acceptodds
Under review as a conference paper at ICLR 2027

LatentTime: Efficient Long-context Time Series Foundation Models via Latent Forecasting

Abstract

Time series foundation models (TSFMs) have emerged as a promising approach to general-purpose forecasting. As public pre-training corpora expand, recent TSFMs increasingly combine larger model capacities with longer input contexts. However, fine-grained patching and padding of variable-length sequences introduce substantial computational overhead. To solve this problem, we propose LatentTime, an efficient long-context forecasting framework that combines continuous latent forecasting with a Length-Aware Bucket strategy. Its lightweight TS-AE encoder compresses long input contexts by , enabling forecasting in the resulting latent space. The decoder then performs temporal super-resolution on the latent predictions to restore the original resolution. Together, these operations approximately halve input tokens and double the forecast span per prediction block. The Length-Aware Bucket strategy further reduces padding overhead and controls training exposure across context lengths. Experiments on the GIFT-Eval benchmark demonstrates that LatentTime achieves forecasting accuracy comparable to that of raw-space PatchTST with lower forward FLOPs, while TS-AE maintains high reconstruction fidelity for long contexts across diverse time series.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Forecasting, Time Series Foundation Model, Latent Forecasting Model

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_hjddOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes LatentTime, a framework for efficient long-context time-series foundation models that reduces computational overhead from long input contexts and variable-length padding. Specifically, LatentTime introduces (i) Latent Forecasting, which uses a lightweight autoencoder to compress input sequences by 2× and reconstruct forecasts through temporal super-resolution, and (ii) Length-Aware Bucketing, which reduces padding overhead and controls training exposure across different context lengths. Experiments on the GIFT-Eval benchmark demonstrate forecasting performance comparable to raw-space PatchTST with lower computational costs.

Critical strengths
  • While most time-series foundation models focus on autoregressive forecasting and Transformer architectures in the raw input space [1, 2], this paper explores a meaningful direction toward improving long-context efficiency through latent-space forecasting.

  • The Length-Aware Bucketing strategy is a practical approach that improves computational efficiency over conventional padding methods by reducing unnecessary computation, while also demonstrating improved forecasting performance in the experiments.

Critical weaknesses
  • I believe that demonstrating how to construct effective latent representations is crucial to this paper. However, the latent representation design is not explored in sufficient depth, as the proposed method largely relies on a conventional autoencoder. While the spectral alignment loss is introduced to improve latent representations, Figure 11 only provides visualizations of selected examples. Quantitative comparisons are needed to demonstrate its effectiveness in learning suitable latent representations for forecasting.

  • The effectiveness of latent forecasting is not sufficiently clear. As shown in Figure 5, LatentTime achieves nearly identical performance to PatchTST-WB, with only a 0.002 difference in MASE. However, LatentTime switches to raw-space forecasting for context lengths below 512, and the proportion of such cases in the evaluation is not reported. If these cases account for a substantial portion of the benchmark, the overall results may mask a larger performance gap between latent-space and raw-space forecasting.

Questions for authors
  • Could you provide a quantitative ablation study on the effectiveness of the spectral alignment loss, particularly for latent representation quality and downstream forecasting performance?

  • Could you provide a separate comparison between LatentTime and PatchTST baselines in Figure 5 using only evaluation cases with context lengths of at least 512? I think it is important to isolate the performance of latent forecasting from raw-space forecasting.

  • Have you explored alternative latent compression methods, including but not limited to cross-attention-based approaches [3, 4]? How do they compare with the proposed autoencoder-based design?

[1] Das et al., "A Decoder-Only Foundation Model for Time-Series Forecasting", ICML 2024
[2] Shi et al., "Time-MoE: Billion-Scale Time Series Foundation Models with Mixture of Experts", ICLR 2025
[3] Lee et al., "TimePerceiver: An Encoder-Decoder Framework for Generalized Time-Series Forecasting", NeurIPS 2025
[4] Liu et al., "UniTST: Effectively Modeling Inter-Series and Intra-Series Dependencies for Multivariate Time Series Forecasting", TMLR 2025

9 comments

Dear Reviewer,

Thank you for your time and effort in providing constructive feedback on our paper. We have further explored the module architecture and pre-training loss of TS-AE in the paper. In our latest updated PDF of the paper, we have marked the revisions in blue. All section references below are to the revised manuscript. Below, we will answer your questions one by one.

W1 (1): I believe that demonstrating how to construct effective latent representations is crucial to this paper. However, the latent representation design is not explored in sufficient depth, as the proposed method largely relies on a conventional autoencoder.

