acceptodds
Under review as a conference paper at ICLR 2027

Rethinking Universal Time Series Foundation Models from Unified Representations to Shared Dynamics (UniTFU)

Abstract

Time series foundation models (TSFMs) have attracted increasing attention in cross-domain forecasting. Existing works mainly rely on large-scale pre-training and general modeling paradigms. However, the former heavily relies on the diversity and domain coverage of pre-training data. Meanwhile, the latter learns unified representations across heterogeneous domains without explicitly identifying the temporal regularities shared among them, limiting robust cross-domain generalization. This work explores cross-domain generalization by characterizing commonalities in recurring local dynamics across domains, rather than relying solely on implicitly learned temporal representations. To capture such cross-domain commonalities, we propose **Time Series Foundation Units (TFUs)** as reusable building blocks discovered from multi-scale temporal segments, each characterizing a recurring local dynamic pattern shared across domains. Building on TFUs, we develop a unified time series foundation model (**UniTFU**) that explicitly reuses and composes shared temporal regularities across domains for forecasting. Experiments on seven public datasets show that UniTFU substantially outperforms existing TSFMs in cross-domain forecasting and can efficiently adapt to target-domain dynamics by updating its reusable components with limited target data. UniTFU also achieves strong forecasting performance compared with state-of-the-art methods.

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

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_5zWtOpenReview ↗
Rating
3
Confidence
2(−1)
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes a time series foundation model UniTFU, that incorporates unit-level Time Series Foundation Units (TFUs) and their aggregated scene context into the attention mechanism to guide temporal dynamics modeling. The contributions include (1) introduce TFUs as reusable and composable prototypes of recurring local dynamics discovered from multi-scale temporal segments (2) UniTFU, a TSFM that explicitly integrates the discovery, reuse, and composition of TFUs into forecasting, enabling shared local dynamics to be leveraged across domains (3) systematically evaluate UniTFU in cross-domain settings and on standard forecasting benchmarks, demonstrating robust cross-domain generalization and strong forecasting performance.

Critical strengths
  • The authors proposed a novel Time Series Foundation Units (TFUs), where the motivation is clear and the computation is straightforward.
  • Based on TFUs, the authors introduced TFU attention that incorporates the learned prototype and contextualized information for time series forecasting.
  • The proposed method is shown to be effective in cross-domain forecasting and performs better than other foundation models.
Critical weaknesses
  • The discussion of Adaptive TFU Prototype Bank can be improved. Maybe consider the case with and without new dynamics. The authors introduced a few new terms in the paragraph after Eq. 6, such as anchor and G\mathcal{G}, and it would be better to make it clearer.
  • According to Table 3, the gap between UniTFU and other forecasting modules is small (ETTh2, ETTm2, ECL, Weather and Solar). Given the efficiency analysis in Table 6, it is not quite easy to tell why the seemingly more heavily loaded model UniTFU is better than the standard Forecasting Model.
Questions for authors
  • Is each prototype a learnable vector, randomly initialised?
  • For line 232, the prototype-conditioned TFU sequence U is written as a sequence of uu. If only one prototype matches one position, does it indicate u=pku = p_k?
  • The merge buffer and novelty buffer part is a bit unclear. Does this process (Eq. 5) change the total number of prototypes?
6 comments
AuthorsOpenReview ↗

Response to Reviewer 5zWt

Questions 1–3 and Weakness 1: Adaptive TFU Prototype Bank.

Thank you for these constructive comments. We clarify prototype initialization, TFU composition, and adaptive dictionary maintenance as follows:

  • Prototype definition and initialization. The TFU bank stores a collection of D-dimensional vectors representing recurring local temporal patterns, rather than raw waveforms. Each segment is encoded from the mean, peak, and temporal change of its token representations. The bank starts empty, with the first retained segment embedding initializing the first TFU. These prototypes are refined through grouped EMA updates (Eq. 6), rather than gradient descent. During cross-domain adaptation, the bank is initialized from the source checkpoint.

  • Single-prototype match. When only one TFU pkp_k is active, all retained candidates match it, yielding ut=pku_t=p_k at covered positions, even with overlapping segments. More generally, a single candidate covering position tt also yields ut=pku_t=p_k, while overlapping candidates assigned to different prototypes produce a confidence-weighted combination (Eq. 4).

