acceptodds
Under review as a conference paper at ICLR 2027

Bridging Horizon-Specific Representations for Time Series Forecasting

Abstract

Existing time series forecasting methods typically train separate models for each forecasting horizon, as different horizons demand distinct representations. However, such a horizon-specific method limits its flexibility in real-world scenarios with diverse forecasting demands. In this paper, we provide a theoretical analysis, which reveals that optimal representation can vary with the horizon: short-term forecasting relies on local dynamic patterns, while long-term forecasting depends more on global trends. We thereby propose HorizonBridge, a novel framework that leverages a representation translator to bridge horizon-specific representations for cross-horizon forecasting. Specifically, HorizonBridge shares a base representation encoder across different horizons to preserve common structure. Then, we devise a horizon-conditioned representation translator to generate adaptive residual adjustments for the base representation, thereby adapting it to each target horizon. To further improve cross-horizon representation compatibility, we incorporate an adaptive patch embedding mechanism that partitions time series according to local information density and leverages temporal periodicity to construct cyclical embeddings. Extensive experiments show that HorizonBridge achieves competitive forecasting performance across multiple benchmark datasets, while offering improved efficiency and flexibility. The source code and models will be released.

Primary area:
learning on time series and dynamical systems
Keywords:
time series forecasting, cross-horizon prediction, representation learning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_iPTVOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how to reuse time-series representations trained at one forecast horizon to support multiple horizons. The authors first analyze the horizon dependence of optimal representation subspaces under a low-rank linear representation model. They show that reusing a source representation incurs positive excess risk when the optimal subspaces differ and the target spectrum has an eigengap. HorizonBridge trains a base encoder at H=96, freezes the embedding module and encoder, and learns residual adjustments through a dynamic mixture of basis matrices conditioned on the horizon and context, together with a shared prefix forecasting head. The model also uses MDL-based adaptive segmentation and periodic embeddings. The experiments cover eight standard forecasting datasets, four training horizons, three additional backbones equipped with the translator, and unseen intermediate horizons. Table 1 reports competitive average accuracy. The H=128/256/480 experiments in Appendix E evaluate horizon interpolation within the training range, rather than extrapolation beyond the maximum training horizon.

Critical strengths
  1. The deployment problem is practically relevant, and the accuracy results have reasonable breadth. Table 1 and Appendix Table 6 show strong results on datasets such as Weather, Electricity, and Traffic while reusing the encoder trained at H=96. For example, the average Electricity MSE in Table 1 is 0.153, compared with 0.161 for PhaseFormer. Coverage across datasets and horizons gives the findings broader practical relevance than adaptation to a single forecasting task.
  2. Several evaluations address the central design choices. Figure 4 separately examines removing the translator, conditioning, and adaptive patching. Table 2 compares prefix and autoregressive strategies, and Table 3 tests the translator with PatchTST, Crossformer, and TimeXer. The appendix also reports averages over three seeds and results at unseen intermediate horizons. The theoretical lower bound explicitly requires different optimal subspaces and an eigengap.
Critical weaknesses
  1. The theory does not establish the proposed mechanism. Pages 4–5 model the source representation as a reduced-dimensional z=Rμ. The translator receives only this frozen representation, its pooled context, and H, so it cannot recover predictive information already discarded by z. In the centered jointly Gaussian case, E[Y_H|z] is linear in z. Equation (4) already optimizes over all linear readouts; a nonlinear translator cannot improve on the Bayes risk attained by that conditional mean. The theorem therefore motivates a problem without showing that the architecture resolves it in this setting. Deep representations may retain information that the translator makes easier to read out, but this needs empirical support. The unconditional-MLP ablation changes both conditioning and dynamic parameter generation, leaving horizon adaptation, nonlinear readout capacity, and joint multi-horizon training insufficiently separated.
  2. Adaptive segmentation is not isolated from the embedding module. The w/o Adaptive Patch ablation (Figure 4, p. 8) replaces an embedding module combining MDL segmentation and periodic embedding, without clarifying whether periodic embedding is retained. It supports the module as a whole, but does not isolate MDL boundary selection or its contribution to cross-horizon compatibility. The period-length sensitivity analysis (Figure 5, p. 9) does not resolve this. Compare MDL with uniformly spaced boundaries while holding patch count, periodic embedding, backbone, head, and training loss fixed, and report results by horizon. Equation (10) specifies only a minimum segment length, whereas Algorithm 1 fixes N and the Flatten head requires consistent token counts. The implemented optimization problem and values of N and q need clarification.
  3. Statistical reporting and cost definitions affect the claimed advantages. Appendix F (p. 25) claims 53 wins across 64 settings, but Table 6 (p. 27) includes Avg rows in that count. Following its red first-place markings, excluding Avg rows yields 22 MSE and 19 MAE wins, or 41/64; 12 Avg-row wins produce 53. This follows the marked rankings, without inferring ties from rounded values. Table 7 should clarify whether Wilcoxon tests exclude Avg rows and account for dependence among horizons within datasets. Time in Tables 3 and 9 sums per-iteration times across stages or independent models, rather than total training wall-clock time, and offline MDL segmentation adds cost. These measurements cannot establish a 58.92% reduction in total training cost. Both issues affect the claimed breadth and practical value of the gains.
Questions for authors
  1. Mechanism clarification and controls. In the low-rank model, how can the translator recover predictive directions discarded by the source representation? Should the theory motivate the problem while the mechanism is framed as improving readout of retained information? Please compare against a parameter-matched nonlinear head and simple horizon-conditioned modulation under the same multi-horizon loss.
  2. Cost clarification. Please report actual steps and total training times separately for stages one and two and the four independent horizon-specific models, including segmentation preprocessing. Does the reported cost reduction hold with matched validation budgets and supervised losses?
3 comments
AuthorsOpenReview ↗