Thank you for your question. We would like to clarify the following five points:

  1. Conventional autoencoder architecture. The basic architecture of an autoencoder is well defined and has been supported by a clear mathematical framework since its introduction. Subsequent work has largely modified the neural network modules within this basic architecture.
  2. A more lightweight and efficient design. We acknowledge that the design of TS-AE draws inspiration from AE and VAE architectures used in computer vision, such as DC-AE for image generation and Wan-VAE for video generation. However, we also introduce designs tailored to the characteristics of time series, making TS-AE lightweight enough to compress and reconstruct diverse time series by 2×2 \times with only 1.8M parameters. These designs include: (1) linear attention along the temporal dimension to improve global modeling; (2) additional causal convolutions to model temporal dependencies; and (3) multi-scale convolutions for local feature extraction. These are our improvements to the conventional autoencoder.
  3. Integration with bucketing. Autoencoders previously used in time series analysis are often limited to fixed-length sequences or sequences of the same length within a batch. In contrast, both the training and architecture of TS-AE are closely integrated with our bucketing strategy. This enables TS-AE to handle variable-length sequences and substantially improves parallel processing efficiency through bucketing.
  4. Progressive exploration of model architectures. Our initial autoencoder used a Transformer architecture rather than a convolutional architecture. However, we observed substantially different latent representations from the generative Transformer and time-domain convolutions, which can be viewed as temporal filters. This issue is discussed further in our response to Q3. After changing to a convolutional architecture, we still found that the latent space did not preserve some details sufficiently. We also observed that the latent representations were inverted. To improve detail preservation, we introduced a spectral alignment loss.
  5. Utility of the latent space. We conducted extensive experiments to evaluate the effectiveness of TS-AE's latent space: (1) latent-space forecasting in RQ1; (2) reconstruction evaluation and architectural ablations in RQ2; (3) t-SNE visualization and evaluation of preserved forecastability in RQ3; and (4) evaluation of inter-channel correlation preservation in RQ4.

We sincerely hope that these responses address your concerns.

W1 (2): While the spectral alignment loss is introduced to improve latent representations, Figure 11 only provides visualizations of selected examples. Quantitative comparisons are needed to demonstrate its effectiveness in learning suitable latent representations for forecasting.

Thank you for your valuable feedback. We have added an ablation experiment for this loss in Section 4, RQ2, and the specific experimental results are shown in the updated Figure 7. In our response to Q1, we provided a detailed analysis of the problem, experimental results, and reasons.

W2: The effectiveness of latent forecasting is not sufficiently clear. As shown in Figure 5, LatentTime achieves nearly identical performance to PatchTST-WB, with only a 0.002 difference in MASE. However, LatentTime switches to raw-space forecasting for context lengths below 512, and the proportion of such cases in the evaluation is not reported. If these cases account for a substantial portion of the benchmark, the overall results may mask a larger performance gap between latent-space and raw-space forecasting.

Thank you for your valuable comments. We agree that this is an important question and have discussed it in detail in Appendix D.3. Among the 97 evaluation tasks in the GIFT-Eval benchmark, approximately 19 tasks, or 20%, use [RAW] for contexts shorter than 512. The remaining 78 tasks, or approximately 80%, have minimum context lengths greater than 512 and use only the [LAT] token for latent-space forecasting. So the majority of benchmark tasks are forecast entirely in the latent space. We provide further quantitative analysis of this issue in our response to Q2.

Q1: Could you provide a quantitative ablation study on the effectiveness of the spectral alignment loss, particularly for latent representation quality and downstream forecasting performance?

Thank you for your constructive suggestion. We have added an ablation study in RQ2 to evaluate the effect of the spectral alignment loss. The results are shown as "w/o Align" in the updated Figure 7. We conduct the ablation from three perspectives: reconstruction loss, robustness to missing values, and downstream forecasting. The results show that this loss has a limited effect on the reconstruction performance of TS-AE. However, when the model is applied to downstream forecasting, its forecasting performance declines noticeably without the loss.

This loss was designed to address the inverted latent representations produced by the original autoencoder and the insufficient detail in their periodic components. By constraining the power spectral density, it helps the model preserve some linear correlations during compression, thereby substantially improving the details retained in the latent representations. For this reason, our forecasting performance in the latent space has also been greatly enhanced. We have revised Section 4.2 to address this issue.

Q2: Could you provide a separate comparison between LatentTime and PatchTST baselines in Figure 5 using only evaluation cases with context lengths of at least 512? I think it is important to isolate the performance of latent forecasting from raw-space forecasting.

