acceptodds
Under review as a conference paper at ICLR 2027

Latent Propagation for Continual-Context Time-Series Forecasting

Abstract

Time series forecasting is a fundamental task in machine learning, aiming to predict future horizons from historical observations. Prior methods have largely been developed under a *fixed-length formulation*, where a continuous time series is divided into input-output pairs of fixed-length input and prediction windows. This formulation, however, treats each forecasting instance independently, relying only on the observations within the fixed-length input window without maintaining context beyond it. As a result, it cannot leverage useful information accumulated from past observations over time. In this paper, we consider a *continual-context formulation*, where historical context can be accumulated over time beyond the fixed-length input window. Based on this formulation, we propose **CoLP**, which constructs **Co**ntext-carrying **L**atents for each instance and **P**ropagates them along the instance stream, enabling context to accumulate over time. Specifically, our framework treats queries corresponding to future horizons as latent contexts that aggregate prediction-relevant historical information and carry it forward to later instances. Through extensive experiments, we empirically demonstrate the superiority of our framework over baselines across various benchmarks, highlighting the potential of the continual-context formulation.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Forecasting, Transformer, Attention

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 4 reviews, above 85% of the others.
Reviewer_Y54xOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Traditional time series prediction models use a fixed-length input window for each prediction and cannot explicitly retain information accumulated from previous prediction instances. The paper proposes CoLP to address this limitation. The experiments cover 8 commonly used benchmark datasets and 4 prediction lengths. The paper reports that CoLP achieves 13 best and 2 second-best results in the aggregated results of 16 datasets. This proves the effectiveness of this paradigm.

Critical strengths
  1. The problem setting has practical significance. In reality, time series data arrives in a streaming manner, and historical observations may indeed provide useful information for subsequent predictions.

  2. CoLP uses predicted position-aligned query tokens to represent the potential state, and then reads the historical state through cross-attention. This design makes the memory units have clear predictive semantics compared to directly passing the hidden states of traditional RNNs.

  3. The experiment coverage is relatively comprehensive.

Critical weaknesses
  1. If CoLP predicts every LL time steps, while the baseline still uses a standard sliding window and generates predictions at each time step, the sets of prediction starting points used to compute the test error may differ. This can affect the statistical meaning of the metrics and also the comparability of the results.

  2. Numerical differences exist between the main results and ablation experiments, such as weather MSE 0.215 in table 1 and 0.229 in table 2.

Questions for authors
  1. Are all models making predictions at exactly the same timestamps? What is the window stride during the test phase? Are MSE and MAE averaged based on the prediction start points or all prediction points?

  2. The core claim of the paper is that latent states can continuously accumulate and reuse information across multiple prediction instances. However, the training stage uses a limited context length G=4G=4, while the inference propagation length experiments on page 9 only show Tprop∈1,2,4,8T_{\mathrm{prop}}\in{1,2,4,8}. Can the model maintain performance and stability without resetting and during long-term operation?

3 comments
AuthorsOpenReview ↗

Response to Reviewer Y54x

Dear Reviewer Y54x,

We sincerely appreciate your careful review and thoughtful comments. In what follows, we address your concerns one by one. Note that we only report MSE results in this rebuttal due to the space limit. Please refer to the revised manuscript, as all our responses follow its updated content and numbering.


[W1] If CoLP predicts every LL time steps, while the baseline still uses a standard sliding window and generates predictions at each time step, the sets of prediction starting points used to compute the test error may differ. This can affect the statistical meaning of the metrics and also the comparability of the results.

[Q1] Are all models making predictions at exactly the same timestamps? What is the window stride during the test phase? Are MSE and MAE averaged based on the prediction start points or all prediction points?

[Response to W1, Q1] Thank you for raising this important point of clarification. We clarify that CoLP and all baselines follow the same sliding-window evaluation protocol with a stride of 1, ensuring predictions at exactly the same timestamps. Specifically, CoLP propagates latent states between non-overlapping windows separated by LL time steps and maintains independent latent states for each window offset, allowing predictions at every time step. MSE and MAE are computed over the same prediction starting points and prediction horizons for all models, ensuring a fair comparison under identical evaluation conditions. We include these details in the revised manuscript (Appendix A.1).


[W2] Numerical differences exist between the main results and ablation experiments, such as weather MSE 0.215 in table 1 and 0.229 in table 2.

[Response to W2] Thank you for pointing this out. The reported values are not inconsistent, as the differences arise from different experimental settings. As stated in Section 5.1, Table 1 reports the best performance selected from multiple input lengths (i.e., input length is a hyperparameter and selected through validation), whereas Table 3 in the revised manuscript and all other ablation studies use a fixed input length of L=96L=96. Therefore, the reported values are consistent with their respective experimental settings.