Response to Reviewer iPTV

Thank you for your careful evaluation and constructive feedback on our paper. We address your concerns below and have updated the manuscript accordingly, with revisions highlighted in red.

W1 & Q1: Mechanism clarification and controls.

1. Gap between the idealized analysis and the actual encoder: We agree that predictive information discarded by the encoder matrix RR cannot be recovere by any subsequent transformation in the idealized linear compression model. However, our theoretical analysis aims to reveal cross-horizon representation mismatch within a mathematically tractable framework, thereby motivating representation adaptation. It is not intended to directly characterize the encoding process in practical forecasting models. Specifically, the posterior mean of the latent state μT\mu_T depends on the underlying dynamical system and its parameters, whereas real-world time-series dynamics are often complex and unknown. Consequently, practical encoders learn high-dimensional representations directly from historical observations rather than applying a fixed low-rank compression to a known posterior latent state. These representations do not necessarily follow the compression assumption in our theoretical model and may retain predictive information relevant to different forecasting horizons, which the Translator can better exploit through horizon-conditioned adaptation. Our newly added prediction-relevant subspace analysis in Section 5.4 (p. 9) further supports this interpretation, showing that the Translator improves representation similarity to target-horizon references.

2. Additional Mechanism Controls: We respectfully note that the w/o Condition variant in the original Fig. 4 (p. 8) already uses a parameter-matched MLP in place of the Translator, serving as a nonlinear prediction head baseline. According to your suggestion, we have further added a simple horizon-conditioned modulation baseline based on Feature-wise Linear Modulation (FiLM). As shown in the table below, the FiLM Translator variant achieves some improvement over w/o Condition but still underperforms HorizonBridge (Full results across different horizons are provided in Table 12 on page 32).

DatasetMetricHorizonBridgew/o ConditionFiLM Translator
ETTh1MSE0.3900.4090.394
MAE0.4170.4320.418
WeatherMSE0.2150.2190.217
MAE0.2460.2520.249
SolarMSE0.1800.1880.185
MAE0.2140.2190.218

W2: Adaptive Patching vs. Uniform Patching.