Following your suggestion, we have added an analysis in Appendix D.3. Based on the use of [RAW] and [LAT], we identified 78 evaluation tasks whose minimum context length is greater than 512 and which use only [LAT] for latent-space forecasting. In the remaining tasks, 15 use only [RAW] for forecasting in the original space, while 4 use a mixture of the two modes. The dataset statistics and partitioning are provided in Tables 27 and 28. Following the original metric computation procedure, we calculated results for these two task groups, as shown below and in Table 29. The 78-task group isolates [LAT] tasks. It does not include [LAT] windows from the four mixed-mode tasks.

Minimum context lengthModelGmean MASEMedian MASEGmean CRPSMedian CRPS
>512>512 (78)PatchTST0.7950.7550.4440.423
PatchTST-WB0.7930.7320.4340.412
LatentTime0.7950.7410.4370.408
PatchTST-P32-WB0.8110.7810.4490.475
PatchTST-P320.8140.7660.4660.467
≤512\le 512 (19)PatchTST0.5130.6250.7990.84
PatchTST-WB0.4920.6170.7650.78
LatentTime0.4940.6210.7680.785
PatchTST-P32-WB0.4930.6080.7490.796
PatchTST-P320.4910.640.7760.764

On the 78 LAT-only tasks, LatentTime matches PatchTST in geometric-mean MASE at the reported precision and improves on its other three metrics. It also outperforms both patch-size-32 variants on all four metrics. Compared with PatchTST-WB, LatentTime remains close in geometric-mean MASE (0.795 vs. 0.793) and CRPS (0.437 vs. 0.434), while achieving a lower median CRPS (0.408 vs. 0.412). LatentTime achieves these results with a 2×2 \times compressed input context; the RQ6 measurements show lower FLOPs and latency for longer contexts under the tested settings.

These results support the effectiveness of LatentTime's latent-space forecasting, while the RQ6 experiments provide evidence of its efficiency in processing long-context sequences.

Q3: Have you explored alternative latent compression methods, including but not limited to cross-attention-based approaches [3, 4]? How do they compare with the proposed autoencoder-based design?

This is an interesting question. We did try different architectures and ultimately chose the one we are currently using. Most of the references you provided use Transformer architectures. Our initial compression model was also Transformer-based. Compared with our current convolutional model, its main advantage is that compressed patch representations can be retained directly as downstream tokens. Transformers also have better support for software and hardware acceleration techniques such as FlashAttention and KV caching.

However, we observed a serious issue: compression with the Transformer architecture appeared to be generative. Both latent compression by the encoder and reconstruction from the compressed latent representation by the decoder seemed to require the model to "generate" a new output. In contrast, time-domain convolutions act as natural filters: each convolution can be understood as extracting and filtering features from the existing representation through a linear transformation, rather than generating a new result in the same way. We analyze this phenomenon in detail in Appendix B.6. Figure 14 compares the latent spaces constructed by the convolutional and Transformer architectures. The Transformer-generated representations appear overly smooth and fail to reflect trend variations and periodic details.

We also conducted reconstruction and forecasting experiments on the Transformer-based TS-AE, following an approach similar to that used in Q1. The results are presented in RQ2 of Section 4. Although its latent-space characteristics are noticeably weaker than those of the convolutional model, the Transformer-based model can still achieve good 2×2\times compression and reconstruction of time series. However, it performs poorly on downstream time-series generation. We believe this may be because its latent representation is generated rather than obtained through natural filtering. In other words, details in the latent space are also critical to whether the forecasting model can learn predictive representations and successfully forecast in the latent space. For this reason, we ultimately chose a convolutional network as the main architecture of TS-AE. We have included the paper you mentioned in the discussion of our paper.

We thank you again for your time and constructive feedback. We hope that our responses and revisions address your concerns, and we look forward to your further feedback.

Reviewer_y1cHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes LatentTime to reduce the computational cost of long-context time series foundation models. It uses a 1.8M-parameter deterministic convolutional autoencoder TS-AE to compress the context by 2x in time, forecasts in the latent space with a 21M-parameter PatchTST predictor, and applies 2x temporal super-resolution in the decoder, activating latent mode only when all sequences in a batch have at least 512 valid steps while otherwise falling back to raw space. It also employs a Length-Aware Bucket strategy to reduce padding waste. In zero-shot evaluation on GIFT-Eval, LatentTime achieves a MASE and CRPS of 0.724 and 0.488, outperforming PatchTST-P32-WB at 0.736 and 0.498 and performing close to PatchTST-WB at 0.722 and 0.485.

