- 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
- 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.
- 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
- 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.
- 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.
- 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
- 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.
- 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
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 (), demonstrating the robustness of its gains.
| Comparison | Metric | Wilcoxon Statistic | p-value | Significant |
|---|---|---|---|---|
| HorizonBridge vs PhaseFormer | MSE | 0 | 7.813 × 10⁻³ | Yes |
| HorizonBridge vs VPNet | MSE | 0 | 7.813 × 10⁻³ | Yes |
| HorizonBridge vs PhaseFormer | MAE | 3 | 3.906 × 10⁻² | Yes |
| HorizonBridge vs VPNet | MAE | 1 | 1.563 × 10⁻² | Yes |
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.
| Dataset | Model | Original Time (s) | Original Epochs | Original Iterations | +Translator Time (s) | +Translator Epochs | +Translator Iterations |
|---|---|---|---|---|---|---|---|
| Solar | PatchTST | 26999.1 | 56 | 62466 | 9455.4 | 23 | 25672 |
| Crossformer | 13314.0 | 45 | 50279 | 2248.5 | 16 | 17840 | |
| TimeXer | 7137.2 | 48 | 53660 | 3685.6 | 22 | 24570 | |
| Weather | PatchTST | 6014.3 | 55 | 61586 | 2540.9 | 25 | 28035 |
| Crossformer | 4312.2 | 41 | 45941 | 1254.4 | 16 | 17888 | |
| TimeXer | 1652.7 | 29 | 32470 | 848.2 | 15 | 16760 | |
| Traffic | PatchTST | 32981.7 | 29 | 81278 | 13728.7 | 15 | 41742 |
| Crossformer | 21293.1 | 49 | 137879 | 7693.6 | 22 | 61514 | |
| TimeXer | 47936.0 | 43 | 120401 | 22954.4 | 20 | 55936 |
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
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 R 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 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).
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. 10, the summation over K patches already reflects the fixed patch-count setting, which also ensures compatibility with the flatten head. To avoid confusion, we have replaced K with N in Eq. 10 for consistency with Algorithm 1 in the Appendix. The dataset-specific values of N and q are provided in the table below.