1. Ablation of Adaptive Patch Embedding: We would like to clarify that MDL-based adaptive patching and periodic embedding are structurally coupled in our framework, rather than being two components that can be independently removed. Specifically, MDL-based patching produces variable-length patches, which conventional fixed-length patch embedding cannot directly handle without additional mechanisms. Conversely, replacing MDL-based patching with uniform patching allows standard patch embedding to be applied directly, eliminating the need for a separate periodic embedding mechanism. Accordingly, the w/o Adaptive Patching variant in the original Fig. 4 (p. 8) already provides this uniform patching baseline, while keeping all other settings (e.g., the number of patches) consistent with HorizonBridge. The results show consistent performance improvements with our proposed method.

2. MDL Implementation Details: In the original Eq. 1010, the summation over KK patches already reflects the fixed patch-count setting, which also ensures compatibility with the flatten head. To avoid confusion, we have replaced KK with NN in Eq. 1010 for consistency with Algorithm 1 in the Appendix. The dataset-specific values of NN and qq are provided in the table below.

DatasetETTh1ETTh2ETTm1ETTm2WeatherSolarElectricityTraffic
qq33666336
NN3030151215151030
AuthorsOpenReview ↗

Response to Reviewer iPTV

W3 & Q2: Statistical Analysis and Training Efficiency.

1. Correction of Statistical Errors: Thank you for pointing out this statistical error. The original win count mistakenly included the Avg rows, and the correct count is 41/64 wins (22 for MSE and 19 for MAE). Following your comment, we carefully rechecked the statistical results throughout the manuscript and also identified an error in the reported reduction in cumulative per-iteration training time, which should be 65.04% rather than the originally reported 58.92%. We have made the corresponding corrections in the revised manuscript.

2. Wilcoxon statistical analysis: We agree that results across different forecasting horizons within the same dataset are not independent. Therefore, we have adopted a dataset-level Wilcoxon signed-rank test, using the average performance difference across the four horizons for each of the eight datasets. We have updated Table 8 accordingly on page 30, with the Avg rows excluded from the test. As shown in the table below, HorizonBridge significantly outperforms both PhaseFormer and VPNet in terms of MSE and MAE (p<0.05p < 0.05), demonstrating the robustness of its gains.

ComparisonMetricWilcoxon Statisticp-valueSignificant
HorizonBridge vs PhaseFormerMSE07.813 × 10⁻³Yes
HorizonBridge vs VPNetMSE07.813 × 10⁻³Yes
HorizonBridge vs PhaseFormerMAE33.906 × 10⁻²Yes
HorizonBridge vs VPNetMAE11.563 × 10⁻²Yes

3. Training Cost Analysis: Following your suggestion, we have added Table 11 on page 31 to report end-to-end training times, training epochs, and training iterations for the plug-and-play experiments. As shown in the table below, the Translator reduces the average end-to-end training time by 60.91%, demonstrating HorizonBridge's efficiency advantage in practice. Notably, training epochs vary across models due to early stopping, while iterations per epoch also differ slightly because different forecasting horizons yield different numbers of training samples. Our original per-iteration runtime metric provides a more direct comparison of computational efficiency by excluding differences in training epochs and overhead from validation and checkpoint I/O. Together with the newly reported end-to-end training times, these measurements provide a more comprehensive evaluation of practical training efficiency.

DatasetModelOriginal Time (s)Original EpochsOriginal Iterations+Translator Time (s)+Translator Epochs+Translator Iterations
SolarPatchTST26999.156624669455.42325672
Crossformer13314.045502792248.51617840
TimeXer7137.248536603685.62224570
WeatherPatchTST6014.355615862540.92528035
Crossformer4312.241459411254.41617888
TimeXer1652.72932470848.21516760
TrafficPatchTST32981.7298127813728.71541742
Crossformer21293.1491378797693.62261514
TimeXer47936.04312040122954.42055936
AuthorsOpenReview ↗

Follow-up to Reviewer iPTV

Dear Reviewer iPTV,

We appreciate your careful review and the insightful suggestions for our work. As the discussion period nears its conclusion, we would like to kindly follow up on our responses and revised manuscript.

