acceptodds
Under review as a conference paper at ICLR 2027

TSRouter: Training-Free Capability Landscape Discovery and Routing for Growing Time-Series Foundation Model Zoos

Abstract

Time-series foundation models (TSFMs) offer zero-shot forecasts, but their strengths vary across forecasting tasks and change as new models appear. This creates a two-sided selection problem: forwarding every candidate can improve task-specific selection but makes inference expensive as the model zoo grows, whereas supervised selectors avoid online forwarding but require candidate-dependent label updates and retraining after model arrivals. A practical router must retain competitive quality while keeping both inference and adaptation costs bounded. How can a router reuse capability evidence without training a selector for each evolving model zoo? We hypothesize that model preference is locally transferable: forecasting windows with similar temporal structure tend to favor similar models. We introduce TSRouter, a training-free framework for discovering, retrieving, and updating a capability landscape. Representative, source-stratified probe anchors identify model-specific capability regions. A shared representation lets unseen task windows retrieve this evidence, and window/channel rank fusion turns local preferences into a task-level choice. New models are incorporated into the same capability space using existing anchors, preserving previously discovered evidence. On 97 forecasting configurations from 23 datasets and a 20-model zoo, TSRouter achieves the best mean absolute scaled error among the evaluated deployable selectors. Holdout diagnostics and ablations support local capability transfer and evidence aggregation, while release-ordered evaluation demonstrates adaptation to changing candidates. A runtime-aware variant provides a complementary accuracy–latency trade-off.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Foundation Models, Model Routing, Zero-Shot Learning, Capability Discovery, Time Series Forecasting

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

This paper proposes TSRouter, a training-free framework for selecting time series foundation models from an evolving model zoo. TSRouter consists of three stages. PROFILE constructs model-specific capability regions using representative probe anchors and measured forecasting losses. ROUTE retrieves local capability evidence through a fixed Fourier-based representation and aggregates window/channel rankings to select a model. INSERT incorporates newly released models by evaluating them on existing anchors without retraining a selector. Experiments on 97 forecasting configurations from 23 datasets and 20 TSFMs show that TSRouter achieves a MASE of 1.508, outperforming the strongest evaluated baseline.

Critical strengths
  1. The separation into PROFILE, ROUTE, and INSERT is appealing. Reusing a fixed capability representation avoids retraining a supervised selector when the candidate pool changes.
  2. TSRouter-main improves MASE from 1.558 to 1.508 over the strongest evaluated deployable baseline, while achieving substantially lower routing overhead than online candidate probing.
Critical weaknesses
  1. TSRouter relies on 3000 anchors to characterize model capabilities and guide routing decisions. However, it remains unclear whether performance on these anchors reliably reflects model capabilities across diverse and unseen tasks.
  2. The proposed capability region construction relies on hard assignments of anchors to the best models. It is unclear whether this design offers advantages over other simpler designs, such as KNN-based aggregation of model losses on nearby anchors.
Questions for authors
  1. Why can anchors selected based on coverage in the time-series feature space reliably characterize the relative capabilities of different TSFMs? Have the authors compared alternative anchor selection strategies? Additionally, why does increasing the number of anchors from 3,000 to 5,000 lead to worse routing performance, as shown in Table 3?
  2. Could the authors compare TSRouter with simpler alternatives, such as kNN-based loss aggregation, to justify the proposed capability-region design?
1 comment
AuthorsOpenReview ↗

Anchor Transfer, Budget Effects, and Direct-Loss Routing

Thank you for the constructive questions about anchor coverage and capability-region design.

Q1a — Why can anchors describe TSFM capability on unseen tasks?

Anchors and unseen task windows occupy the same representation space. Anchors attach measured model preferences to locations in this space, turning abstract task/model differences into measurable distances to model-specific support. Appendix E, Table 9 supports this transfer mechanism: identity features → Fourier with random anchors → Fourier with clustered anchors raises held-out anchor winner recovery from 0.076 → 0.396 → 0.446, and task recovery from 0.206 → 0.247 → 0.268. Thus, the representation and representative coverage jointly improve local preference transfer.


