acceptodds
Withdrawn from ICLR 2027

S2G-HAMC: Forecast-Calibrated Context Adaptation for Frozen Time-Series Foundation Models

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Time-series foundation models (TSFMs) can forecast without task-specific retraining. However, every frozen forecaster still depends on the numerical context it receives. Deployment observations may not align with all coordinates in that context. The system must therefore map timestamped observations to a canonical context. We introduce S2G-HAMC, a forecast-calibrated context-adaptation framework for frozen TSFMs. Its S2G-Adapter builds a support-aware canonical context. It learns two calibration scalars on synthetic signals by using forecast loss from the frozen forecaster. It does not fit real benchmark targets. HAMC is a fixed robustness module. It evaluates nearby S2G context views and combines their forecasts across the prediction horizon. It does not change the backbone or add trainable parameters. S2G-HAMC improves 35/36 backbone–dataset pairs across 12 real-world datasets and three frozen forecasting families. Its mean Weighted Quantile Loss (WQL) gains are 7.62–8.60% over the corresponding validation-selected Linear interface. S2G alone yields 7.39–8.24% mean gains, and the complete system has a larger mean gain than S2G alone on all three backbones. The S2G-HAMC gain also increases as observation retention falls from 0.40 to 0.10. At the sparsest level, the gains reach 9.61–15.15%. With all adaptation choices fixed, the framework also improves quantile loss in 22/24 cells across three previously unused domains. These results show that S2G-HAMC can adapt frozen TSFMs through their numerical context. Across 12 datasets with 80% missing histories, the proposed adapter achieves the best mean Weighted Quantile Loss (WQL) rank among eight frozen/pretrained forecasting models and ranks first on 6 of 12 datasets. The same frozen adapter further transfers to held-out domains without retuning. These results demonstrate that explicitly modeling sampling geometry provides an effective forecast-oriented context interface for robust probabilistic forecasting under severe observation sparsity.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Foundation Models, Time Series Forecasting, Context Adaptation, Context Construction, Probabilistic Forecasting

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_FewJOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes S2G-HAMC, a context calibration method that keeps the backbone TSFM frozen, constructs multiple canonical input contexts, and combines the TSFM's forecasts on these contexts with fixed weights to improve forecast robustness. The method consists of two components. The S2G Adapter adaptively smooths the input window according to observation density to build the canonical context for the frozen TSFM. HAMC combines the forecasts from multiple contexts with different bandwidth scales using fixed, horizon-dependent weights. Throughout this process, the backbone parameters are left entirely unchanged, and only two S2G scalars are trained. With this approach, the method improves the WQL of TSFMs by 7.62–8.60% when 80% of the observations are missing, and S2G alone achieves improvements of 7.39–8.24%. These results show that the proposed method can improve TSFM performance through context calibration alone, without modifying the TSFM weights.

Critical strengths
  • The method trains only two scalars, which represent the base bandwidth and the sensitivity to observation density, and uses a fixed-weight combination scheme. It thereby improves TSFM forecasting performance without any additional trainable parameters.
  • Instead of fitting to the real benchmark datasets used for evaluation, the scalars are calibrated on synthetic data containing a trend, seasonal components, a chirp, and heteroskedastic noise. This has the advantage of not fitting directly to the evaluation datasets.
  • Experiments on diverse datasets from various domains show that the proposed method can indeed improve TSFM performance.
Critical weaknesses
  • W1. Poor Readability
    • Many key terms are used without being defined. For example, "canonical context" is mentioned repeatedly throughout the paper, starting from its first appearance in the Abstract. However, apart from an indirect description in Sec. 3.2, it is never explained, and even Sec. 3.2 does not define it precisely.
    • Many notations are used without definition. For example, hminh_{min} and hmaxh_{max} in Eq. (4) are not defined in the main text, and their units are not specified.
    • Terms are used inconsistently. The term "query" appears in several forms, such as "model queries", "canonical query", "query", and "query-local", so readers must infer what each one refers to.
  • W2. Lacks Details
    • The proposed architecture involves many hyperparameters, such as the support score clip, bandwidth clip, and kernel function. Table 7 reports sensitivity for only some of them (temperature, bandwidth clip, and ridge), and only for Chronos-2 on four datasets. No rationale or impact analysis is provided for the others, including the support score clip, the kernel, and the fallback constants.
    • In particular, for HAMC, a main component, the paper does not explain why the forecasts are combined with the stated ratios, including 0.95 / 0.05 in Eq. (6).
    • The paper states that θa\theta_a and θb\theta_b are trained on synthetic data. However, the synthetic data generation process is described only as an equation, and it is not specified what data is used or how it is generated.
    • The candidate grid, selection criteria, and results corresponding to "Benchmark validation selects…" in Sec. 3.4 and "as described in the appendix" in Sec. 4.1 are not provided in the appendix.
  • W3. Limited Reproducibility
    • The source code is not released, so it cannot be verified whether the implementation matches the description.
    • Context length appears to be a key hyperparameter, but only the candidate lengths are given (Table 9); the actually selected values are not reported.
  • W4. Lacks Diverse Experimental Results
    • No experimental results are provided for different forecast horizons.
    • The paper states that confidence intervals are provided, but the full raw results are not reported; only averages are presented.
    • Table 3 presents various TSFMs as baselines, but the proposed method is applied only to Chronos-2. Its performance when applied to the other models is not reported.
    • The method has a large number of hyperparameters, but no ablation is provided for each of them.
Questions for authors
  • Q1. As mentioned in W3, could you release the source code through the supplementary material or an anonymous GitHub repository?
  • Q2. Including the points raised in W2 and W4, could you provide the design choice, final selected value, ablation results, and other details for every hyperparameter? In particular, please include why clipping is applied and what effect it has, as well as an ablation on the combination ratios in HAMC.
  • Q3. As noted in W1, could you provide a consolidated list of all terminology and notation used in the paper?
  • Q4. Could you also report results under a fixed setting in which the context length and HAMC constants are not selected using benchmark validation data?
Reviewer_CkbkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes an S2G-HAMC framework for adapting frozen time-series foundation models to sparse or irregular observations without modifying the model parameters. S2G-HAMC contains two parts: S2G is designed to select the local interpolation bandwidth based on the observation support at each time step, producing a regular context for the frozen forecaster. The bandwidth parameters are calibrated using synthetic signals and the forecaster’s own prediction loss. HAMC then combines forecasts generated from multiple context scales. Extensive experiments show improvements over a Linear context interface in 35 of 36 model–dataset pairs, with average WQL gains of 7.62–8.60%. However, most of the improvement comes from S2G, while HAMC provides only a modest additional gain at the cost of multiple backbone forward passes.