To better substantiate the proposed mechanism, we have introduced a prediction-relevant subspace analysis and additional comparisons with a horizon-conditioned FiLM baseline. We have also clarified the scope of our theoretical analysis and the implementation details of adaptive patching. Furthermore, we have corrected the statistical reporting, updated the Wilcoxon significance tests, and supplemented the efficiency evaluation with end-to-end training cost comparisons.

We would greatly appreciate your feedback on whether these additional experiments and clarifications address your main concerns. If any questions remain, we would be happy to provide further clarification before the discussion ends.

Thank you once again for your valuable time and consideration.

Best regards,
Authors of Submission 30237

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

This paper proposes HorizonBridge, a framework for cross-horizon time series forecasting without retraining the entire model. The authors provide a theoretical analysis of representation mismatch across forecasting horizons and introduce a horizon-conditioned translator to adapt the representations learned by a frozen Transformer encoder. They also propose an adaptive patch embedding mechanism combining MDL-based patching and periodic embeddings. Experiments on eight benchmark datasets demonstrate competitive forecasting performance with reduced training costs and model parameters. Additional ablation studies and plug-and-play experiments are conducted to validate the proposed framework.

Critical strengths
  1. The paper addresses an important problem in time series forecasting, where different forecasting horizons usually require training separate models. Supporting multiple horizons without retraining the entire model can be useful for real-world applications and reduce computational costs.
  2. The proposed horizon-conditioned representation translator is an interesting approach to adapting learned representations across different forecasting horizons. The two-stage training strategy is also reasonable, and the experimental results show competitive performance with reduced training costs.
Critical weaknesses
  1. The proposed translator is trained and evaluated on predefined forecasting horizons (96, 192, 336, and 720). However, the paper does not investigate whether the model can generalize to unseen horizons. Since cross-horizon forecasting is the main motivation of this work, I believe this is a significant limitation that weakens the main contribution of the paper.

  2. The base encoder is trained only on H=96 and then frozen for other forecasting horizons. However, the authors do not provide sufficient justification for selecting this horizon or any sensitivity analysis. It is unclear whether using other source horizons, such as H=336 or H=720, would improve or degrade the performance. Since the entire framework relies on the frozen base encoder, I believe this design choice needs stronger experimental validation.

Questions for authors
  1. What is the justification for selecting H=96 as the source horizon? Have the authors investigated whether using H=336 or H=720 as the source horizon would improve or degrade the performance across different target horizons?

  2. Can HorizonBridge generalize to forecasting horizons that were not included during training, such as H=300 or H=600? Since cross-horizon forecasting is the main motivation of this work, I believe this experiment is important to support the main contribution.

  3. The paper argues that different forecasting horizons require different representations. Can the authors provide quantitative evidence of this mismatch, such as representation similarity or subspace analysis? Also, how can we verify that the proposed translator actually reduces this mismatch?

  4. Proposition 1 is derived based on a linear dynamical system, while the proposed framework uses a nonlinear Transformer encoder and translator. Can the authors explain how the theoretical findings support the proposed architecture and whether the same representation mismatch can be observed in the actual model?

  5. Why is another unified multi-horizon forecasting model not included in the baseline comparisons? Since these methods also support different forecasting horizons, I believe such comparisons are important to evaluate the effectiveness of HorizonBridge.

  6. The proposed framework combines MDL-based adaptive patching, periodic embedding, and representation translation. Can the authors provide separate ablation results for MDL-based patching and periodic embedding to clarify how much each component contributes to the final performance?

  7. Does the reported training time include the initial training of the base encoder and the computation required for adaptive patching? Since computational efficiency is one of the main advantages claimed in the paper, I believe a complete comparison of training costs is necessary.

3 comments
AuthorsOpenReview ↗

Response to Reviewer aDXV

Thank you for your careful evaluation and constructive feedback on our paper. We address your concerns below and have updated the manuscript accordingly, with revisions highlighted in red.


Q1: Why Choose H=96 as the Source Horizon?