Q1b — Were other anchor-selection strategies compared?

Yes—Table 4 explicitly ablates clustered versus random anchor selection. At the same 3,000-anchor budget and source allocation, random → clustered improves MASE 1.5172 → 1.5076 and CRPS 0.2385 → 0.2335, supporting representative sampling beyond anchor count alone. Table 4 also separates source-removal and budget controls, showing how coverage and cost affect the operating point.


Q1c — Why do 5,000 anchors perform worse?

Increasing the anchor budget rebuilds the supporting neighborhoods and can change model rankings; here, a few costly model switches outweigh gains elsewhere. The task-level diagnosis in Appendix H.4 (new analysis) shows that, after increasing from 3,000 to 5,000 anchors, 59 of 97 tasks retain the same choice; among the 38 changed choices, 15 improve and 23 worsen. One COVID task accounts for 77% of the net MASE increase, so the higher average does not reflect uniformly worse routing. The MASE-difference 95% interval [−0.00252, 0.05372] includes zero. This illustrates a trade-off between anchor coverage, routing accuracy, and construction cost: a larger support set does not necessarily yield more accurate model selection.


Q2 — Does region retrieval improve over kNN-loss?

Thank you for suggesting direct loss aggregation. We add a TSRouter-kNN variant that replaces region retrieval with nearby-loss aggregation, using the same 3,000 anchors, historical CRPS matrix, features, queries, and fusion (Appendix H.2, Tables 11–13 (new experiment/analysis)). TSRouter-main improves MASE, while kNN has better CRPS point estimates. At the fixed k=10 with inverse-distance weighting:

RuleMASE ↓CRPS ↓sMAPE ↓
TSRouter-main1.507590.233500.43613
TSRouter-kNN1.565920.222080.43942

The MASE advantage survives multiplicity correction.

This is a difference in how anchors are used within the retrieval step: TSRouter-main retrieves model-specific winning support, while TSRouter-kNN aggregates losses over common neighbors. Both preserve the framework's core mechanism—reusing fixed anchor evidence for training-free routing and incorporating new models without retraining the selector. The comparison therefore clarifies a retrieval-rule trade-off within this framework.

Reviewer_Gb3WOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TSRouter, a training-free routing method for a growing pool of time-series foundation models. The method first selects representative anchors from historical time-series data and evaluates candidate models on these anchors. It then constructs Capability Regions based on which model performs best on each anchor. For a new task, TSRouter extracts time-series features using a fixed encoder, retrieves the corresponding Capability Regions, and selects a suitable model through multi-window and multi-channel rank aggregation. When a new model arrives, only the new model needs to be evaluated on the existing anchors to update the Capability Regions, without retraining the router. Experiments on 20 time-series foundation models and 97 forecasting configurations demonstrate the effectiveness of the method in model selection and incremental adaptation.

Critical strengths
  1. Simple and practical method. TSRouter does not require training an additional router and supports incremental integration of new models, making it practically useful.

  2. Comprehensive experiments. The paper evaluates multiple time-series models and datasets, with extensive baseline comparisons, ablation studies, and efficiency analyses that support the effectiveness of the proposed method.

Critical weaknesses
  1. Limited novelty. Previous works, such as ZooCast (2025) and TimeRouter (2026), have explored similar ideas for time-series model routing. However, this paper does not mention these related works or clarify how its core idea differs from them, which raises concerns about its novelty.

  2. Insufficient justification for Capability Regions. The paper already maintains a complete Capability Map (a Model × Anchor prediction error matrix). For a new task, the historical prediction errors could be directly weighted according to its similarity to each anchor, and the model with the lowest estimated error could be selected. The paper does not sufficiently explain why Capability Regions are additionally needed or provide a direct comparison with this simpler approach.

  3. Unclear meaning of Capability Regions. A Capability Region is essentially a set of anchors on which a model performs best. Although the experiments demonstrate that this grouping helps model selection, the paper does not clearly explain how these regions correspond to specific model capabilities.