  • Dictionary maintenance with or without new dynamics. Eq. (5) only routes candidates, without changing the number of active TFUs. Existing dynamics enter the merge buffer for EMA refinement. Novel candidates enter the novelty buffer, where an anchor candidate defines a group of similar embeddings. A new TFU is created only when the group contains sufficient candidates and capacity permits. Thus, prototype birth increases the active count, while consolidation of redundant prototypes decreases it; the maximum capacity remains fixed.

These mechanisms are further clarified in Section 3.1 and Appendix A.3 of the revised version.

Weakness 2. Forecasting Accuracy and Inference Cost.

Thank you for raising this concern. We additionally evaluated DLinear, PatchTST, iTransformer, and PhaseFormer under the same inference setting (672→336, batch size 1). The results are reported in Table 7 of the revised version, with forecasting errors in Appendix Table 12. We acknowledge that UniTFU incurs substantially higher inference costs for modest supervised forecasting gains (e.g., on Weather, MSE: 0.256 vs. 0.251; latency: 402.177 vs. 8.964 ms, compared with PatchTST).

However, UniTFU's primary benefit lies in cross-domain transfer rather than inference efficiency. In direct transfer, TFU learning reduces average MSE by 13.4% relative to w/o TFU and 4.3% relative to a fixed VQ dictionary (Section 4.2, Table 2), demonstrating the value of reusable local dynamics. We have clarified this trade-off in the revised manuscript.

AuthorsOpenReview ↗

Summary of Response to Reviewer 5zWt

Thank you again for your thoughtful and constructive comments, which have helped us improve our paper. We clarified the adaptive TFU mechanism and extended the efficiency analysis to examine the accuracy–efficiency trade-off. We kindly invite you to reconsider the contributions and insights of our work. We welcome any further questions or discussion.

AuthorsOpenReview ↗

Eagerly Await Your Response

Dear Reviewer 5zWt,

Thank you again for your valuable and constructive review. We have clarified the adaptive TFU mechanism and provided additional accuracy–efficiency comparisons, explicitly acknowledging the computational trade-offs in our revised manuscript.

We would greatly appreciate it if you could review our responses and share any remaining concerns. We would be happy to provide further clarifications before the rebuttal period ends. If our responses satisfactorily address your concerns, we kindly hope you could reconsider your rating.

Thank you again for your support and constructive feedback.

Best regards,

Authors

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

The paper proposes UniTFU, a TSFM designed to improve cross-domain forecasting by explicitly discovering and reusing shared local temporal dynamics. The authors argue that existings models rely heavily on diverse pre-training data or unified representations without explicitly modeling temporal regularities shared across domains.

To address these challenges, UniTFU introduces Time Series Foundation Units (TFUs), which represent recurring multi-scale temporal patterns organized in an adaptive prototype dictionary. These TFUs are integrated into a forecasting Transformer through scene context aggregation and TFU-guided attention.

Experiments demonstrate improved cross-domain transfer and competitive forecasting performance.

Critical strengths
  1. The introduction of adaptive and reusable temporal representations is the major contribution of this paper. The proposed TFU dictionary can dynamically update, expand, and consolidate temporal protypes. Unlike fixed-size codebooks, this adaptive mechanism provides greater flexibility when transferring across heterogeneous domains.
  2. The experiment demonstrates that updating TFUs with limited target-domain data imporves foreasting accuracy compared with keeping them frozen. At the same time, the model achieves competitive restuls across seven forecasting datasets, outperforming several established TSFM in both cross-domain transfer and standard forecasting tasks.
Critical weaknesses
  1. The novelty of UniTFU appears limited, as the idea of leveraging shared or domain-aware representations for cross-domain time series forecasting has already been explored in prior studies [1]. Although the adaptive prototype dictionary is potentially useful, the paper does not sufficiently establish how TFUs differ conceptually from existing prototype-based or vector-quantization methods [2-5].
  2. The experimental evaluation overlooks relevant cross-domain time series forecasting methods, particularly UniTime [1], which explicitly incorporates domain-specific information into a unified forecasting framework.