[Q2] The core claim of the paper is that latent states can continuously accumulate and reuse information across multiple prediction instances. However, the training stage uses a limited context length G=4G=4, while the inference propagation length experiments on page 9 only show Tprop∈1,2,3,4T_{prop}\in\\{1,2,3,4\\}. Can the model maintain performance and stability without resetting and during long-term operation?

[Response to Q2] Thank you for this important question. To examine CoLP's stability during long-term operation, we evaluate its performance on the final 30% of the test stream with an extended range of inference propagation lengths Tprop∈1,2,4,8,32,128,∞T_{prop}\in\\{1,2,4,8,32,128,\infty\\}.

As shown below, performance improves as TpropT_{prop} increases and then stabilizes at longer propagation lengths, consistent with our findings in Section 4.5. In particular, the no-reset setting Tprop=∞T_{prop}=\infty maintains strong performance, demonstrating that CoLP can effectively operate over long inference streams without requiring periodic resets. We attribute this stability in part to our training strategy and latent propagation design. The training segment length G=4G=4 allows the model to learn how to accumulate and retain historical context across consecutive instances, rather than limiting context accumulation to four instances. As shown in Figure 5(a), G=4G=4 provides a good balance between forecasting performance and training efficiency, as performance gains begin to saturate with longer training segments. Furthermore, our cross-attention-based propagation mechanism allows the current latent state to selectively incorporate relevant information from the accumulated historical context.

We include this additional analysis (Section A.6) and extend Figure 5(b) to cover the expanded range of propagation lengths in the revised manuscript.

TpropT_{prop}ETTm1WeatherSolar
10.4610.3130.319
20.4190.3060.267
40.4110.2980.210
80.4080.2960.208
320.4080.2970.207
1280.4080.2960.208
∞0.4070.2960.208

We hope that our responses have addressed your concerns and clarified the key aspects of our framework. We would be happy to address any additional questions you may have.

Thank you for your valuable time.

Best regards,
Authors

Thank you for your reply, which solved my problem. I will keep my positive score.

AuthorsOpenReview ↗

Response to Reviewer Y54x

Dear Reviewer Y54x,

Thank you very much for your thoughtful comments and for taking the time to review our manuscript. We are glad that we were able to address and improve upon the points you raised, including the necessary clarifications of our paper.

We hope our work will inspire further research toward practical time-series forecasting.

If you have any further questions, we would be happy to hear them.

Best regards,
Authors

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

COLP studies forecasting when earlier observed windows remain available through a persistent state. Prediction queries first summarize current patches, then attend to the previous latent state. Training propagates detached states across chronological windows; inference continues this process. The evaluation compares state designs, history lengths, and continual-context baselines.

Critical strengths
  1. A credible operational problem with a concrete solution. Sliding-window evaluation can discard useful earlier observations, whereas persistent state can retain them with bounded storage. Prediction-query propagation connects the carried representation to future prediction and is more specific than merely adding memory.
  2. The experiments directly investigate the proposal. Propagation ablations, history/reset-length studies, a stated matched conventional-state control in Table 4, and continual SegRNN/MambaTS comparisons support useful history reuse. Multi-seed results and the plug-in study further strengthen the evidence.
Critical weaknesses
  1. The distinctive benefit of prediction-aligned memory remains insufficiently isolated. The horizon queries introduce an architectural bias, but conventional recurrent states trained for forecasting are also shaped by prediction loss. The methodological question is therefore why the carried memory should itself be organized around prediction horizons. Table 4 provides positive evidence for the query-based design, but the description does not establish whether its advantage comes specifically from horizon alignment or from differences in memory update and readout. Table 3 compares different recurrent architectures and cannot resolve this distinction.
  2. What performance is obtained when the scale token is retained but cross-instance propagation is disabled, with the remaining architecture and training unchanged? Without this control, the additional gains cannot be clearly attributed to preserving historical scale information rather than improving the representation of the current window.
Questions for authors
  1. Table 4 states that the conventional-state control is matched. What are its actual state-update operator and decoder/readout, so that the comparison can establish the distinctive value of prediction-aligned query states?
  2. What performance is obtained when the scale token is retained but cross-instance propagation is disabled, with the remaining architecture and training unchanged?
3 comments
AuthorsOpenReview ↗

Response to Reviewer XquA

Dear Reviewer XquA,

We sincerely appreciate your thoughtful feedback and for highlighting important points. In what follows, we address your concerns one by one. Note that we only report MSE results in this rebuttal due to the space limit. Please refer to the revised manuscript, as all our responses follow its updated content and numbering.