  4. Incomplete computational cost analysis. Although the paper evaluates routing and model insertion costs, it lacks a comprehensive comparison of the initial Capability Landscape construction cost and long-term cumulative overhead. Therefore, its overall efficiency advantage over trained routers is not fully established.

Questions for authors
  1. ZooCast (2025) and TimeRouter (2026) have explored similar ideas for time-series model routing. Could the authors clarify the key differences between TSRouter and these works, and explain the main novelty of the proposed method?

  2. Why not directly use the complete Model × Anchor prediction error matrix stored in the Capability Map for model selection? For example, the historical prediction errors could be weighted based on the similarity between a new task and each anchor, and the model with the lowest estimated error could be selected. Could the authors provide a comparison using the same anchors and feature encoder to demonstrate the necessity of Capability Regions?

  3. Capability Regions are defined only by the anchors on which each model performs best. What exactly does "Capability" refer to in this context? How does this grouping reflect actual forecasting capabilities rather than simply recording which model performs best on historical samples?

  4. Could the authors report the initial cost of constructing the Capability Landscape and compare the cumulative computational costs of TSRouter and trained routers under the same experimental conditions to better support the claimed efficiency advantages?

2 comments
AuthorsOpenReview ↗

Related Routers, Capability Regions, and Cumulative Costs

We apologize for the delay in our response. We sought guidance from the Program Chairs on the anonymity and arXiv-related issues associated with your Q1. We have now received clear guidance and privately conveyed the relevant context to the Area Chair. Thank you for your careful review; we respond to your comments below.

Thank you for recognizing the practical value of incorporating new models without router retraining.


Q1 — What differs from ZooCast and TimeRouter?

Compared with ZooCast, TSRouter removes representation training, preserves local model support, and maintains a growing zoo through explicit evidence updates. ZooCast already uses offline evaluations for model matching and incorporates newcomers while reusing old scores. The differences are: training a shared encoder → fixed Fourier features; averaging each model's advantage subset into one embedding → retaining multiple winning anchors; appending a model summary → updating anchor ownership and the retrieval index from the newcomer's losses. Both reuse historical evaluations; TSRouter additionally measures release-ordered adaptation and separates initial, insertion, and cumulative costs (Appendix H.3, Tables 14–15 (new experiment/analysis)).

TimeRouter cites ZooCast as related embedding-based Top-K selection, but uses a different mechanism: supervised routing heads with candidate CV scores and forecast outputs. Thus TimeRouter selects using candidate computations on the current task; TSRouter selects before candidate forwarding from observed windows and cached evidence. Model arrival entails routing-head retraining in TimeRouter versus profile/index updates in TSRouter. TimeRouter's fallback can combine multiple forecasts, whereas TSRouter executes one selected model: their accuracy and latency therefore reflect different inference budgets, not interchangeable single-model routing results. We have added both citations and these distinctions in Section 2.


Q2 — Why use regions rather than aggregate losses?

Thank you for suggesting direct loss aggregation. We add a TSRouter-kNN variant that replaces region retrieval with nearby-loss aggregation, using the same 3,000 anchors, historical CRPS matrix, features, queries, and fusion (Appendix H.2, Tables 11–13 (new experiment/analysis)). On GIFT-Eval, TSRouter-main achieves lower MASE than TSRouter-kNN. At the fixed k=10 with inverse-distance weighting:

RuleMASE ↓CRPS ↓sMAPE ↓
TSRouter-main1.507590.233500.43613
TSRouter-kNN1.565920.222080.43942

The MASE advantage survives multiplicity correction.

This is a difference in how anchors are used within the retrieval step: TSRouter-main retrieves model-specific winning support, while TSRouter-kNN aggregates losses over common neighbors. Both preserve the framework's core mechanism—reusing fixed anchor evidence for training-free routing and incorporating new models without retraining the selector. The comparison supports model-specific winning support for the main MASE objective within this shared framework.