[1] UniTime: A Language-Empowered Unified Model for Cross-Domain Time Series Forecasting [2] TOTEM: TOkenized Time Series EMbeddings for General Time Series Analysis [3] ProtoTS: Learning Hierarchical Prototypes for Explainable Time Series Forecasting [4] A Wave Is Worth 100 Words: Investigating Cross-Domain Transferability in Time Series [5] APT: Affine Prototype-Timestamp for Time Series Forecasting Under Distribution Shift

Questions for authors

refer to the weaknesses above

5 comments
AuthorsOpenReview ↗

Response to Reviewer UQQH (Part 1)

Q1. Cross-domain modeling and TFU representations.

Thank you for recognizing the value of our adaptive TFU dictionary. We clarify the distinctions from prior work in cross-domain modeling and prototype-based representations:

1. Cross-domain modeling vs UniTime. UniTime [1] uses domain instructions to guide a unified forecasting model. In contrast, UniTFU explicitly identifies recurring local dynamics shared across domains, allowing source-learned TFUs to represent target-domain subsequences through temporally aligned composition. Thus, cross-domain knowledge reuse is grounded in shared temporal structures rather than explicit domain instructions.

2. TFUs vs existing methods. TFUs differ from the representations in [2–5] in three aspects:

  • Representation: TOTEM [2] learns a VQ-based tokenizer, while WQ4TS [4] constructs a shared wavelet-based spectral representation. UniTFU learns multi-scale local dynamic prototypes and composes overlapping matches into a temporally aligned TFU sequence.
  • Forecasting integration: ProtoTS [3] uses hierarchical forecasting prototypes, while APT [5] employs timestamp-conditioned prototypes to generate affine parameters. UniTFU instead combines local TFU representations with window-level scene context to condition token representations and modulate temporal attention.
  • Dictionary adaptation: Beyond prototype matching, UniTFU supports EMA refinement, novelty-driven prototype birth, and redundancy-based consolidation, adapting both prototype representations and the active vocabulary.

The table below summarizes the methods in [2–5]. Table 1 in the revised version presents the full comparison.

MechanismTOTEMProtoTSWQ4TSAPTUniTFU
Subsequence Matching✓✗✗✗✓
Learned Dictionary✓✓✗✓✓
Cross-Domain Forecasting✓✗✓✗✓
Scene-Conditioned Attention✗✗✗✗✓
Prototype Birth & Consolidation✗✗✗✗✓

The key novelty of UniTFU lies beyond prototype matching: it explicitly reuses and composes multi-scale local dynamics across domains, integrating them into scene-conditioned forecasting. Its adaptive dictionary further supports the refinement and reorganization of these reusable temporal units.

AuthorsOpenReview ↗

Response to Reviewer UQQH (Part 2)

Q2. Comparison with Cross-Domain Forecasting Methods.

Thank you for your valuable suggestion. We additionally evaluate three relevant cross-domain forecasting methods: UniTime [1] (language-guided domain modeling), LangTime [6] (language-guided forecasting with policy optimization), and LSTPO [7] (preference-guided meta-learning). Following the experimental protocol in Section 4.2, we evaluate direct transfer and 10% few-shot adaptation across 20 source–target pairs. The results are summarized below:

Direct Transfer

Target Domain (MSE / MAE)UniTFUUniTimeLangTimeLSTPO
ETTh10.537 / 0.4900.608 / 0.5440.597 / 0.5390.607 / 0.528
ETTm10.879 / 0.6010.875 / 0.6210.856 / 0.6040.855 / 0.602
Weather0.245 / 0.2870.255 / 0.3050.253 / 0.3010.250 / 0.302
ECL0.579 / 0.5600.694 / 0.6350.686 / 0.6230.682 / 0.619
Solar0.620 / 0.5620.885 / 0.7530.867 / 0.7340.861 / 0.748
Average0.572 / 0.5000.663 / 0.5710.652 / 0.5600.651 / 0.560

Target-Domain Adaptation (10% Few-shot)

Target Domain (MSE / MAE)UniTFUUniTimeLangTimeLSTPO
ETTh10.447 / 0.4530.479 / 0.4730.47 / 0.4790.497 / 0.489
ETTm10.555 / 0.4740.486 / 0.450.52 / 0.4560.469 / 0.437
Weather0.177 / 0.2230.185 / 0.2310.182 / 0.2280.185 / 0.235
ECL0.164 / 0.2590.172 / 0.2740.171 / 0.2680.176 / 0.281
Solar0.202 / 0.2410.221 / 0.2750.221 / 0.2640.228 / 0.284
Average0.309 / 0.3300.309 / 0.3410.313 / 0.3390.311 / 0.345

These comparisons are included in Section 4.2 (Tables 2–3) of the revised version, with complete results in Appendices B.1–B.2.