Critical strengths

The motivation is clear and the design is coherent, introducing latent-space modeling and 2x decoder super-resolution to shorten input tokens and autoregressive steps while systematically evaluating TS-AE reconstruction fidelity up to length 8192 and inter-channel correlation, and candidly motivating the raw-versus-latent switch from weak short-context reconstruction.

The baseline comparisons are targeted, using PatchTST-P32-WB as an equal-token alternative, separating the gains of bucketing from latent compression, and openly acknowledging in the appendix that parallelism, effective batch size, and length diversity are confounded.

Critical weaknesses
  1. The central selling point is efficiency, but direct, realistic evidence of efficiency is missing, and accuracy shows no advantage. The abstract and conclusion repeatedly claim lower forward FLOPs, but the FLOPs curves in Fig. 2 involve Base/Large predictors, which Appendix C.1 says were never trained; only the 21M Small predictor is used for all forecasting experiments. The actual savings for the evaluated model are therefore not clearly stated in the main text, and the paper itself notes that savings grow with predictor size and context length, implying the Small model benefits the least.

  2. I could not find end-to-end wall-clock latency, throughput, memory or training-time comparisons. The real cost of TS-AE (convolutions plus linear attention) and of sequential bucket processing (the appendix concedes reduced throughput under single-threaded execution) could offset theoretical FLOP savings; FLOPs do not equal speed.

  3. The motivation is long context, yet most GIFT-Eval tasks have moderate contexts, and latent mode is activated only when all sequences in a batch have ≥512 valid steps. The paper does not report what fraction of GIFT-Eval predictions actually use [LAT], nor does it separate [LAT]/[RAW] results or provide accuracy/cost curves as context length grows. The long-context claim is thus not directly tested.

Questions for authors
  1. There are some compilation issues in the paper, and some content is displayed as garbled characters, such as the table in Figure 6.

  2. For the Small (21M) predictor, please report measured end-to-end inference latency, throughput and memory of LatentTime vs. PatchTST-WB at several context lengths (e.g., 1K/2K/4K/8K), plus training wall-clock including TS-AE overhead and sequential bucket processing. Do the Base/Large FLOPs in Fig. 2 correspond to actual models?

3 .What fraction of GIFT-Eval windows/tasks actually use the [LAT] path? Please report MASE/CRPS separately on [LAT] and [RAW] subsets for LatentTime vs. PatchTST-WB, and accuracy as a function of context length, to substantiate long-context gains.

5 comments

Dear Reviewer,

We sincerely thank you for your valuable feedback and assistance with our paper. In response, we first added a dedicated experimental section to discuss the efficiency of latent space prediction and bucketing strategies used in LatentTime for long input contexts. Then, we primarily added specific metrics such as actual latency, throughput, computational complexity, and peak memory usage through practical quantification. Following your suggestions, we then revised some other details. In our latest updated PDF of the paper, we have marked the revisions in blue. All section references below are to the revised manuscript. Below, we will address your questions one by one.

W1: The central selling point is efficiency, but direct, realistic evidence of efficiency is missing, and accuracy shows no advantage. The abstract and conclusion repeatedly claim lower forward FLOPs, but the FLOPs curves in Fig. 2 involve Base/Large predictors, which Appendix C.1 says were never trained; only the 21M Small predictor is used for all forecasting experiments. The actual savings for the evaluated model are therefore not clearly stated in the main text, and the paper itself notes that savings grow with predictor size and context length, implying the Small model benefits the least.

Thank you for your valuable suggestion. We have added RQ6 in Section 4 to report the computational cost of LatentTime-Small, the model used in our forecasting experiments. We also compare the computational efficiency of the Base and Large variants. Specifically, we report total FLOPs, latency, throughput, and peak GPU memory, extending the computational complexity analysis in Figure 2(a). As shown in Figure 11, we measure these four metrics across input context lengths for PatchTST predictors of different sizes and the corresponding LatentTime models equipped with the 1.8M-parameter TS-AE. The forecast length is fixed at 512. The results show that LatentTime's efficiency advantages become more apparent as the input context grows and as the predictor size increases. For a fair comparison, all experiments were conducted in FP32 on V100 GPUs with 32 GB of memory.

