution estimation, and gated anchor fusion. Experiments on eight forecasting benchmarks demonstrate competitive performance, particularly in terms of MSE.

Critical strengths
  1. Interesting formulation. The paper highlights the importance of horizon-dependent historical evidence and proposes an intuitive patch-to-horizon contribution framework for long-term forecasting.

  2. Reasonable architectural design. The combination of multi-resolution representations, sparse horizon-specific routing, and signed contribution estimation provides a structured approach to forecasting.

  3. Competitive empirical performance. PHL achieves strong MSE results across eight benchmarks, and the ablation studies provide preliminary evidence supporting several design choices.

Critical weaknesses
  1. Limited methodological novelty. The proposed framework primarily integrates existing techniques, including wavelet decomposition, attention-like routing, sparse patch selection, and weighted prediction aggregation. The distinction between the proposed contribution learning and conventional attention-based or horizon-conditioned forecasting is not sufficiently established.

  2. Insufficient validation of the contribution-learning claim. The paper interprets learned routing weights and signed estimates as predictive contributions. However, these quantities are optimized through forecasting loss without direct supervision or validation of their faithfulness. The contribution visualizations alone do not establish that the identified patches are genuinely responsible for predictions.

  3. Incomplete ablation studies. The current ablations mainly evaluate individual architectural components on two datasets. More fundamental comparisons are missing, particularly horizon-independent routing, conventional attention-based aggregation, and variants without the signed contribution estimator or anchor fusion. Consequently, the improvements cannot be clearly attributed to the proposed PHCL formulation.

  4. Limited experimental rigor. Although PHL achieves strong MSE results, its MAE performance is notably weaker than some baselines on datasets such as Solar and Traffic. The paper also lacks repeated-run statistics, significance analysis, and computational efficiency comparisons, making it difficult to assess the robustness and practical trade-offs of the proposed method.

Questions for authors

Q1. What fundamentally distinguishes PHCL from horizon-specific cross-attention or weighted patch aggregation? Could the authors clarify the mathematical or optimization-level difference and provide a comparison against an appropriately matched attention-based baseline?

Q2. How can the learned patch contributions be verified as faithful explanations of predictions? For example, does removing or perturbing patches with high estimated contributions produce larger forecasting changes than perturbing low-contribution patches? Such experiments would substantially strengthen the interpretability claim.

Q3. Since the recency prior produces one of the largest improvements in Table 2, how much of the performance gain comes from the proposed contribution-learning mechanism rather than this handcrafted inductive bias? Would a simpler forecasting architecture with the same prior achieve comparable results?

Q4. In Eq. (9), the fusion gate is horizon-dependent but appears independent of the input instance. How does it adapt to the reliability of historical evidence for individual predictions, as claimed in Section 3.6? Could the authors clarify its behavior and report results without the anchor branch?

Q5. Why does PHL achieve stronger MSE but noticeably weaker MAE than SEER on Solar and Traffic? Does this indicate a trade-off involving outliers or error distributions? Additionally, could the authors report variance across multiple runs and training/inference costs to better establish the practical advantages of PHL?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.