acceptodds
Under review as a conference paper at ICLR 2027

One Model, Three Views: Input-Condition Integration for Probabilistic Time Series Foundation Models

Abstract

Recent time series foundation models (TSFMs) can condition on known future covariates, but this conditioning helps some targets and samples while hurting others, and which ones cannot be told in advance. By varying only its inputs, a single frozen TSFM yields three views of the same forecast: from the target's history alone (target-only), with past covariates added (multivariate-history), and with known future covariate values added (known-future). We show that replaying how forecasts change across these views on completed historical events reveals whether richer conditioning was beneficial, harmful, or insufficient. Based on this finding, we propose input-condition integration, which corrects the known-future forecast using the other two views as references rather than as candidates to select or average. The correction is learned from the observed input history alone and applied only when it proves reliable there, leaving the TSFM unchanged. Across three TSFM backbones, input-condition integration improves probabilistic accuracy over the known-future forecast on real-world benchmarks and better recovers covariate-driven responses in controlled synthetic tasks. These results suggest that the value of additional inputs lies not only in the information they supply, but in how the model responds to them.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Foundation Models, Probabilistic Forecasting, Covariate Conditioning, Input-Condition Integration, Forecast Calibration

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_tsiZOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes Input-Condition Integration (ICI) for frozen covariate-aware time series foundation models. It replays historical events under three input conditions, fits event-aligned quantile corrections, and uses chronological validation to select native or corrected forecasts. Evaluation covers controlled synthetic relationships and real-world benchmarks with three backbones. However, the principal real-world results are explicitly labeled as illustrative placeholders and reused in the main analysis, substantially undermining the credibility and significance of the empirical claims and raising serious concerns about the integrity of the experimental reporting.

Critical strengths
  1. Separating historical forecast correction from conditioning-response diagnostics is a sensible design choice that allows correction without requiring a strong backbone response.
  2. The synthetic setup includes null and relationship-shift conditions, making the method's dependence on transferable historical relationships explicit.
Critical weaknesses
  1. The central real-world evidence is presented as both experimental results and drafting placeholders. Table 2 repeats all 18 corresponding values from Table 7, whose caption explicitly designates the numbers as placeholders; Table 8 does likewise for the earlier cohort. Section 4.3 and Appendix E.1 interpret these results as empirical findings, and Table 9 repeats Table 7's aggregate values. No replacement set of measured results is identified in the manuscript. This is a serious reporting-integrity concern that prevents reliance on the real-world conclusions. Treating the captions as obsolete drafting text would not adequately resolve the contradiction without substantiating the reported results.
  2. The method is not compared with the cited TFMAdapter (CIKM 2025), and no real-world event-only ablation isolates the value of three-view guidance. Most improvement in the synthetic ablation comes from the base correction, leaving the added mechanism's practical contribution unclear.
Questions for authors
  1. Are the values in Tables 2, 7, 8, and 9 measured results or placeholders? If the captions are obsolete, what existing experimental evidence substantiates the reported values and their interpretation?
  2. Under the same available history and covariates, how does ICI compare with TFMAdapter and its own event-only correction on real-world tasks?
2 comments

Thank you for your careful review. We appreciate your observation that separating historical forecast correction from conditioning-response diagnostics allows the method to apply a correction even when the backbone responds weakly. We also appreciate your point that the null and relationship-shift experiments make the method’s reliance on transferable historical relationships explicit.

During drafting, we generated Tables 7 and 8 in Appendix E with temporary values, and the table-generation process automatically added a note to each caption identifying the values as placeholders. We replaced those values with results from completed experiments before submission. We overlooked the notes when updating the tables. Table 2 reports selected task-level results from Table 7.

We also found errors in some of the reported macro averages in Tables 7 and 8. The affected averages from Table 7 also appear in the "All cases" column of Table 9. We apologize for leaving the placeholder notes in the captions and for the errors in the reported averages.

Thank you for suggesting real-world comparisons with TFMAdapter and with our event-only correction under the same available history and covariates. As you point out, the base correction accounts for most of the gains in the synthetic ablation. Your suggestions give us clear priorities for a future revision, particularly for assessing what three-view guidance adds on real-world tasks.

Dear authors, the rebuttal phase is nearing its end. If there is no further information to provide, I will maintain my current ratings and recommend rejection.

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

The paper proposes Input-Condition Integration (ICI) for adapting a frozen covariate-aware forecaster. It replays completed historical events under three input conditions: target history alone, multivariate history, and known future covariates. An event-aligned quantile correction is first fitted to historical errors. Differences between the three forecasts then guide a further residual correction when historical outcomes support their usefulness. Chronological validation selects between the native forecast, the base correction, and the guided correction. Evaluation is presented on controlled synthetic tasks and selected fev-bench tasks with three backbones.