Critical strengths
  1. The proposed S2G provides a useful adaptation interface between irregular observations and the canonical context expected by a frozen TSFM. S2G is lightweight and potentially practical for deployment.

  2. Extensive results support the effectiveness of proposed context-construction strategy with various datasets and backbones.

Critical weaknesses
  1. The major weakness is that the proposed method appears heuristic. S2G combines local polynomial interpolation with a support-dependent bandwidth, while HAMC averages forecasts from several manually specified context scales. There is a lack of strong theoretical or conceptual justification for why choosing these design and why they would work. Also, HAMC contributes only a very small additional gain over S2G in the final holdout. This makes the overall novelty and necessity of the proposed framework questionable.

  2. The empirical evidence does not establish superiority over sufficiently strong alternatives. Most improvements are reported relative to a Linear context interface, rather than against stronger methods for irregular or missing time-series observations or more competitive context-adaptation approaches. In addition, the key calibration parameters are learned from synthetic signals whose relationship to the real benchmark domains is unclear. Thus, the reported gains may be credited to the choice of baseline and evaluation protocol rather than a broadly effective adaptation principle.

Questions for authors
  1. Could the authors compare S2G-HAMC with stronger baselines for irregular or missing time-series forecasting, such as learned imputation/context-adaptation methods and recent frozen-forecaster adaptation approaches, rather than primarily the Linear interface?

  2. How sensitive are the results to calibrating the S2G parameters on synthetic signals? Will the synthetic calibration transfer across different temporal patterns, noise levels, sampling processes, and real-world domains?

  3. Given that HAMC provides only a 0.452% additional gain over S2G in the final holdout while requiring four backbone forward passes, what evidence supports its practical necessity? Also, could the authors report the accuracy-inference-cost tradeoff and clarify why the final-holdout improvement is much smaller than the gains reported in the main experiments?

Reviewer_3DQvOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes S2G-HAMC, a lightweight context-adaptation framework for frozen time-series foundation models (TSFMs) that maps sparse or irregularly sampled observations onto a canonical input grid using a support-aware local polynomial projection calibrated via synthetic forecast loss.

Critical strengths
  1. Treating the observation-to-context mapping as an adaptation surface that requires calibrating only two scalar parameters without modifying backbone weights is practically valuable because it transfers seamlessly across diverse TSFM architectures.
  2. The proposed method improves Weighted Quantile Loss and shows monotonically increasing gains as observation retention substantiating the effectiveness of local support conditioning under severe data sparsity.
Critical weaknesses
  1. The experiments are relatively limited, and comparisons with other adapter methods are absent. This makes it unclear to me how much gain this method can achieve compared with other state-of-the-art approaches.
  2. HAMC requires 4x the backbone forward passes for a marginal ~0.23–0.47% WQL improvement over S2G alone, and the supposedly gradient-learned calibration scalars in Table 5 are inexplicably identical to four decimal places for two completely different architectures, casting doubt on the synthetic calibration procedure.
Questions for authors
  1. Could you compare your method with the methods mentioned in your paper, such as δ\delta-Adapter and COSA? This will help me clearly understand the improvements brought by the method in this paper.
  2. Why did AdamW optimization of (θa,θb)(\theta_a, \theta_b) in Table 5 converge to the exact same four-decimal round values (a=0.2000,b=0.3800a = 0.2000, b = 0.3800, and a=0.75a = 0.75 for b=0b = 0) across two completely different model families (TimesFM-3 and TiRex-2)?
  3. How does S2G perform against stronger, non-linear or density-aware interpolation baselines and native TSFM NaN-masking interfaces?

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

Related papers

Discussion (0)

Sign in to comment.