[W1] The distinctive benefit of prediction-aligned memory remains insufficiently isolated. The horizon queries introduce an architectural bias, but conventional recurrent states trained for forecasting are also shaped by prediction loss. The methodological question is therefore why the carried memory should itself be organized around prediction horizons. Table 4 provides positive evidence for the query-based design, but the description does not establish whether its advantage comes specifically from horizon alignment or from differences in memory update and readout. Table 3 compares different recurrent architectures and cannot resolve this distinction.

[Q1] Table 4 states that the conventional-state control is matched. What are its actual state-update operator and decoder/readout, so that the comparison can establish the distinctive value of prediction-aligned query states?

[Response to W1, Q1] Thank you for raising this important point. We would like to note that the experiment (Table 5 in the revised manuscript) was specifically designed to isolate the effect of horizon alignment in the propagated memory. The two designs differ only in how the propagated memory is constructed, while using the same forecasting queries and decoder/readout.

Specifically, CoLP uses the representations generated from horizon-aligned queries as the memory propagated to the next forecasting instance. In contrast, the conventional design constructs its memory using separate learnable tokens without explicit horizon alignment, while retaining the same horizon-aligned queries for forecasting. Both designs use the same cross-attention operations for latent construction (Eq. 8) and propagation (Eq. 9), as well as the same state size, linear decoder (Eq. 10), and training setup.

As shown in Table 5, CoLP achieves better performance under these controlled conditions, supporting the benefit of explicitly aligning the propagated memory with individual prediction-horizon patches. We further clarify these implementation details in the revised manuscript (Section 5.4), where we have also replaced “prediction-aligned” with “horizon-aligned” for clarity.


[W2] What performance is obtained when the scale token is retained but cross-instance propagation is disabled, with the remaining architecture and training unchanged? Without this control, the additional gains cannot be clearly attributed to preserving historical scale information rather than improving the representation of the current window.

[Q2] What performance is obtained when the scale token is retained but cross-instance propagation is disabled, with the remaining architecture and training unchanged?

[Response to W2, Q2] Thank you for this important question. To better understand the contribution of the scale-aware token, we additionally evaluate its effect without latent propagation, while keeping the remaining architecture and training setup unchanged.

As shown below, adding the scale-aware token reduces MSE by 2.6% on average without propagation, while the improvement increases to 4.8% when propagation is enabled. These results support the role of the scale-aware token in preserving useful historical scale information across instances, as its contribution becomes greater when combined with latent propagation, beyond its benefit to the current input representation. We include these additional results and the corresponding discussion in the revised manuscript (Section 5.3).

Componentsw/o scale, w/o propw/ scale, w/o propw/o scale token, w/ propw/ scale token, w/ prop
ETTh10.4230.4100.4200.398
ETTm20.2770.2760.2730.270
Weather0.2420.2450.2290.229
Solar0.2650.2430.2260.197

We hope that our responses have addressed your concerns. If you have any further questions or would like to discuss any remaining points, we would be glad to continue the discussion.

Thank you for your valuable time.

Best regards,
Authors

Thanks for the authors' efforts. I decided to maintain current score.

AuthorsOpenReview ↗

Response to Reviewer XquA

Dear Reviewer XquA,

Thank you very much for your thoughtful comments and for taking the time to review our manuscript. We are glad that our additional clarifications and experiments have addressed your concerns.

We hope our work will encourage future research to explore continual-context forecasting as a promising direction.

If you have any further questions, we would be happy to hear them.

Best regards,
Authors

Reviewer_p2MQOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates TSF under a continual-context formulation. It aims to overcome the limitations of conventional fixed-length forecasting, which discards historical information beyond the current input window. The authors propose CoLP, which constructs prediction-aligned latent tokens and propagates historical information across consecutive forecasting instances through attention mechanism. A scale-aware token is further introduced to preserve amplitude information across windows. Experiments on several TSF benchmarks demonstrate competitive performance against existing baselines, supported by ablation studies examining the effectiveness of latent propagation and related design choices.

Critical strengths

This paper explores continual-context time series forecasting, addressing the limitations of conventional fixed-length forecasting in retaining and leveraging historical information. The proposed CoLP framework incorporates prediction-aligned latent tokens and a latent propagation mechanism to accumulate historical information across consecutive forecasting instances. The overall architecture is easy to understand. The proposed method achieves competitive results across eight widely used time series forecasting benchmarks. The authors also provide several ablation studies to investigate the effectiveness of the proposed components and training strategies.

Critical weaknesses
  1. My primary concern is the limited novelty of the proposed method. The core idea appears conceptually similar to existing memory-based or pattern-prototype-based approaches, where a set of tokens is used to retain global historical information or store representative temporal patterns. I encourage the authors to clarify the key methodological distinctions and novel contributions of CoLP compared with existing works, such as Attraos, RAFT, and MemCast.