Our available training budget was limited to 8 NVIDIA V100 GPUs. Considering the computational resources required to train time series foundation models such as Chronos-2 and TiRex-2, we were unable to pre-train the Base and Large variants to a high standard. So we focused full pre-training and forecasting evaluation on the Small model to demonstrate the effectiveness and feasibility of latent-space forecasting. The Base and Large predictors in Figure 2(a) and RQ6 are used only for computational efficiency analysis. Their results show how efficiency varies with predictor size and context length; they do not represent the forecasting performance of trained larger models. Motivated by the scaling laws of DNNs, we may pre-train larger predictors to pursue better performance when sufficient computational resources are available.

W2: I could not find end-to-end wall-clock latency, throughput, memory or training-time comparisons. The real cost of TS-AE (convolutions plus linear attention) and of sequential bucket processing (the appendix concedes reduced throughput under single-threaded execution) could offset theoretical FLOP savings; FLOPs do not equal speed.

Thank you for your valuable comments. We agree that FLOPs alone cannot establish practical speed. The new RQ6 reports end-to-end latency, throughput, peak GPU memory, and FLOPs for complete forecasts, including TS-AE encoding and decoding. As shown in Figure 11, latent-space forecasting improves computational efficiency for long contexts, and this effect becomes more apparent as the model size and input context length increase

In Figure 12, we further evaluate the effect of bucketing on computational speed and related metrics. The results show that sequential bucketing reduces FLOPs and single-GPU latency at batch sizes of 128 and above, while parallel bucketing further reduces wall-clock latency when four GPUs are used. These results distinguish the effects of reduced padding from the additional benefits of hardware parallelism.

LatentTime is obtained by training the pre-trained PatchTST-WB for an additional 500K steps to adapt it to latent-space forecasting. Using the configuration in Table 4 and 8 V100 GPUs, this process takes approximately 53 hours. However, pre-training time is generally less of a focus for foundation models because they are often used for zero-shot, plug-and-play forecasting in practice. We have added the total time recorded in the training logs to Appendix C.1.

In summary, the results from end-to-end wall-clock latency, throughput, and memory usage demonstrate that the latent space prediction and bucketing algorithm used in this paper can significantly improve the efficiency of processing long contexts.

W3: The motivation is long context, yet most GIFT-Eval tasks have moderate contexts, and latent mode is activated only when all sequences in a batch have ≥512 valid steps. The paper does not report what fraction of GIFT-Eval predictions actually use [LAT], nor does it separate [LAT]/[RAW] results or provide accuracy/cost curves as context length grows. The long-context claim is thus not directly tested.

Thank you for your valuable quections. We have added a discussion in Appendix D.3 of the proportions of latent-space and raw-space forecasting on GIFT-Eval. Based on the use of [RAW] and [LAT], we identified 78 evaluation tasks with minimum context lengths greater than 512, which use only [LAT] for latent-space forecasting. Another 15 tasks use only [RAW] for forecasting in the original space, while four use a mixture of the two modes. The dataset statistics and partitioning are provided in Tables 27 and 28. Following the original metric computation procedure, we calculated forecasting results for the long-context latent-space group and the shorter-context group, which includes both RAW-only and mixed-mode tasks. The results are shown below and in Table 29.

PartitioningModelGmean MASEMedian MASEGmean CRPSMedian CRPS
>512>512 (78)PatchTST0.7950.7550.4440.423
PatchTST-WB0.7930.7320.4340.412
LatentTime0.7950.7410.4370.408
PatchTST-P32-WB0.8110.7810.4490.475
PatchTST-P320.8140.7660.4660.467
≤512\le 512 (19)PatchTST0.5130.6250.7990.84
PatchTST-WB0.4920.6170.7650.78
LatentTime0.4940.6210.7680.785
PatchTST-P32-WB0.4930.6080.7490.796
PatchTST-P320.4910.640.7760.764

These results show that LatentTime uses [LAT] for latent-space forecasting on the majority of GIFT-Eval tasks and achieves good performance on long contexts. They support the effectiveness of LatentTime's latent-space forecasting and its efficiency in processing long-context sequences.

Besides, the primary motivation of this work is the efficiency of processing contexts rather than the effectiveness of longer contexts themselves. Model performance can be affected by many factors, among which input context length may be an important one. Based on this quections, we examined how forecasting performance changes with historical context length in Appendix C.13. Our results show that MASE generally improves to some extent as the historical context grows, while CRPS, used to evaluate prediction uncertainty, appears more robust to changes in context length.

Q1: There are some compilation issues in the paper, and some content is displayed as garbled characters, such as the table in Figure 6.

Thank you for your careful review of our paper. We have reviewed and updated it, and checked for the compilation errors.