[1] UniTime: A Language-Empowered Unified Model for Cross-Domain Time Series Forecasting.

[2] TOTEM: TOkenized Time Series EMbeddings for General Time Series Analysis.

[3] ProtoTS: Learning Hierarchical Prototypes for Explainable Time Series Forecasting.

[4] A Wave is Worth 100 Words: Investigating Cross-Domain Transferability in Time Series.

[5] APT: Affine Prototype-Timestamp For Time Series Forecasting Under Distribution Shift.

[6] LangTime: A Language-Guided Unified Model for Time Series Forecasting with Proximal Policy Optimization.

[7] Preference Guided Meta-Learning for Cross Domain Time Series Forecasting.

AuthorsOpenReview ↗

Summary of Response to Reviewer UQQH

Thank you again for your valuable and constructive comments, which have helped us improve our paper. We clarified the distinctions from existing methods and expanded our cross-domain forecasting comparisons. We kindly invite you to reconsider the contributions and insights of our work. We welcome any further questions or discussion.

Reviewer_FqqNOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes using reusable and composable local dynamic prototypes (TFUs) to explicitly organize temporal patterns shared across domains, and introduces a framework called UniTFU. The novelty of this approach is open to debate.

Critical strengths

1、The method presented in this paper is simple and easy to understand.

Critical weaknesses

1、The novelty of the proposed method and how it differs from VQ (Vector Quantization), prototype-based learning, dictionary learning, and shapelet/motif-based approaches need to be clarified. Specifically, the relationship between the TFU prototype library and VQ codebooks, prototype networks, dictionary learning, and shapelet/motif methods should be more explicitly defined. Although the paper uses a "w/ VQ" (with VQ) configuration as a baseline for comparison, the TFU approach is, in essence, merely an improvement upon a VQ codebook.

2、For the experimental setup corresponding to Table 1, the evaluation of transfer learning should involve relevant baseline models; otherwise, the results would lack any substantive significance. The same applies to the comparative experiments in Table 2.

3、The evidence supporting the contribution of the TFU adaptive dictionary remains weak.In Table 2, "Updated-TFU" shows an improvement over "Frozen-TFU"; however, this small-scale training comparison is insufficient to definitively demonstrate the gains derived from the update process. Such comparisons should instead be conducted within a large-scale dataset pre-training framework. Furthermore, it remains to be investigated whether the gains from TFU updates are dependent on the initial vocabulary size.

4、There is insufficient validation regarding interpretability and "local dynamic semantics." While labels such as "rise," "fall," "oscillation," "switch," and "precursor" appear in Figure 2, the paper does not clarify whether these semantic descriptors are manually assigned, derived from visual interpretation, or represent verifiable properties. Figures 3 and 4 present only a single cross-domain case (ETTh1→Weather); furthermore, the improvement in similarity shown in Figure 3b may simply indicate that the updated prototypes are "closer" to the candidates, which does not necessarily translate into predictive gains. Additional evidence is required, including results across more source-target pairs, comparisons against random or frequency-based baselines, an analysis of the correlation between prototype usage rates and prediction errors, and quantitative data on local shape correlations.

5、What is the specific experimental design for UniTFU in Tables 3 and 4? Why do the model evaluation results differ between the two tables? Is data from downstream tasks included in the fine-tuning process?

6、In the ablation studies, the performance gains attributed to each ablated module differ only at the third decimal place; such figures call into question the credibility of these ablation experiments.

Questions for authors

See Weaknesses.

6 comments
AuthorsOpenReview ↗

Respnse to Reviewer FqqN (Part 1)

Q1. Technical contribution of TFUs.

Thank you for your valuable feedback. We acknowledge that TFU assignment shares the nearest-prototype matching principle with VQ. However, UniTFU is not merely an extension of VQ; its contribution lies in organizing, composing, and adapting reusable local dynamics for cross-domain forecasting. We clarify its distinctions from the five categories of related methods:

  • Cross-domain structural reuse. ShapeX [1] (shapelet-based) learns shapelets for classification explanations, while L-MAP [2] (motif-based) discovers recurring multivariate motifs. UniTFU instead organizes multi-scale local dynamics into a shared TFU dictionary, enabling source-domain units to represent target sequences through temporally aligned composition of overlapping matches.