  2. Table 1 does not indicate the publication years of the compared models, making it difficult to assess whether the selected baselines represent the latest and most competitive methods. Furthermore, several baseline models are not adequately discussed in the Related Work section, and the paper does not clearly explain the criteria used for baseline selection. More importantly, the experimental comparisons lack representative memory-based forecasting approaches that explicitly leverage historical knowledge or stored temporal patterns, as mentioned in my first concern. Such comparisons would be important for establishing the effectiveness and novelty of CoLP.

  3. The paper has several issues regarding its organization and presentation. First, the Related Work section is placed near the end of the paper, and more importantly, it lacks a systematic discussion of existing approaches involving memory mechanisms, representative pattern storage, and techniques for leveraging historical streaming data in other domains. Second, the Introduction would benefit from a concise, itemized list of the main contributions, rather than presenting them in several lengthy paragraphs.

  4. Additionally, the Experiments section has some presentation issues. For example, why are Lines 300–308 structured as a series of questions to introduce the entire section? This presentation appears rather formulaic and gives the impression of AI-generated writing without sufficient human revision.

Questions for authors
  1. Please clarify how your work differs from existing techniques for historical pattern memorization, storage, and retrieval/replay, both in time series forecasting (TSF) and in other research domains. In particular, what are the key methodological innovations of CoLP compared with these existing approaches?

  2. Please reconsider the organization of the Related Work,Experiments, and Introduction sections. Additionally, the Experiments section has some presentation issues.

  3. Please clarify the rationale and criteria for selecting the competing models in Table 1, particularly the omission of representative memory-based forecasting methods.

Among the concerns raised above, clarifying the methodological novelty is my highest priority. I currently consider this paper to be borderline. If the authors can adequately address most of my concerns during the rebuttal, I would be willing to increase my rating.

8 comments
AuthorsOpenReview ↗

Response to Reviewer p2MQ [1/2]

Dear Reviewer p2MQ,

We sincerely thank you for your helpful feedback and insightful comments. In what follows, we address your concerns one by one. Note that we only report MSE results in this rebuttal due to the space limit. Please refer to the revised manuscript, as all our responses follow its updated content and numbering.


[W1] My primary concern is the limited novelty of the proposed method. The core idea appears conceptually similar to existing memory-based...

[Q1] Please clarify how your work differs from existing techniques for historical pattern memorization, storage, and retrieval/replay, both in time series forecasting (TSF) and in other research domains. In particular, what are the key methodological innovations of CoLP compared with these existing approaches?

[Response to W1, Q1] Thank you for raising this important point. The key distinction of CoLP lies in how historical context is accumulated and updated across forecasting instances. We clarify the differences from memory-based approaches in two aspects.

Comparison with recurrent and state-space approaches. Methods such as Attraos [1], SegRNN [2], and MambaTS [3] accumulate historical information through memory or recurrent mechanisms, but operate within the fixed-length formulation. Their memory states are constructed within each forecasting instance (single forecasting instance = window of length L), whereas CoLP propagates its latent state across consecutive instances, allowing historical context to accumulate beyond the current input window. Although recurrent models can be extended to the continual-context formulation, our experiments in Section 4.4 show that directly propagating their states across instances degrades performance, highlighting the importance of CoLP's latent propagation design.

Comparison with retrieval-based approaches. Another line of work, including RAFT [4], MemCast [5], and GTR [6], retrieves historical patterns or learned representations constructed from training data for each prediction, rather than propagating historical context across forecasting instances. In contrast, CoLP continuously updates and propagates its latent state as new observations arrive, allowing historical context to accumulate over time. Since these methods also leverage information beyond the current input window, we agree that they are relevant baselines for comparison.

Please refer to our response to [W2-2, Q3] and Table 2 in the revised manuscript for the additional experimental comparisons.


[W2-1] Table 1 does not indicate the publication years of the compared models, making it difficult to assess whether the selected baselines represent the latest and most competitive methods.

[Response to W2-1] Thank you for this helpful suggestion, which improves the clarity of our manuscript. We have added the publication years of all compared models to Table 1, 2 in the revised manuscript, making it easier to identify when each model was published.

[W2-2] Furthermore, several baseline models are not adequately discussed in the Related Work section, and the paper does not clearly explain the criteria used for baseline selection. More importantly, the experimental comparisons lack representative memory-based forecasting approaches...

[Q3] Please clarify the rationale and criteria for selecting the competing models in Table 1, particularly the omission of representative memory-based forecasting methods.