Q2: For the Small (21M) predictor, please report measured end-to-end inference latency, throughput and memory of LatentTime vs. PatchTST-WB at several context lengths (e.g., 1K/2K/4K/8K), plus training wall-clock including TS-AE overhead and sequential bucket processing. Do the Base/Large FLOPs in Fig. 2 correspond to actual models?

Thank you for your constructive comments on LatentTime's efficiency for long-context forecasting. We have added RQ6 to the paper to investigate this issue. As discussed in our responses to W1 and W2, we compare LatentTime and PatchTST-WB in terms of FLOPs, end-to-end latency, throughput, and peak GPU memory across context lengths from 512 to 4096 (max context in pre-training). The results are shown in Figure 11, where the gray dashed line represents the overhead of TS-AE itself. In Figure 12, we further compare the practical effects of the bucketing strategy. The results show that the advantages of sequential and parallel bucketing become more apparent as batch size increases. The performance improvement mainly comes from reducing the number of redundant tokens introduced by zero-padding, as illustrated in Figure 2(b). We have also reported the total time required under our current pre-training resources and configuration in Appendix C.1.

Finally, our current hardware resources are insufficient for adequate pre-training of Base and Large-scale models. Therefore, as we analyzed in W2, these two specifications do not correspond to actual models but are analyzed theoretically to demonstrate that our method will have a more context-efficient advantage for larger-scale prediction models. However, our pre-training of small-scale models has already demonstrated the feasibility of our latent space forecasting.

We thank you again for your time and constructive feedback. We hope that our responses and revisions address your concerns, and we look forward to your further feedback.

Reviewer_gSx7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

LatentTime compresses historical sequences with a deterministic autoencoder, forecasts in latent space, and decodes predictions at the original resolution. It reports accuracy close to P16 bucketed PatchTST with lower forward FLOPs, while retaining a raw-space path for short contexts.

Critical strengths
  1. Learned temporal compression offers a practical alternative to larger patches. The P32 and bucketed controls are useful, and reconstruction ablations plus forecasting probes provide evidence beyond reconstruction fidelity alone.
Critical weaknesses

The evaluation does not support the stated zero-shot claim. The predictor uses GIFT-Eval training splits (§4; Appendix C.2), which conflicts with the benchmark’s strict zero-shot definition. Comparisons with external checkpoints should distinguish their training-data exposure. Checkpoint selection also needs clarification: Appendix C.2 describes testing several validation-shortlisted checkpoints and taking “the best result,” leaving unclear whether test scores influenced the final selection.

The practical benefit of the added complexity is not sufficiently established. LatentTime adds three-stage autoencoder training and 500K predictor adaptation updates, but improves MASE over P32-WB only from 0.736 to 0.724. A control receiving equivalent additional training is missing. Aggregate results also mix RAW and LAT modes because short contexts bypass compression. End-to-end latency, peak memory, and accuracy restricted to LAT windows are needed to establish whether the savings justify the training and architectural complexity.

Questions for authors

Given the use of GIFT-Eval training splits, how should the results be distinguished from strictly zero-shot evaluation? Were final checkpoints selected solely by validation performance, or did test scores influence selection among the shortlisted checkpoints?

Does the improvement over P32-WB persist when that baseline receives an equivalent additional training budget? Please report end-to-end latency, peak memory, the proportion of windows using LAT mode, and forecasting accuracy on those windows.

4 comments

Dear Reviewer,

Thank you for the time and effort you devoted to reviewing our paper and for your constructive comments and questions. In response, we have clarified the evaluation setting and checkpoint selection procedure, extended the training of the baselines with 32 patch size, analyzed [LAT] token usage and forecasting performance across context lengths, and added experiments on inference efficiency. In our latest updated PDF of the paper, we have marked the revisions in blue. Below, we respond to each comment and question point by point. All section references below are to the revised manuscript.

W1 (1): The evaluation does not support the stated zero-shot claim. The predictor uses GIFT-Eval training splits (§4; Appendix C.2), which conflicts with the benchmark's strict zero-shot definition. Comparisons with external checkpoints should distinguish their training-data exposure.

Thank you for pointing out the issue concerning zero-shot generalization in our experiments. We have revised RQ1 in Section 4 by removing the zero-shot claim, which had been mistakenly retained. Regarding comparisons with external checkpoints, all the comparison models shown in Figure 5, as well as the LatentTime and PatchTST models we validated, were strictly divided into "Zero-Shot" and "PreTrained" types according to the official requirements of GIFT-Eval.