  • Context-conditioned forecasting. TOTEM [3] (VQ-based) learns discrete time-series tokens, while ProtoTS [4] (prototype-based) learns hierarchical prototypes for interpretable forecasting. UniTFU instead composes matched local dynamics into window-level scene context, jointly conditioning token representations and modulating temporal attention.

  • Adaptive dictionary organization. EE2E-SAE [5] (dictionary learning) learns sparse dictionaries of neural activation features. UniTFU maintains temporal prototypes through confidence-based routing, grouped EMA refinement, novelty-driven birth, and redundancy-based consolidation, adapting both prototype representations and the active vocabulary.

The table below summarizes these distinctions. Table 1 in the revised version presents the full comparison.

MechanismShapeXL-MAPTOTEMProtoTSE2E-SAEUniTFU
Subsequence Matching✓✓✓✗✗✓
Learned Dictionary✓✗✓✓✓✓
Cross-Domain Forecasting✗✗✓✗✗✓
Scene-Conditioned Attention✗✗✗✗✗✓
Prototype Birth & Consolidation✗✗✗✗✗✓

Crucially, VQ is a representation discretization mechanism, whereas UniTFU uses learned prototypes as reusable and composable units that explicitly condition forecasting across domains. Its contribution therefore extends beyond prototype assignment to structured temporal reuse, context-conditioned prediction, and adaptive dictionary maintenance.

Q2. Cross-domain forecasting baselines.

Thank you for your valuable suggestion. We have additionally evaluated three cross-domain time series forecasting methods in Table 2 and 3 of the revised manuscript: UniTime [6], LangTime [7], and LSTPO [8]. Here are the results:

Direct Transfer

Target Domain (MSE/MAE)UniTFUUniTimeLangTimeLSTPO
ETTh10.537 / 0.4900.608 / 0.5440.597 / 0.5390.607 / 0.528
ETTm10.879 / 0.6010.875 / 0.6210.856 / 0.6040.855 / 0.602
Weather0.245 / 0.2870.255 / 0.3050.253 / 0.3010.250 / 0.302
ECL0.579 / 0.5600.694 / 0.6350.686 / 0.6230.682 / 0.619
Solar0.620 / 0.5620.885 / 0.7530.867 / 0.7340.861 / 0.748
Average0.572 / 0.5000.663 / 0.5710.652 / 0.5600.651 / 0.560

Target-Domain Adaptation (10% few-shot)

Target Domain (MSE/MAE)UniTimeLangTimeLSTPOUniTFU
ETTh10.479 / 0.4730.470 / 0.4790.497 / 0.4890.447 / 0.453
ETTm10.486 / 0.450.52 / 0.4560.469 / 0.4370.555 / 0.474
Weather0.185 / 0.2310.182 / 0.2280.185 / 0.2350.177 / 0.223
ECL0.172 / 0.2740.171 / 0.2680.176 / 0.2810.164 / 0.259
Solar0.221 / 0.2750.221 / 0.2640.228 / 0.2840.202 / 0.241
Average0.309 / 0.3410.313 / 0.3390.311 / 0.3450.309 / 0.330

These comparisons are included in Section 4.2 (Tables 2–3), with complete per-transfer results in Appendices B.1–B.2 of the revised manuscript.

AuthorsOpenReview ↗

Respnse to Reviewer FqqN (Part 2)

Q3. Dictionary Adaptation After Large-Scale Pretraining.

Thank you for your suggestion. Using the pretrained UniTFU checkpoint from Section 4.4, we compare Frozen-TFU and Updated-TFU with 100% and 50% retention of pretrained active prototypes. Both variants share the same initial prototypes, frozen backbone, and 10% target training data; only Updated-TFU updates the dictionary. Results are averaged over four forecasting horizons.