[Response to W2-2, Q3] Thank you for this valuable comment. Please note that we divide memory-based approaches into two categories, (i) recurrent and state-space approaches and (ii) retrieval-based approaches, as in our response to [W1, Q1].

First, we select representative forecasting models from diverse architectural families, including Transformer-based, MLP-based, recurrent, and state-space models. In particular, Table 1 already includes recurrent and state-space architectures such as SegRNN [2], MambaTS [3], and S-Mamba [7]. These methods maintain historical information through recurrent or state-space mechanisms within the fixed-length formulation, providing relevant comparisons with CoLP's latent propagation framework.

Second, we agree that retrieval-based approaches are also important baselines, as they leverage historical information beyond the current input window. Following your suggestion, we extend our comparisons to include representative retrieval-based methods [4, 6] under the same input window length. As shown in Table 2 in the revised manuscript, CoLP achieves 13 best and 3 second-best results across 16 comparisons, demonstrating its strong forecasting performance even against methods that leverage additional historical information.

We have clarified the related methods (Section 2) and baseline selection criteria (Section 5.1), and incorporated these comparisons (Section 5.2) in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer p2MQ [2/2]

[W3] The paper has several issues regarding its organization and presentation. First, the Related Work section is placed near the end of the paper, and more importantly, it lacks a systematic discussion of existing approaches involving memory mechanisms, representative pattern storage, and techniques for leveraging historical streaming data in other domains. Second, the Introduction would benefit from a concise, itemized list of the main contributions, rather than presenting them in several lengthy paragraphs.

[Response to W3] Thank you for these constructive suggestions. We have revised the organization and presentation of our manuscript to make our contributions clearer.

First, we have moved the Related Work section immediately after the Introduction. We believe this organization helps readers better understand the differences between CoLP and existing approaches before presenting the formal formulation and methodology. We have also expanded the discussion of related methods, as detailed in our responses to [W1, W2].

Second, we have revised the Introduction to present our main contributions in a concise, itemized list, making the key contributions easier to identify.

We sincerely appreciate your suggestions and would be happy to hear any further feedback on these revisions.


[W4] Additionally, the Experiments section has some presentation issues. For example, why are Lines 300–308 structured as a series of questions to introduce the entire section? This presentation appears rather formulaic and gives the impression of AI-generated writing without sufficient human revision.

[Response to W4] Thank you for sharing your concern. We respectfully clarify that organizing the Experiments section around research questions was an intentional choice to help readers clearly identify the purpose of each experiment and locate the corresponding results. This structure was not AI-generated, and similar approaches have been adopted in well-established research papers [8, 9] to improve the clarity and organization of experimental evaluations. We believe this structure helps readers understand how each experiment addresses a specific research question.

We appreciate your feedback on the presentation and remain open to further discussion.


[Q2] Please reconsider the organization of the Related Work,Experiments, and Introduction sections. Additionally, the Experiments section has some presentation issues.

[Response to Q2] Thank you for your suggestions. We have addressed these concerns in our responses to [W3] and [W4], where we discuss the revisions to the Related Work and Introduction sections and clarify the organization of the Experiments section. Please refer to those responses for further details.

We sincerely appreciate your thoughtful feedback, which has helped improve the clarity and presentation of our manuscript.


References

[1] Hu et al., Attractor Memory for Long-Term Time Series Forecasting: A Chaos Perspective, NeurIPS 2024
[2] Lin et al., SegRNN: Segment Recurrent Neural Network for Long-Term Time Series Forecasting, IEEE Internet of Things Journal 2026
[3] Cai et al., MambaTS: Improved Selective State Space Models for Long-Term Time Series Forecasting, Pattern Recognition 2026
[4] Han et al., Retrieval Augmented Time Series Forecasting, ICML 2025
[5] Tao et al., MemCast: Memory-Driven Time Series Forecasting with Experience-Conditioned Reasoning, ICML 2026
[6] Cao et al., Enhancing Multivariate Time Series Forecasting with Global Temporal Retrieval, ICLR 2026
[7] Wang et al., Is Mamba Effective for Time Series Forecasting?, Neurocomputing 2025
[8] Cai et al., LightGCL: Simple yet Effective Graph Contrastive Learning for Recommendation, ICLR 2023
[9] Yu et al., Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think, ICLR 2025


We appreciate your valuable feedback and hope that our responses have addressed your concerns. We would be happy to continue our discussion to further improve our paper, so please do not hesitate to share any additional thoughts.

Thank you for your valuable time.