We selected the shortest forecasting horizon, Hs=96H_{s}=96, as the source horizon for two main reasons:

  • First, shorter-horizon forecasting generally involves lower prediction difficulty, reducing the uncertainty associated with long-horizon forecasting and facilitating the learning of stable temporal representations by the base encoder.
  • Second, uniformly using this horizon avoids dataset-specific tuning and contributes to a unified experimental setting.

To empirically validate this design choice, we conducted a sensitivity analysis evaluating HorizonBridge with four source horizons Hs∈{96,192,336,720}H_s \in \lbrace 96, 192, 336, 720 \rbrace, which has been added to Fig. 6 (p. 25). As shown in the table below, the overall MSE and MAE vary by only 2.11% and 1.96%, respectively, with Hs=96H_s=96 achieving the best aggregated performance.

DatasetMetricHs=96H_s = 96 (Default)Hs=192H_s = 192Hs=336H_s = 336Hs=720H_s = 720
ETTm2MSE0.25000.24880.25740.2464
MAE0.30480.30420.30810.3049
WeatherMSE0.21460.21930.22010.2211
MAE0.24640.25230.25330.2542
SolarMSE0.18040.18260.18190.1794
MAE0.21370.21630.21800.2171
AverageMSE0.21500.21690.21980.2156
MAE0.25500.25760.25980.2587

Q2: Generalization to Unseen Forecasting Horizons.

We respectfully refer you to Fig. 7 in Appendix E, which already reports the generalization results on unseen forecasting horizons. In this experiment, HorizonBridge is evaluated on H=128,256 and 480, none of which were used during training. These results demonstrate that HorizonBridge can generalize to unseen forecasting horizons without retraining.


Q3: Quantitative Analysis of Representation Mismatch.

Thank you for your valuable suggestion. Based on your advice, we added a prediction-relevant subspace analysis in Section 5.4 on page 9, and further visualized it via UMAP in Figs. 9–10, Appendix F. Specifically, we construct target-horizon reference representations from the ground-truth sequences via the pseudoinverse of the prediction head. Then, we project original and translated representations onto the row space of the prediction head to exclude prediction-irrelevant components, and compute their cosine similarity to the reference representation (see Appendix C.3 for details). As shown in the table below, translation consistently improves similarity across all datasets, with larger gains at longer horizons. These results suggest that our proposed method helps alleviate cross-horizon representation mismatch in the actual forecasting model.

DatasetMetricH = 192H = 336H = 720Avg.
ETTh2Original0.3670.2860.1930.282
+Translator0.4020.3330.2530.329
IMP9.54%16.13%30.72%18.80%
ETTm2Original0.3390.2530.1220.238
+Translator0.4520.3570.2050.338
IMP33.19%40.82%68.22%47.41%
SolarOriginal0.5980.4640.2590.441
+Translator0.6570.5250.3150.499
IMP9.80%13.09%21.46%14.78%
WeatherOriginal0.4010.2950.1520.283
+Translator0.5220.4000.2390.387
IMP30.19%35.61%56.53%40.77%
AuthorsOpenReview ↗

Response to Reviewer aDXV

Q4: Applicability of Theoretical Analysis to Nonlinear Models.

We would like to clarify that the linear dynamical system is introduced to model the underlying dynamics of time series, rather than to impose linearity constraints on the forecasting encoder. Although this assumption simplifies the complex dynamics of real-world time series, similar idealized assumptions have been widely adopted in theoretical analyses of time-series forecasting [1,2,3]. Our theoretical analysis is intended to provide a theoretical motivation for cross-horizon representation adaptation, rather than to fully characterize the complex dynamics of real-world time series. Furthermore, the subspace analysis in Section 5.4 and the improvements in forecasting performance provide empirical support for cross-horizon representation mismatch and the effectiveness of representation translation in nonlinear forecasting models.

Reference

[1] Taga, E. O., et al. "TimePFN: Effective Multivariate Time Series Forecasting with Synthetic Data." AAAI, 2025.

[2] Hu, Y., et al. "Bridging Past and Future: Distribution-Aware Alignment for Time Series Forecasting." ICLR, 2026.