Target Domain (MSE / MAE)Frozen (100%)Updated (100%)Frozen (50%)Updated (50%)
ETTh10.356 / 0.3710.349 / 0.3700.372 / 0.4000.363 / 0.393
ECL0.147 / 0.2190.140 / 0.2120.153 / 0.2350.149 / 0.235
Weather0.219 / 0.2350.220 / 0.2340.220 / 0.2450.218 / 0.238
Solar0.186 / 0.2350.182 / 0.2310.200 / 0.2470.195 / 0.247
Average0.227 / 0.2650.223 / 0.2620.236 / 0.2820.231 / 0.278

Dictionary updates improve average performance at both initial sizes, with a larger relative MSE gain at 50% retention (2.27% vs. 1.68%). The updated full dictionary achieves the lowest overall error. These results are further analyzed in Section 4.6 (Fig. 5), which reports horizon-wise MSE under the four dictionary configurations.

Q4. Local Dynamic Semantics.

Thank you for highlighting the distinction between matching a local pattern and improving a forecast. We clarify the labels in Fig. 2 and examine these two questions separately.

(1) Labels in Fig. 2. The names rise, fall, oscillation, switch, and precursor, together with their curves, are manually constructed illustrations, not semantic labels assigned to learned TFUs. UniTFU learns prototypes from multi-scale segments and matches them by representation similarity (Eq. 3), without semantic supervision. This is clarified in the Fig. 2 caption of the revised manuscript.

(2) Cross-domain shape retrieval. We extend the ETTh1 → Weather analysis to four transfer directions. The table reports waveform correlations after linear detrending and normalization for frozen / updated dictionaries. Embedding-based nearest-neighbor retrieval achieves the highest correlation in 7/8 settings, except for frozen Weather → ECL, where frequency-based retrieval performs best. We have summarized these findings in Section 4.6 and provided the detailed cross-domain shape retrieval results in Appendix G of the revised manuscript.

Source → TargetEmbedding NNFrequency NNRaw-stat NNRandom
ETTh1 → Weather0.789 / 0.8250.679 / 0.6940.656 / 0.6640.040 / 0.041
ETTh1 → ETTm10.688 / 0.7260.606 / 0.6090.564 / 0.5650.023 / 0.028
ECL → Solar0.699 / 0.7470.618 / 0.6330.568 / 0.5980.042 / 0.046
Weather → ECL0.585 / 0.6040.607 / 0.5920.533 / 0.5420.035 / 0.040

(3) Prototype usage and forecasting error. Across all 20 source–target transfers (672→96), prototype usage is negatively correlated with window-level forecasting MSE in 17/20 directions. We summarize these correlations alongside changes in matching similarity and forecasting errors.

TargetUsage–Error ρSimilarity GainFrozen MSEUpdated MSE
ETTh1−0.109+0.0180.4820.478
ETTm1−0.119+0.0180.4470.440
Weather−0.154+0.0140.1900.186
ECL−0.185+0.0220.2350.217
Solar−0.257+0.0270.2590.233
Overall−0.162+0.0200.3220.310

Usage–Error ρ is computed across test windows within each transfer and summarized by the median across four sources per target (20 transfers overall). Other columns report the corresponding means.

Mean matching similarity increased in all 20 transfers, while MSE decreased in 18. Across transfers, larger similarity gains were associated with larger MSE reductions (Spearman ρ = 0.721). However, ECL → ETTh1 and ECL → ETTm1 show increased similarity but worsened MSE. Thus, closer prototype matches are associated with predictive improvements across transfers but do not guarantee better forecasts. These findings are summarized in Section 4.6, with complete per-transfer results and metric definitions provided in Appendix G.

AuthorsOpenReview ↗

Respnse to Reviewer FqqN (Part 3)

Q5. Experimental Design for Original Tables 3 and 4.

Thank you for raising this question. The original Tables 3 and 4 (Tables 4 and 5 in the revised version) use the same UniTFU architecture and rolling prediction setting (lookback 672; horizons 96, 192, 336, and 720), but differ in their training settings:

  • Table 4 (Supervised; Appendix C, Table 12): UniTFU is trained from scratch separately on each target dataset's training split (dictionary capacity 128) and evaluated on its held-out test split.
  • Table 5 (Zero-shot; Appendix D, Table 13): A single checkpoint pretrained on UTSD and LOTSA (dictionary capacity 1024) is evaluated on all target test splits without downstream fine-tuning or dictionary updates. All seven evaluation datasets are excluded from pre-training.