Q3 — What does a capability region represent?

Capability means a model's relative measured advantage on nearby profiling windows. Direct loss aggregation uses the same notion of capability: nearby anchors' measured losses estimate which model is locally preferable. TSRouter-main expresses this evidence through model-specific winning support. Appendix E, Table 9 shows that random → clustered Fourier anchors improve held-out winner recovery 0.396 → 0.446 and task recovery 0.247 → 0.268, supporting a useful link between representation locality and model preference.


Q4 — Does the cost advantage include initial construction?

The cost advantage is in selector-state maintenance without retraining; total offline and end-to-end costs remain workload-dependent. Appendix H.3, Tables 14–15 (new experiment/analysis) separates these components. Selector-state initialization and updates total 2.33 s for TSRouter versus 249.89 / 781.19 / 51,862.05 s for AutoForecast / AutoXPCR / SimpleTS, excluding model profiling and downstream inference; SimpleTS also lacks a recorded encoder-pretraining cost. This quantifies the low cost of maintaining the region index as models arrive.

Including the recorded research pipeline and optional pool diagnostics, TSRouter / AutoXPCR require 506.81 / 485.67 s initially and 9,323.96 / 9,329.19 s for subsequent arrivals, making their offline totals comparable. We therefore locate the efficiency benefit in the selector-maintenance component and absence of retraining, while retaining profiling and inference in the full cost accounting.

AuthorsOpenReview ↗

Follow-up on Q1–Q4

Thank you again for your constructive review. We have responded to Q1–Q4 with related-work clarifications, the same-evidence kNN comparison, capability diagnostics, and initial/cumulative cost accounting. Our updated Summary of new experiments and the revised PDF also incorporates the transfer analysis from our follow-up to Reviewer bcWz, Clarification on transfer beyond the profiling anchors. Could you let us know whether these additions address your concerns, or whether any specific issue still requires clarification before the discussion closes?

Reviewer_XCw5OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper proposes TSRouter that selects among a growing pool of TSFMs. It measures model losses on fixed anchors and assigns each anchor to its best model. Fixed Fourier features link observed task windows with these capability regions. Here, Rank aggregation across windows and channels selects one model, and new models require evaluation on the existing anchors and updates to region assignments, without any retraining. In GIFT-Eval, the authors report lower MASE than the baselines, showing improved performance.

Critical strengths
  1. INSERT evaluates only the new model and reuses stored losses for existing models. The representation remains fixed, and the selector requires no retraining making them highly useful in real world use cases.

  2. TSRouter-main has the lowest mean model rank among the tested deployable selectors. Tables 3 and 8 provide additional evidence for the value of Fourier features and representative anchors. These results support the usefulness of the proposed method within the tested setting.

Critical weaknesses
  1. While the paper has a comprehensive evaluation, it does not compare against [1], which also uses a similar idea of capability evidence in a shared representation space and adds models with repeated evaluation of existing models

  2. The paper does not report confidence intervals or statistical significance tests for the main comparisons. Thus, the reliability of the reported performance gains remains unclear.

  3. While GIFT-Eval is definitely comprehensive, it is important to also evaluate the performance gain of TSRouter on other benchmarks like fev-bench.

[1] Shi, Hao-Nan, Ting-Ji Huang, Lu Han, De-Chuan Zhan, and Han-Jia Ye. "One-embedding-fits-all: Efficient zero-shot time series forecasting by a model zoo." arXiv preprint arXiv:2509.04208 (2025).

Questions for authors

Can you please compare TSRouter with that of [1] as discussed above?

3 comments
AuthorsOpenReview ↗

ZooCast Clarification, Statistical Reliability, and External Evaluation

We apologize for the delay in our response. We sought guidance from the Program Chairs on the anonymity and arXiv-related issues associated with your Q1. We have now received clear guidance and privately conveyed the relevant context to the Area Chair. Thank you for your careful review; we respond to your comments below.

We have added uncertainty estimates and an external-benchmark evaluation.