[3] Benechehab, A., et al. "AdaPTS: Adapting Univariate Foundation Models to Probabilistic Multivariate Time Series Forecasting." ICML, 2025.


Q5: Comparison with Unified Multi-Horizon Forecasting Models.

We agree that including unified multi-horizon forecasting models would provide a valuable comparison. According to your comment, we have included a representative mask-based multi-horizon forecasting model ElasTST as an additional baseline. As shown in the updated Table 1 on page 7, HorizonBridge consistently outperforms ElasTST, achieving average MSE and MAE reductions of 9.40% and 7.59%, respectively. Furthermore, we have extended the multi-horizon strategy analysis in Table 2 on page 8 to include the mask-based strategy, alongside the existing Prefix and AR strategies.


Q6: Separate Ablations of MDL-based Patching and Periodic Embedding.

We would like to clarify that MDL-based adaptive patching and periodic embedding are structurally coupled in our framework, rather than being two components that can be independently removed. Specifically, MDL-based patching produces variable-length patches, which conventional fixed-length patch embedding cannot directly handle without additional mechanisms. Conversely, replacing MDL-based patching with uniform patching allows standard patch embedding to be applied directly, eliminating the need for a separate periodic embedding mechanism. Accordingly, the w/o Adaptive Patching variant in the original Fig. 4 (p. 8) already provides this uniform patching baseline, while keeping all other settings consistent with HorizonBridge. The results demonstrate the effectiveness of our proposed embedding approach.


Q7: Does the Reported Training Time Include All Training Stages and Adaptive Patching?

1. Training Efficiency Evaluation: Fig. 8 in Appendix E separately reports the per-iteration training time of the base encoder training (Stage 1) and the subsequent training of the Translator (Stage 2). We adopt per-iteration runtime to enable a more direct comparison of computational efficiency, as the official baseline implementations differ in early-stopping criteria, validation frequencies, and checkpointing procedures, all of which affect end-to-end training time. This evaluation method has also been widely adopted in prior time series forecasting studies [4,5,6]. To provide a more comprehensive comparison under a unified training pipeline, we have added the end-to-end training costs of plug-and-play experiments to Table 11 on page 31.

2. Adaptive Patching Overhead: MDL-based patching is performed once as an offline preprocessing step using dynamic programming. The resulting patch partitions are reused throughout training without repeated computation. Therefore, the preprocessing cost is not included in the reported per-iteration runtime. To quantify this overhead, we additionally measured the MDL preprocessing time across all eight datasets, as shown in the table below. Results are reported as mean ± standard deviation over 10 runs.

DatasetPreprocessing (s)
ETTh18.25 ± 0.31
ETTh27.80 ± 0.46
ETTm110.44 ± 0.29
ETTm27.42 ± 0.27
Weather27.37 ± 0.55
Solar148.21 ± 2.99
Electricity129.73 ± 5.45
Traffic266.25 ± 7.96

Reference

[4] Hu, Y., et al. "Bridging Past and Future: Distribution-Aware Alignment for Time Series Forecasting." ICLR, 2026.

[5] Peng, Z., et al. "ProtoTS: Learning Hierarchical Prototypes for Explainable Time Series Forecasting." ICLR, 2026.

[6] Yue, W., et al. "OLinear: A Linear Model for Time Series Forecasting in Orthogonally Transformed Domain." NeurIPS, 2025.

AuthorsOpenReview ↗

Follow-up to Reviewer aDXV

Dear Reviewer aDXV,

Thank you again for your careful review and constructive feedback. As the author–reviewer discussion period draws to a close, we would like to kindly follow up on our responses and revised manuscript.

Following your suggestions, we have conducted a source-horizon sensitivity analysis and a quantitative analysis of prediction-relevant subspaces, providing further empirical evidence of cross-horizon representation mismatch and the effectiveness of the Translator. In addition, we have included comparisons with ElasTST, a unified multi-horizon forecasting model, and provided a more comprehensive efficiency evaluation covering end-to-end training costs and adaptive patching preprocessing overhead.