Specifically, we changed the question from "How do 2×2\times context compression with TS-AE and the Length-Aware Bucket strategy affect zero-shot generalization in time series forecasting?" to "How do 2×2\times context compression with TS-AE and the Length-Aware Bucket strategy affect time series forecasting performance?". No other zero-shot claims remain in the paper, including in the main conclusions of Section 4.1.

In addition, to further distinguish our experimental setup for the forecasting task, we added a description in the Experiment Setup section of Section 4.1, explicitly specifying that our model belongs to the "Pretrained" type rather than zero-shot, in order to completely eliminate any misunderstandings about this issue.

W1 (2): Checkpoint selection also needs clarification: Appendix C.2 describes testing several validation-shortlisted checkpoints and taking “the best result,” leaving unclear whether test scores influenced the final selection.

Thank you for your constructive comments and concerns regarding checkpoint selection. First, our forecasting models underwent large-scale pre-training: PatchTST was trained for 800K steps, and LatentTime was trained for an additional 500K steps. We saved model checkpoints every 1K steps. Consequently, we had a large number of candidate checkpoints after pre-training. Then, we used the validation split for preliminary screening and selected the five best checkpoints based on validation performance. This step substantially narrowed the range of candidate checkpoints and no test data were used for screening during this process. Finally, we evaluated these five checkpoints using the GIFT-Eval benchmark and presented the best result among them.

In summary, the test datasets were used only to further evaluate the five best checkpoints that had already been selected. We mainly showcased the results that demonstrated better performance in the tests. Our validation split had already excluded a large number of checkpoints and eliminated the influence of the test datasets on the overall direction of checkpoint screening. Compared with the supervised learning experiments, this process is broadly consistent with checkpoint selection and evaluation for large-scale pre-trained foundation models.

To explain this process in greater detail, we have revised the description of checkpoint selection in Appendix C.2.

W2: The practical benefit of the added complexity is not sufficiently established. LatentTime adds three-stage autoencoder training and 500K predictor adaptation updates, but improves MASE over P32-WB only from 0.736 to 0.724. A control receiving equivalent additional training is missing. Aggregate results also mix RAW and LAT modes because short contexts bypass compression. End-to-end latency, peak memory, and accuracy restricted to LAT windows are needed to establish whether the savings justify the training and architectural complexity.

Thank you for your constructive suggestions regarding model pre-training and the effectiveness of LatentTime. Accordingly, we have made the following revisions:

  1. We retrained the two patch size 32 controls, PatchTST-P32 and PatchTST-P32-WB, while keeping all other experimental conditions unchanged. We increased their training duration to 1,300K steps, corresponding to the 800K PatchTST pre-training steps plus the 500K latent-space adaptation steps of LatentTime. We have updated the benchmark results shown in Figures 3 and 5 near the beginning of the paper to ensure a fair comparison. The results show that PatchTST-P32's evaluation results are nearing saturation after 800K steps of training. Therefore, the performance improvement brought by the additional 500K steps is minimal. A detailed analysis is provided in our response to Q2 (1).
  2. We analyzed the proportion of GIFT-Eval tasks using the [LAT] token and latent-space forecasting according to their input context lengths. Among the 97 evaluation tasks, 78 have a minimum input context length greater than 512. Thus, approximately 80% of the evaluation tasks are forecast entirely in the latent space by LatentTime. We discuss this issue further in Q2 (2) and analyze and compare the performance of different models according to whether [LAT] is used for latent-space forecasting. We have added Appendix D.3 to discuss this issue and present the full results.
  3. In Q2 (2), we further analyze the efficiency of latent forecasting and the bucketing strategy for long input contexts. Specifically, we have added RQ6 to the experimental section in Section 4 to analyze computational complexity. For a complete forecast, we report the total floating-point operations in Figure 11(a), latency in Figure 11(b), throughput in Figure 11(c), and peak memory in Figure 11(d).

Q1: Given the use of GIFT-Eval training splits, how should the results be distinguished from strictly zero-shot evaluation? Were final checkpoints selected solely by validation performance, or did test scores influence selection among the shortlisted checkpoints?

Thank you for your helpful comments and questions. First, We have removed the incorrect wording from RQ1 in Section 4 and clarified that our model type is "Pretrained" rather than "Zero-Shot". Then, we mainly explore whether the latent space prediction paradigm and bucketing strategy affect prediction performance, with controllable variables. We have also clarified in Section 4.1 that all models pre-trained in this work, including LatentTime and its comparison variants, are categorized as Pretrained.