Q1 — How does TSRouter relate to ZooCast?

Compared with ZooCast, TSRouter removes representation training, preserves local model support, and maintains a growing zoo through explicit evidence updates. ZooCast already uses offline evaluations for model matching and incorporates newcomers while reusing old scores. The differences are: training a shared encoder → fixed Fourier features; averaging each model's advantage subset into one embedding → retaining multiple winning anchors; appending a model summary → updating anchor ownership and the retrieval index from the newcomer's losses. Both reuse historical evaluations; TSRouter additionally measures release-ordered adaptation and separates initial, insertion, and cumulative costs (Appendix H.3, Tables 14–15 (new experiment/analysis)).


Q2 — Are the main comparisons statistically supported?

Yes—we have added uncertainty estimates for all main-table deployable comparisons in Appendix H.1, Table 10 (new experiment/analysis), which support the MASE advantage over several principal baselines. Against AutoXPCR, MASE decreases by 0.05007; the multiplicity-adjusted interval is [−0.13769, −0.00470], entirely below zero. Adjusted intervals also favor TSRouter over AutoForecast, SimpleTS, Profile-probe, and Random.


Q3 — Does TSRouter generalize beyond GIFT-Eval?

We added full fev-bench results to the main text (Table 3, new experiment); TSRouter-main has the lowest mean MASE and SQL among the nine compared selectors. The evaluation covers 100 tasks and 1,308 official forecast origins with the same 20-model zoo and original 3,000 anchors, using target history only. Selected comparisons are:

MethodMASE ↓SQL ↓Win rate ↑Skill ↑
TSRouter-main1.619801.352390.6460.294
SimpleTS1.627901.362710.5480.292
AutoForecast1.769791.509390.3710.253
AutoXPCR1.719931.436480.5020.276
Task-probe1.664451.406120.5640.294

Win rate is recomputed within the full nine-method panel. Appendix C.3 (new protocol) gives the settings; Appendix I.1, Tables 16–18 (new analysis) retains all methods, source-grouped intervals and paired comparisons. These are supplementary results; the paired MASE interval against SimpleTS includes zero.

Thank you very much for addressing my primary concerns, and I particularly appreciate adding the fev-bench results in the main manuscript. I'm leaning towards increasing the score, as the added uncertainty bounds as well as benchmark results substantially increases my confidence.

Thank you for your careful reassessment. We are glad that the additional uncertainty analysis and full fev-bench results addressed your primary concerns and strengthened your confidence in the work. We appreciate your constructive feedback, which helped us improve the manuscript.

Comments on the paper

3 comments
Area_Chair_7n5BOpenReview ↗

please engage

Dear reviewers,

Please join the discussion with the authors.

Best, AC

AuthorsOpenReview ↗

Summary of new experiments and the revised PDF

We thank the reviewers and Area Chair for their constructive feedback. The revised PDF has been uploaded to OpenReview. Added or replaced text is marked blue. New experiments address every question.

The new evidence supports model-specific capability retrieval on GIFT-Eval, quantifies its selector-maintenance advantage, and extends the evaluation to full fev-bench. On the supplementary 100-task fev evaluation, TSRouter-main has the lowest mean MASE and SQL among the nine compared selectors; the table summarizes the main additions.