Best regards,
Authors

Reviewer_p2MQOpenReview ↗

Response to authors

Thank you very much to the authors for their responses. While they have addressed some of my concerns, the authors also noted in their rebuttal that "the key distinction of CoLP lies in how historical context is accumulated and updated across forecasting instances." This appears to draw heavily upon, or leverage, methodologies from continual learning, online learning, or incremental learning. However, the rebuttal does not seem to clarify what makes CoLP innovative compared to these existing approaches, which are already well-established in fields like NLP and CV. Consequently, I remain skeptical regarding the novel contribution this paper offers to the TSF community. Therefore, I maintain my original recommendation. Best regards.

Reviewer_6NLmOpenReview ↗
Rating
3(+1)
Confidence
3(−1)
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The authors consider time-series forecasting in a continual-context setting, where useful historical information may extend beyond the current fixed-length input window. They propose CoLP, which carries and propagates latent context across consecutive forecasting instances so that earlier observations can inform later predictions. The reported results suggest that this approach can improve forecasting performance. However, the extent to which these gains come from latent propagation itself, rather than simply providing access to a longer effective history, requires closer examination.

Critical strengths

1)The paper addresses a meaningful limitation of fixed-length forecasting: useful historical information may extend beyond the current input window. 2)Propagating latent context across consecutive forecasting instances is well aligned with this motivation.

Critical weaknesses
  1. CoLP has access to historical information beyond the current input window, whereas the main fixed-length baselines do not. A comparison with methods given a comparable amount of historical information would better isolate the benefit of the proposed propagation mechanism. 2) CoLP does not achieve the best result in every setting. For example, other methods obtain lower MSE on Traffic, Solar, and ETTh2. The performance claims should therefore be stated more carefully. 3)The main component ablation covers only four of the eight datasets, leaving the consistency of each component's contribution across all benchmarks less clear. 4)Training propagates context over a short segment of consecutive instances, while inference can carry the latent state over a much longer stream. Further analysis of long-run state drift and reset strategies would strengthen the continual-context claim. 5) The authors should additionally report CoLP’s results with an input length of 336 across all datasets and prediction horizons. This commonly used setting would make its performance easier to compare with prior work and help assess whether the reported gains depend on the selected input length.
Questions for authors

My questions for the authors are detailed in the weaknesses above. In particular, I would appreciate clarification on the fairness of the historical-context comparison and additional CoLP results with an input length of 336.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 6NLm [1/3]

Dear Reviewer 6NLm,

We sincerely appreciate your thoughtful feedback and constructive suggestions. In what follows, we address your concerns one by one. Note that we only report MSE results in this rebuttal due to the space limit. Please refer to the revised manuscript, as all our responses follow its updated content and numbering.


[W1] CoLP has access to historical information beyond the current input window, whereas the main fixed-length baselines do not. A comparison with methods given a comparable amount of historical information would better isolate the benefit of the proposed propagation mechanism.

[Response to W1] Thank you for this important suggestion. We conduct additional experiments comparing CoLP with fixed-length baselines given the same amount of historical information. Specifically, CoLP uses Tprop=4T_{prop}=4 consecutive windows with L=384L=384, while TimePerceiver and PatchTST use an input window of L=1536L=1536. As shown below, CoLP consistently outperforms both baselines across all six datasets. These results demonstrate that CoLP's improvements are not simply due to access to more historical information, but also support the effectiveness of latent propagation in utilizing that information for forecasting. We include these results in the revised manuscript (Appendix A.6).

ModelsOurs (4×384)TimePerceiver (1536)PatchTST (1536)
Electricity0.1560.1620.166
Weather0.2160.2220.224
ETTh10.3980.4140.441
ETTh20.3390.3560.348
ETTm10.3340.3530.358
ETTm20.2490.2550.276

[W2] CoLP does not achieve the best result in every setting. For example, other methods obtain lower MSE on Traffic, Solar, and ETTh2. The performance claims should therefore be stated more carefully.

[Response to W2] Thank you for the valuable comment. To provide further evidence for our performance claims, we compare the performance and variance of CoLP with TimePerceiver, the second-best baseline in Table 1. As shown below, CoLP achieves better performance in 25 of 32 comparisons and exhibits smaller variances overall, demonstrating its strong and stable forecasting performance. We would like to note that the performance gaps in the main table appear relatively small because the results are averaged across multiple prediction horizons, whereas the individual comparisons show that CoLP performs better in most cases. We believe these results provide further evidence of CoLP's strong and stable forecasting performance and help address the concern regarding our performance claims.

CoLP results.