We would greatly appreciate your feedback on whether these additional experiments and clarifications address your main concerns. If any questions remain, we would be happy to provide further clarification before the discussion ends.

Thank you again for your time and thoughtful consideration.

Best regards,
Authors of Submission 30237

Reviewer_aPGzOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes HorizonBridge for time-series forecasting across multiple prediction horizons. The method first trains embedding and encoder modules on a base prediction horizon and freezes their parameters. Conditioned on the target horizon and input context, a representation transformer generates residual adjustments, while a shared prefix prediction head produces the forecasts. The paper further introduces minimum-description-length-based adaptive patching and periodic embeddings.

The authors conduct experiments on eight datasets and four prediction horizons, including component ablations, evaluations on unseen horizons, and adaptation experiments across multiple backbone architectures. The results indicate that the proposed method is competitive and can reduce the redundant training overhead associated with multi-horizon settings.

Critical strengths

(1) The study is practically relevant. Different applications often require forecasts at multiple horizons, and the need to train and maintain separate models increases costs. By sharing a base encoder, the proposed approach reduces redundant training while maintaining forecasting accuracy, which has practical value.

(2) The technical rationale is clear. The method separates base representation extraction from target-horizon adaptation, generates dynamic residual adjustments conditioned on the horizon and context, and uses a shared prediction head to support multi-horizon outputs. This design is well motivated.

Critical weaknesses

(1) The independent contributions of some design choices are not sufficiently disentangled. The ablation in Figure 4 changes both MDL-based patching and periodic embeddings simultaneously, making it difficult to quantify the performance gain attributable to each component. Equation (33) in the appendix introduces a hybrid time-domain/frequency-domain loss, but its contribution is not analyzed.

(2) The efficiency analysis and result reporting require greater rigor. Tables 3 and 9 mainly aggregate the per-iteration runtime of each training stage, which cannot be directly equated with end-to-end training time. In the main experiments and plug-and-play adaptation experiments, the same baseline reports different values under identical dataset and prediction-horizon settings. The differences in experimental configurations and training protocols should be clarified.

Questions for authors

(1) In Figure 4, the ablation replaces the adaptive patching embedding as a whole. Could the authors disentangle the respective contributions of MDL-based patching and periodic embeddings, and provide an ablation of the hybrid loss in Equation (33)? These results could materially affect the assessment of the claimed component-wise benefits.

(2) Which specific modules correspond to the channel-independent setting mentioned in Section 4? Why is spatiotemporal attention used in Appendix C.1? The boundary between these two designs should be clarified to support reproducibility.

(3) When computing the periodic phase, does the index t denote the relative position within the input window or the global timestamp in the original series? In addition, Appendix C.1 introduces both learnable and sinusoidal horizon embeddings. Which embedding is used in the main experiments, and which is used in the unseen-horizon experiments?

3 comments
AuthorsOpenReview ↗

Response to Reviewer aPGz

Thank you for your careful evaluation and constructive feedback on our paper. We address your concerns below and have updated the manuscript accordingly, with revisions highlighted in red.


W1 & Q1: Independent Contributions of Model Components.

1. Adaptive patching vs. periodic embedding: We would like to clarify that MDL-based adaptive patching and periodic embedding are structurally coupled in our framework, rather than being two components that can be independently removed. Specifically, MDL-based patching produces variable-length patches, which conventional fixed-length patch embedding cannot directly handle without additional mechanisms. Conversely, replacing MDL-based patching with uniform patching allows standard patch embedding to be applied directly, eliminating the need for a separate periodic embedding mechanism. Accordingly, the w/o Adaptive Patching variant in the original Fig. 4 already provides this uniform patching baseline, while keeping all other settings consistent with HorizonBridge. The results show consistent performance improvements with our proposed method.

2. Hybrid time–frequency loss: According to your comment, we have added an ablation study on the hybrid time–frequency loss in Eq. 33 to Table 12 on page 32. As shown in the table below, hybrid loss improves HorizonBridge’s performance.