These distinct training and evaluation protocols are clarified in Sections 4.3–4.4 and Appendix A.3 of the revised manuscript.

Q6. Ablation Analysis.

Thank you for this concern. We agree that small differences in mean errors alone are insufficient to establish consistent improvements. We therefore conduct paired comparisons across 7 datasets and 4 random seeds (Appendix Table 14). UniTFU outperforms w/o TFU in 26/28 comparisons for both MSE and MAE, with average reductions of 2.96% and 1.98%, respectively. The results are summarized below.

VariantMSE gain ± std (reduction)MSE W/T/LMAE gain ± std (reduction)MAE W/T/L
w/o TFU0.0069 ± 0.0016 (2.96%)26/0/20.0058 ± 0.0024 (1.98%)26/0/2
w/o Sce0.0046 ± 0.0023 (2.00%)21/2/50.0037 ± 0.0026 (1.28%)21/1/6
w/o Att0.0051 ± 0.0017 (2.20%)22/0/60.0041 ± 0.0018 (1.43%)20/3/5
w/o Res0.0046 ± 0.0015 (2.00%)25/0/30.0038 ± 0.0013 (1.32%)23/1/4
w/ VQ0.0041 ± 0.0011 (1.79%)26/1/10.0036 ± 0.0014 (1.26%)24/3/1

Gain = mean(variant − UniTFU); std is computed across four seed-wise gains, each averaged over seven datasets. Percentage reductions are relative to each variant. W/T/L denotes wins/ties/losses at the reported precision.

The gains over w/o TFU remain positive across all four seeds, supporting the consistency of TFU learning improvements.

[1] ShapeX: Shapelet-Driven Post Hoc Explanations for Time Series Classification Models.

[2] Learnable Matrix Profile for Motif Discovery on Multivariate Time Series.

[3] TOTEM: TOkenized Time Series EMbeddings for General Time Series Analysis.

[4] ProtoTS: Learning Hierarchical Prototypes for Explainable Time Series Forecasting.

[5] Identifying Functionally Important Features with End-to-End Sparse Dictionary Learning.

[6] UniTime: A Language-Empowered Unified Model for Cross-Domain Time Series Forecasting.

[7] LangTime: A Language-Guided Unified Model for Time Series Forecasting with Proximal Policy Optimization.

[8] Preference Guided Meta-Learning for Cross Domain Time Series Forecasting.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

We sincerely thank all reviewers for their insightful and constructive feedback, which has greatly helped us improve the quality of our paper.

In this work, we propose UniTFU, a time series foundation model that enables cross-domain forecasting through reusable and composable local temporal dynamics. By organizing recurring multi-scale patterns into adaptive Time Series Foundation Units (TFUs), UniTFU explicitly models how source-domain dynamics are reused, composed, and updated in target domains, complementing existing approaches based primarily on implicit shared representations.

We appreciate the reviewers' recognition of "the value of reusable temporal prototypes for forecasting" (Reviewers UQQH, 5zWt), "the clear motivation and intuitive design of the TFU mechanism" (Reviewers FqqN, 5zWt), and "UniTFU's strong cross-domain forecasting performance" (Reviewers UQQH, 5zWt).

The reviewers raised insightful and constructive concerns. Over the past seven days, we have carefully addressed these concerns by providing additional experimental evidence, clarifications, and requested results. The major revisions are summarized below:

  • Technical Contributions and Positioning (FqqN, UQQH): We clarified UniTFU's distinctions from VQ, shapelet/motif discovery, prototype learning, and dictionary-based methods, emphasizing explicit local dynamic reuse, composition, and adaptation.
  • Methodological and Experimental Clarifications (FqqN, 5zWt): We clarified the training and evaluation protocols, prototype initialization, matching and dictionary maintenance, and the illustrative nature of local dynamic labels.
  • Comprehensive Experimental Evaluations (FqqN, UQQH, 5zWt): We conducted 120+ additional experiments, including additional cross-domain baselines, few-shot adaptation, dictionary reuse and updating, local shape and prototype usage analyses, and inference efficiency comparisons.

All revisions are highlighted in blue in the revised manuscript. We sincerely appreciate the reviewers' valuable feedback, which has greatly improved our work. We hope our responses adequately address their concerns and welcome any further questions or discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.