Critical strengths
  1. The method makes a useful distinction between a model responding to additional covariates and that response actually improving its forecast. Historical outcomes provide a way to test the latter. The base correction also remains available when the backbone barely responds to an informative event, which addresses a limitation of methods that can only combine existing forecasts.

  2. The chronological validation protocol is carefully specified. Earlier replay targets must be fully observed before a later validation origin, and validation outcomes are excluded from fitting, descriptor scaling, and residual retrieval. The null and relationship-shift experiments are also well chosen: they test whether adaptation is selective and expose the assumption that historical event responses remain relevant.

Critical weaknesses
  1. The status of the real-world experimental results is unresolved in the submitted manuscript. On page 21, the caption of Table 7 explicitly states, “Numbers are illustrative placeholders for manuscript drafting,” and Table 8 carries a similar statement. Several Table 7 values also appear in the main-text Table 2. This could be an editing error, but until it is clarified, I cannot treat these tables as verified empirical support for the real-world claims.
  2. The incremental value of the three-view mechanism is less established than that of historical residual correction. Table 3 reports an average synthetic improvement of 15.3% for the event-only correction and 16.5% for full ICI. The paper acknowledges this difference, but does not provide a corresponding real-world decomposition or a runtime comparison. Consequently, it remains unclear whether the extra replay queries and guidance machinery offer enough benefit over the simpler event-only method.
Questions for authors
  1. Are Tables 7 and 8 actual experimental results or illustrative values? Which other tables depend on these numbers? Please provide the finalized results and corresponding evaluation records.
  2. On verified real-world runs, how much does full ICI improve over the event-only correction using identical histories and validation folds? Please report paired uncertainty and the frequency with which guidance is selected.
  3. Since retrieval can access 2,048 observations while the native backbone receives at most 512, how does ICI compare with a native forecast using the longer history where supported? What is the additional inference cost of E and full ICI?
1 comment

Thank you for the time and care you put into reviewing our paper. We appreciate your comments on the chronological validation protocol and the null and relationship-shift experiments. We also appreciate your point that a forecast may respond to additional covariates without improving, and that event-only correction can still be applied when the backbone responds weakly.

During drafting, the table-generation process automatically added placeholder notes to the captions of Tables 7 and 8 in Appendix E. We replaced the temporary values with results from completed experiments before submission and overlooked the notes. Table 2 reports selected task-level results from Table 7. We also found errors in some of the macro averages in Tables 7 and 8. The affected averages from Table 7 appear in the "All cases" column of Table 9. We apologize for the misleading captions and incorrect averages.

Your review gives us a clear direction for a future revision. Comparing event-only correction with full ICI on real-world tasks using identical histories and validation folds would help assess the benefit of adding guidance. We appreciate your suggestions to report paired uncertainty estimates, guidance selection rates, and inference costs for each method, and to include a native forecast with the longer history where supported.

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

This paper proposes Input-Condition Integration (ICI), a test-time adaptation framework for probabilistic time-series foundation models. ICI examines three forecasting views from a frozen backbone: target-only, history-conditioned, and known-future-conditioned predictions. By replaying completed historical events, it evaluates whether additional covariate information produces beneficial, harmful, or insufficient forecasting responses. The method first learns an event-aligned quantile correction from historical residuals, then uses validated input-condition responses to guide additional residual adjustments. Chronological validation determines whether to retain the native forecast or apply a correction. Experiments with three TSFM backbones suggest improvements in controlled delayed-response settings and report smaller gains on real-world forecasting benchmarks.

Critical strengths
  1. Well-motivated problem. The paper identifies an important limitation of covariate-aware TSFMs: access to known-future covariates does not guarantee their effective utilization. The controlled analysis of delayed covariate responses provides a clear motivation for instance-specific adaptation.

  2. Interpretable input-condition diagnosis. Comparing target-only, history-conditioned, and future-conditioned predictions offers a structured way to examine how a frozen backbone responds to additional information. The distinction between response magnitude and historically validated forecasting utility is particularly meaningful.

  3. Careful temporal evaluation design. The historical replay procedure explicitly restricts fitting and validation to completed past events. The chronological selection mechanism also provides a reasonable safeguard against applying unsupported corrections.