DatasetMetricHorizonBridgew/o Hybrid Loss
ETTh1MSE0.3900.396
MAE0.4170.422
WeatherMSE0.2150.219
MAE0.2460.253
SolarMSE0.1800.186
MAE0.2140.227

Q2: Clarification of Channel Interaction.

The channel-independent formulation is adopted to simplify the mathematical notation in Section 4. Our core designs, including adaptive patching, periodic embedding and the representation translator, do not involve cross-channel interactions, so the channel dimension is omitted. Cross-channel interactions are introduced only in the base encoder through the channel attention, as described in Appendix C.1 (p. 18). We have clarified this in Section 4 to avoid ambiguity.


Q3: Clarification of Embedding Configurations.

1. Periodic Embedding: The index tt used to compute the periodic phase denotes the relative position within the input window, rather than the global timestamp in the original time series. This avoids introducing unnecessary dependence on the starting position of the input window.

2. Horizon Embedding: Both the main experiments and the unseen-horizon experiments use sinusoidal horizon embeddings, enabling the model to generalize to forecasting horizons not encountered during training.

To avoid confusion, we have clarified these details in Section 4.1 (p. 6) and Appendix C.1 (p. 18).

AuthorsOpenReview ↗

Response to Reviewer aPGz

W2: Experimental Consistency and Training Efficiency.

1. Training Efficiency Evaluation: We use per-iteration runtime to exclude differences in training epochs caused by early stopping, as well as overhead from validation, checkpoint I/O, and data loading. This allows a more direct comparison of computational efficiency across models. This evaluation method has also been widely adopted in prior time series forecasting studies [1,2,3]. Nevertheless, we agree that end-to-end training time is also important for assessing practical training costs. Following your suggestion, we have added Table 11 on page 31, which reports the end-to-end training cost of the plug-and-play experiments. As shown in the table below, the Translator reduces end-to-end training time by an average of 60.91% across the nine experimental settings, further confirming its efficiency advantage.

DatasetModelOriginal Time (s)Original EpochsOriginal Iterations+Translator Time (s)+Translator Epochs+Translator Iterations
SolarPatchTST26999.156624669455.42325672
Crossformer13314.045502792248.51617840
TimeXer7137.248536603685.62224570
WeatherPatchTST6014.355615862540.92528035
Crossformer4312.241459411254.41617888
TimeXer1652.72932470848.21516760
TrafficPatchTST32981.7298127813728.71541742
Crossformer21293.1491378797693.62261514
TimeXer47936.04312040122954.42055936

2. Experimental Consistency: The differences in baseline performance between the main and plug-and-play experiments arise from their training configurations: the main experiments follow the official baseline settings, whereas the plug-and-play experiments adopt unified training configurations to ensure fair comparisons, as detailed in Appendix C.2.

Reference

[1] Hu, Y., et al. "Bridging Past and Future: Distribution-Aware Alignment for Time Series Forecasting." ICLR, 2026.

[2] Peng, Z., et al. "ProtoTS: Learning Hierarchical Prototypes for Explainable Time Series Forecasting." ICLR, 2026.

[3] Yue, W., et al. "OLinear: A Linear Model for Time Series Forecasting in Orthogonally Transformed Domain." NeurIPS, 2025.

AuthorsOpenReview ↗

Follow-up to Reviewer aPGz

Dear Reviewer aPGz,

We sincerely appreciate your constructive comments and suggestions on our work. With the discussion period approaching its end, we would like to briefly follow up on the additional experiments and clarifications provided in our rebuttal.

Following your suggestions, we have added an ablation study on the hybrid time–frequency loss and an end-to-end training cost analysis to further evaluate the effectiveness of our components and the practical efficiency of HorizonBridge. We have also clarified the relationship between adaptive patching and periodic embedding, the implementation of channel interactions, the embedding configurations and training settings across experiments to improve clarity and reproducibility.

We hope these additions have helped address your concerns and strengthen the experimental evidence presented in our paper. If any questions remain, we would be happy to provide further clarification before the discussion ends.

Thank you again for your time and thoughtful consideration.

Best regards,
Authors of Submission 30237

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.