AdditionReviewer concern addressedMain conclusion and evidenceLocation
Related-work comparisonXCw5 Q1; Gb3W Q1: overlap with ZooCast and distinction from TimeRouterCompared with ZooCast, TSRouter removes representation training through fixed features and retains multiple winning anchors. Unlike TimeRouter, it selects before candidate forwarding and updates evidence without selector retraining.Section 2, new citations and comparison paragraph
Same-evidence TSRouter-kNN comparisonGb3W Q2; bcWz Q2: why construct regions rather than directly aggregate nearby losses?With identical evidence, TSRouter-main reduces MASE from 1.56592 to 1.50759 versus TSRouter-kNN. The advantage survives multiplicity correction.Appendix H.2, Tables 11–13
External-benchmark evaluationXCw5 Q3; bcWz follow-up: transfer beyond GIFT-EvalOn full fev-bench, TSRouter-main has the lowest mean MASE/SQL: 1.61980/1.35239. After excluding 21 GIFT-Eval-source tasks, it remains lowest on the other 79 tasks: 1.81013/1.51623, using the original 3,000 anchors.Main text, Table 3; Appendix C.3 and I.1; follow-up: Clarification on transfer beyond the profiling anchors
Main-result uncertaintyXCw5 Q2: missing confidence intervals or significance evidenceAgainst AutoXPCR, the MASE difference is −0.05007, with a multiplicity-adjusted interval [−0.13769, −0.00470]. This supports the main MASE advantage.Appendix H.1, Table 10
Initial and cumulative cost accountingGb3W Q4: does the efficiency claim include initialization and model arrivals?Selector initialization/updates take 2.33 s for TSRouter versus 781.19 s for AutoXPCR. Including profiling, their full offline totals are comparable.Appendix H.3, Tables 14–15

Observed E2E P95 is 75.309 s for TSRouter-main and 11.489 s for TSRouter-fast; fast trades MASE 1.61980 → 1.70560 for lower latency. Selection frequencies and measured latency are in Appendix I.2, Table 19.

Taken together, these analyses show how reusable anchor evidence supports model selection and growing-zoo maintenance: the retrieval comparison tests how that evidence is used, the uncertainty analysis supports the main MASE gains, and the cost accounting identifies the low-overhead update component.

AuthorsOpenReview ↗

Clarification on transfer beyond the profiling anchors

Thank you for your prompt follow-up, for acknowledging the kNN comparison and anchor-budget analysis, and for specifying the remaining concern. This question was indirectly addressed in our response to Reviewer XCw5 Q3: the added fev-bench evaluation provides direct evidence that the original anchor bank remains useful beyond the GIFT-Eval workload. In their follow-up, Reviewer XCw5 noted that the added uncertainty analysis and benchmark results substantially strengthened their confidence and updated their assessment accordingly.

The full fev-bench evaluation covers 100 tasks and 1,308 official forecast origins, using the original 3,000 anchors and the same 20-model zoo. Selected comparisons from Table 3, matching those in our response to XCw5, are:

MethodMASE ↓SQL ↓Win rate ↑Skill ↑
TSRouter-main1.619801.352390.6460.294
SimpleTS1.627901.362710.5480.292
AutoForecast1.769791.509390.3710.253
AutoXPCR1.719931.436480.5020.276
Task-probe1.664451.406120.5640.294

To examine whether the result is driven by overlap with GIFT-Eval, we additionally summarized the stored final scores after excluding the 21 tasks from GIFT-Eval sources, following the fev-bench paper's source classification. On the remaining 79 tasks from other sources:

MethodMASE ↓SQL ↓Win rate ↑Skill ↑
TSRouter-main1.810131.516230.6270.299
SimpleTS1.815881.525880.5410.302
AutoForecast1.991611.705050.3740.253
AutoXPCR1.935211.620750.4830.278
Task-probe1.862051.577140.5740.303

Win rate uses all nine selectors; Skill is relative to Seasonal Naive. All four metrics are recomputed on the 79-task subset. The benefit therefore extends beyond GIFT-Eval-source tasks using the original 3,000 anchors, without adding fev histories or using its evaluation targets for profiling or selector fitting.

Appendix E, Table 9 provides complementary evidence: identity → Fourier with random anchors → Fourier with clustered anchors improves profile-holdout winner recovery from 0.076 → 0.396 → 0.446. Together with the fev results, this supports our hypothesis that anchors and task windows share a representation space that preserves transferable model preferences, allowing a compact anchor bank to support selection across heterogeneous evaluated tasks. The subgroup analysis is descriptive.

We hope this additional source-level analysis clarifies the transfer evidence and addresses your remaining concern. We would welcome any specific points that still require clarification.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.