Q2 (1): Does the improvement over P32-WB persist when that baseline receives an equivalent additional training budget?

Thank you for your valuable comments and suggestions. To address this question, we retrained PatchTST-P32 and PatchTST-P32-WB using the same pre-training data, training procedure, optimizer, learning-rate schedule, and hardware environment as before. We increased their training duration to 1,300K steps, corresponding to the combined number of steps used for PatchTST pre-training and LatentTime's latent-space adaptation. The results are as follows:

ModelMASECRPS
PatchTST-P320.739 → 0.7370.505 → 0.514
PatchTST-P32-WB0.736 → 0.7350.498 → 0.496

These results suggest that PatchTST has nearly saturated after approximately 800K training steps. Additional training yields only modest improvements; indeed, the CRPS of PatchTST-P32 deteriorates more substantially. Overall, these variants still lag behind PatchTST with a patch size of 16 and LatentTime with latent-space forecasting. Therefore, while ensuring fairness in computing resources, LatentTime consistently outperforms PatchTST-P32 and -P32-WB.

We have updated the GIFT-Eval benchmark comparison in RQ1 of Section 4.1 and the MASE comparison in Figure 3 (left). We have also provided further details of the pre-training configurations in Appendix C.1.

Q2 (2): Please report end-to-end latency, peak memory, the proportion of windows using LAT mode, and forecasting accuracy on those windows.

Thank you for your constructive suggestion. We have added RQ6 to the experimental section to investigate the computational efficiency of latent forecasting and the bucketing strategy for long input contexts. In addition, we have added an analysis in Appendix D.3 of the use of the [LAT] token for latent-space forecasting on GIFT-Eval.

First, RQ6 extends the experiment in Figure 2(a) from the motivation section. We analyze the computational cost of LatentTime using context lengths of 512, 768, 1024 with 256 step up to 4096. We directly compare LatentTime with PatchTST at the same model size. The only additional component in LatentTime is TS-AE, which compresses the context into the latent space. For a forecast length of 512 at each fixed input context length, we record total FLOPs, latency, throughput, and peak memory in Figure 11. All experiments were conducted in FP32 on a multi-GPU V100 server with 32 GB of memory per GPU. As in Figure 2(a), we evaluate Small, Base, and Large variants to examine how computational cost changes with context length at different model sizes. Figure 11 shows that LatentTime gains clear advantages in computational cost and latency as the input context grows, and these efficiency gains become more pronounced as the size of the PatchTST predictor in the latent space increases. These results support the efficiency of latent representation and forecasting approach for long input contexts.

We also report the proportion of windows using the [LAT] token in Appendix D.3. GIFT-Eval contains 97 evaluation tasks, each with a minimum and maximum input context length. Among them, 78 have input contexts consistently longer than our threshold of 512. Thus, approximately 80% of the evaluation tasks use [LAT] and perform latent-space forecasting throughout. Another 15 tasks have a maximum input context length shorter than 512, while four have a minimum context length shorter than 512 and a maximum context length longer than 512. Together, these account for 20% of the tasks. We list these datasets in Tables 27 and 28, respectively. We further partitioned our experimental results into these two groups and computed the metrics following the original GIFT-Eval procedure:

PartitioningModelGmean MASEMedian MASEGmean CRPSMedian CRPS
>512>512 (78)PatchTST0.7950.7550.4440.423
PatchTST-WB0.7930.7320.4340.412
LatentTime0.7950.7410.4370.408
PatchTST-P32-WB0.8110.7810.4490.475
PatchTST-P320.8140.7660.4660.467
≤512\le 512 (19)PatchTST0.5130.6250.7990.84
PatchTST-WB0.4920.6170.7650.78
LatentTime0.4940.6210.7680.785
PatchTST-P32-WB0.4930.6080.7490.796
PatchTST-P320.4910.640.7760.764

These results show that LatentTime uses [LAT] for latent forecasting on the majority of GIFT-Eval benchmark and achieves good performance on long contexts. The results support the effectiveness of forecasting in latent space and its efficiency in processing long-context time series.

We thank you again for your time and constructive feedback. We hope that our responses and revisions address your concerns, and we look forward to your further feedback.

Dear Reviewer,

As the discussion period is drawing to a close, we would like to ask whether our responses and revisions address your main concerns. We have clarified the evaluation setting and checkpoint selection procedure, extended the training of the P32 baselines, and added analyses of latent-mode forecasting and inference efficiency.

Thank you again for your time and constructive feedback. We look forward to your response.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.