DatasetWeatherSolarElectricityTraffic
960.141±0.0000.164±0.0010.123±0.0010.370±0.001
1920.183±0.0020.177±0.0010.143±0.0020.385±0.001
3360.232±0.0010.188±0.0020.158±0.0030.397±0.003
7200.303±0.0010.200±0.0010.193±0.0000.435±0.002
DatasetETTh1ETTh2ETTm1ETTm2
960.355±0.0020.274±0.0030.276±0.0030.156±0.001
1920.393±0.0010.336±0.0040.315±0.0020.210±0.002
3360.407±0.0030.344±0.0010.345±0.0020.264±0.004
7200.420±0.0030.394±0.0050.391±0.0040.355±0.001

TimePerceiver results.

DatasetWeatherSolarElectricityTraffic
960.146±0.0030.160±0.0030.128±0.0020.371±0.003
1920.185±0.0020.174±0.0020.148±0.0020.386±0.002
3360.237±0.0030.186±0.0020.162±0.0010.399±0.004
7200.315±0.0040.205±0.0040.199±0.0030.434±0.001
DatasetETTh1ETTh2ETTm1ETTm2
960.352±0.0020.278±0.0040.286±0.0080.163±0.004
1920.393±0.0080.332±0.0050.322±0.0050.218±0.007
3360.418±0.0040.345±0.0020.353±0.0020.273±0.003
7200.446±0.0090.404±0.0090.402±0.0040.360±0.001
AuthorsOpenReview ↗

Response to Reviewer 6NLm [2/3]

[W3] The main component ablation covers only four of the eight datasets, leaving the consistency of each component's contribution across all benchmarks less clear.

[Response to W3] Thank you for this valuable feedback. To provide a more comprehensive evaluation, we extend the component ablation analysis to all eight datasets. As shown below, latent propagation consistently improves forecasting performance, demonstrating its effectiveness as the core component of CoLP. The scale-aware token further strengthens these improvements on most datasets, with a clear example illustrated in Figure 4. These results highlight the benefit of accumulating historical context through latent propagation, while the scale-aware token provides additional gains by preserving scale information across instances. For Traffic, where scale variation across instances is limited, capturing fine-grained temporal patterns appears more important than preserving scale information. In this case, latent propagation alone is preferable. We empirically find that one can reliably decide whether to include the scale-aware token based on validation performance, as we did for all evaluated configurations (Appendix A.1).

Componentsw/o scale, w/o propw/o scale token, w/ propw/ scale token, w/ prop
Weather0.2420.2290.229
Solar0.2650.2260.197
ECL0.1760.1680.164
Traffic0.4710.4310.519
ETTh10.4230.4200.398
ETTh20.3760.3730.364
ETTm10.3710.3650.359
ETTm20.2770.2730.270

[W4] Training propagates context over a short segment of consecutive instances, while inference can carry the latent state over a much longer stream. Further analysis of long-run state drift and reset strategies would strengthen the continual-context claim.

[Response to W4] Thank you for this thoughtful suggestion. CoLP can operate beyond the training segment length because it learns how to update accumulated historical context, rather than relying only on a fixed number of preceding instances. As shown in Figure 5(a), performance is relatively poor at G=2G=2, which we interpret as the model primarily learning to incorporate information from the immediately preceding instance. In contrast, performance improves and begins to saturate near its best level at G=4G=4, suggesting that the model has learned to effectively update and retain historical context accumulated across multiple instances. We therefore select G=4G=4 as a sufficient training length for learning context accumulation, rather than as a limit on the historical context available during inference.

To further support this interpretation, we conduct additional experiments on the final 30% of the test stream with an extended range of inference propagation lengths Tprop∈1,2,4,8,32,128,∞T_{prop} \in \\{1, 2, 4, 8, 32, 128, \infty\\}. As shown below, performance improves as TpropT_{prop} increases and then stabilizes at longer propagation lengths. The no-reset setting (Tprop=∞T_{prop}=\infty) also maintains strong performance, supporting the effectiveness of CoLP during long-term inference without requiring periodic resets. We have incorporated this analysis (Section 5.5) and the additional experimental results (Appendix A.5) in the revised manuscript.

TpropT_{prop}ETTm1WeatherSolar
10.4610.3130.319
20.4190.3060.267
40.4110.2980.210
80.4080.2960.208
320.4080.2970.207
1280.4080.2960.208
∞0.4070.2960.208
AuthorsOpenReview ↗

Response to Reviewer 6NLm [3/3]

[W5] The authors should additionally report CoLP’s results with an input length of 336 across all datasets and prediction horizons. This commonly used setting would make its performance easier to compare with prior work and help assess whether the reported gains depend on the selected input length.