  4. Useful synthetic experiments and ablations. Evaluations across three backbones and multiple synthetic regimes demonstrate improvements under transferable event-response relationships. The null and relationship-shift controls expose important failure conditions, while Table 3 separates event-aligned correction from additional input-condition guidance.

Critical weaknesses
  1. Unverified real-world experimental results. Appendix E.1 explicitly labels the numerical results in Tables 7 and 8 as "illustrative placeholders." Moreover, the real-world results summarized in Table 2 exactly match the corresponding entries in Table 7. Consequently, the reported real-world improvements cannot be treated as verified experimental evidence in the current submission. This is a substantial concern regarding the validity and reproducibility of the empirical claims.

  2. Limited incremental contribution of the three-view mechanism. Table 3 shows that event-aligned residual correction alone achieves an average NQS improvement of 15.3%15.3\%, while the complete ICI framework reaches 16.5%16.5\%. Thus, most of the improvement comes from historical residual fitting, with three-view guidance contributing only 1.21.2 additional percentage points. Stronger comparisons with conventional event-conditioned residual regression, distributed-lag models, and local quantile recalibration are needed to establish the distinctive value of the proposed mechanism.

  3. Dependence on stable historical relationships. Table 4 reports performance degradation of 8.4%8.4\%--10.0%10.0\% under relationship shift. This failure is particularly important because the method relies on historical event-response transferability. The chronological validation mechanism does not guarantee protection against changes occurring after the observed history. More realistic nonstationary evaluations and mechanisms for detecting unreliable historical transfer would strengthen the robustness claims.

  4. Limited real-world evaluation transparency and coverage. The method requires sufficiently many comparable completed events, restricting its applicability to selected forecasting instances. The reported coverage statistics are themselves contained in placeholder tables. Furthermore, Appendix A.1 states that the issuance vintage of electricity-price covariates is not independently verified and benchmark-provided future weather variables are accepted as available. These limitations complicate assessment of the real-world evaluation protocol.

  5. Missing efficiency analysis and broader probabilistic validation. ICI requires multiple backbone evaluations for historical event replay, three input conditions, residual fitting, and chronological selection. The inference-time and computational overhead are not quantitatively evaluated. Additionally, the experiments emphasize quantile loss without sufficiently examining predictive interval coverage, calibration stability, or the statistical significance of the incremental three-view gains.

Questions for authors
  1. Experimental validity: Why are Tables 7 and 8 explicitly labeled as illustrative placeholders? Are the values in Table 2 obtained from completed experiments or inherited from these placeholders? Could the authors provide verified results, evaluation artifacts, and uncertainty estimates for all real-world benchmarks?

  2. Incremental contribution: Given that the event-only correction accounts for most of the performance improvement, can the authors compare ICI against standard event-conditioned residual regression, distributed-lag regression, and local quantile recalibration under identical historical information and computational budgets?

  3. Statistical significance: Is the additional 1.21.2-percentage-point average NQS improvement from three-view guidance statistically significant across independent synthetic seeds? How frequently is guided correction selected over the event-only baseline in real-world tasks?

  4. Distribution shift: Can the authors evaluate gradual changes, abrupt reversals, and changes in response delay or magnitude? Is it possible to detect when historically validated corrections are no longer transferable before applying them to new events?

  5. Practical deployment: What are the average numbers of historical replay queries, inference latency, memory requirements, and computational costs per forecast? How does performance vary with the number of available historical analogs, and how is the availability of future covariates verified?

1 comment

Thank you for the time you devoted to both the motivation and the evaluation. We appreciate your recognition of delayed covariate responses as a reason for instance-specific adaptation and of the distinction between the magnitude of a forecast response and its forecasting utility. We are also grateful for your comments on the chronological replay design and on what the null and relationship-shift experiments reveal about the limits of correction.

Before submission, we replaced the temporary values in Tables 7 and 8 of Appendix E with results from completed experiments. The table-generation process had automatically added placeholder notes to both captions during drafting, and we inadvertently left them in place when updating the tables. Table 2 summarizes selected task-level results from Table 7. We also reported some macro averages incorrectly in Tables 7 and 8, and the affected Table 7 averages appear in the "All cases" column of Table 9. We apologize for the caption oversight and the incorrect averages.

Your suggestions give us several priorities for a future revision. Comparisons with event-conditioned residual regression, distributed-lag models, and local quantile recalibration under matched information and compute budgets would help clarify what three-view guidance adds. We also appreciate your suggestions to examine how corrections hold up when relationships change, document covariate availability and adaptation coverage, and report guidance selection rates, inference cost, calibration, and uncertainty around the incremental gains.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.