[Response to W5] Thank you for this valuable suggestion. Following your recommendation, we evaluate CoLP with an input length of L=336 across all eight datasets and four prediction horizons. We compare CoLP with widely used baselines under this setting, including CycleNet [1], PatchTST [2], and DLinear [3], as well as the recent baseline GTR [4]. As shown below, CoLP achieves the best performance across all eight datasets, further demonstrating its effectiveness compared with both established and recent baselines. We include these results in the revised manuscript (Appendix A.6).

ModelsOursGTRCycleNetPatchTSTDLinear
Weather0.2210.2270.2420.2300.245
Solar0.1840.1960.2260.2080.253
Electricity0.1580.1590.1580.1620.166
Traffic0.3980.4050.4130.4010.434
ETTh10.3940.4150.4150.4240.436
ETTh20.3400.3450.3550.3420.421
ETTm10.3330.3450.3550.3540.364
ETTm20.2470.2540.2510.2570.282

References

[1] Lin et al., CycleNet: Enhancing Time Series Forecasting through Modeling Periodic Patterns, NeurIPS 2024
[2] Nie et al., A Time Series Is Worth 64 Words: Long-Term Forecasting with Transformers, ICLR 2023
[3] Zeng et al., Are Transformers Effective for Time Series Forecasting?, AAAI 2023
[4] Cao et al., Enhancing Multivariate Time Series Forecasting with Global Temporal Retrieval, ICLR 2026


We hope that our additional experiments and explanations have addressed your concerns and provided further evidence supporting our framework. If you have any further questions or would like to discuss any remaining points, we would be glad to continue the discussion.

Thank you for your valuable time.

Best regards,
Authors

Comments on the paper

3 comments
AuthorsOpenReview ↗

General Response

Dear Reviewers and AC,

We sincerely appreciate the reviewers’ time and effort, as well as their insightful feedback and constructive suggestions.

We are also grateful for their thoughtful assessments of the key strengths of our work, including but not limited to:

  • Meaningful and practically significant problem setting that addresses the limitations of fixed-length forecasting (Reviewers 6NLm, p2MQ, XquA, Y54x).

  • Well-motivated latent propagation design with horizon-aligned representations (Reviewers 6NLm, p2MQ, XquA, Y54x).

  • Competitive forecasting performance across eight benchmarks (Reviewers p2MQ), supported by comprehensive experiments or ablation studies (Reviewers p2MQ, XquA, Y54x).

To further strengthen our work, we have conducted additional experiments and revised the manuscript accordingly. The key revisions are summarized below:

  • Further Analysis of Latent Propagation and Model Design

    • Clarify the contributions of latent propagation and scale-aware tokens (Section 5.3).
    • Clarify the conventional latent-state control for evaluating horizon-aligned latent design (Section 5.4).
    • Expand the analysis of training segment length and context accumulation (Section 5.5, Appendix A.5).
  • Expand Comparisons with Forecasting Baselines

    • Expand baseline comparisons to include retrieval-based forecasting methods (Section 5.2).
    • Conduct additional comparisons under equal historical context lengths (Appendix A.6).
    • Conduct additional comparisons with a commonly used input length of 336 (Appendix A.6).
  • Improve Organization and Experimental Clarity

    • Reorganize the contributions into bullet points (Section 1).
    • Move and reorganize the Related Work section to better contextualize our approach (Section 2).
    • Clarify the experimental settings and baseline comparisons (Section 5.1, Appendix A.1).

All revisions are highlighted in magenta\textcolor{magenta}{magenta} in the revised manuscript for ease of reference.

We hope that these revisions further strengthen our work and address the reviewers’ concerns. We would be happy to answer remaining questions or discuss further suggestions.

Thank you for your time and effort.

Best regards,
Authors

AuthorsOpenReview ↗

A Gentle Reminder to AC and Reviewers

Dear AC and Reviewers,

We hope this message finds you well. We sincerely appreciate your thoughtful feedback and would like to kindly follow up on our rebuttal.

We have carefully addressed the reviewers’ concerns through detailed clarifications, further analysis, and additional experimental results. We have revised the manuscript accordingly and summarized the key revisions in our General Response. We hope these efforts have helped strengthen our paper and clarify important aspects of our work.

If you have any remaining concerns or would like to discuss any aspect of our work further, we would be happy to address them.

Thank you for your time and thoughtful consideration you have devoted to our work.

Best regards,
Authors

AuthorsOpenReview ↗

A Gentle Reminder on the Final Day of Discussion

Dear AC and Reviewers,

As we approach the final day of the discussion period, we would like to sincerely thank you for your time and thoughtful feedback.

We have carefully addressed all reviewers' concerns through clarifications, additional experiments, and manuscript revisions. We hope our responses have helped resolve the remaining concerns and would greatly appreciate the opportunity for further discussion before the discussion period ends.

Thank you again for your time and careful consideration.

Best regards,
Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.