acceptodds
Under review as a conference paper at ICLR 2027

Adaptive Multi-Resolution Inference for Frozen Time-Series Foundation Models

Abstract

Time-series foundation models (TSFMs) have demonstrated zero-shot forecasting capabilities, but standard inference typically queries the model only once at the original temporal resolution. This may leave predictive information at other time scales underused. We study how multiple temporal resolutions can improve forecasts at the original resolution when the foundation model is fully frozen and the number of queries is limited. To this end, we propose Adaptive Multi-Resolution Inference (AMRI). AMRI aggregates the observed history to multiple temporal resolutions and queries the same frozen TSFM at each resolution. The outputs predict different temporal aggregates of the same future trajectory and therefore share known structural relationships. Using these aggregation relationships and forecast reliability estimated from historical errors before testing, AMRI reconciles the outputs through closed-form weighted least squares to obtain the final forecast at the original resolution. Adaptivity comes from reliability calibration for the target series. The resolution set and calibrated weights remain fixed during testing. The entire procedure requires no gradient-based training, parameter updates, or additional learned forecasting modules. Under unbiased forecast errors and exact population inverse-covariance weighting, we provide an expected squared-error guarantee for the reconciled forecast relative to the original-resolution forecast. Experiments across multiple frozen TSFMs and forecasting benchmarks demonstrate the effectiveness of AMRI. Code is available in the anonymous repository.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Time Series Forecasting, Drought Prediction

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

This paper investigates how multi-resolution queries can improve forecasting accuracy when the foundation model is fully frozen and the number of queries is constrained. Specifically, the observed history is block-averaged at different resolutions (r=1,2,4,8r = 1, 2, 4, 8), and the same frozen TSFM is queried separately at each resolution. Because these outputs represent different aggregations of the same future trajectory, they satisfy known aggregation relationships, represented by matrix AA. Weighted least squares (WLS) can therefore reconcile the multi-resolution forecasts into a forecast at the original resolution. The weights are estimated from historical forecast-error variances on a calibration segment (CAL) and remain fixed during testing. Under the assumptions of unbiased errors and exact population inverse-covariance weighting, the theoretical analysis proves that the expected squared error of the reconciled forecast does not exceed that of the original-resolution forecast. Experiments cover seven frozen TSFMs, eight datasets, and four forecast horizons, yielding 224 evaluation cells.

Critical strengths

Strength 1: The method is conceptually simple and practically useful.

AMRI requires no gradient-based training, parameter updates, or additional learned modules. It improves forecasting accuracy using only multi-resolution queries and closed-form WLS reconciliation. This has clear practical value in deployment settings, especially when the underlying TSFM weights are inaccessible, as with API-based access, or when computational budgets are constrained. Evaluation on seven different TSFMs provides strong support for the method’s model-agnostic applicability.

Strength 2: The connection to classical literature is clear, and the theoretical analysis is well grounded.

The paper explicitly situates multi-resolution reconciliation within the classical forecast reconciliation framework. Proposition 1 directly invokes the Gauss–Markov theorem, and the derivation in Appendix A is logically clear. This positioning as a classical method applied to a new setting helps reviewers understand the work’s intellectual context.

Critical weaknesses

Weakness 1: The core contribution lacks sufficient novelty.

The central technique, weighted least-squares forecast reconciliation, has a history of nearly two decades in the classical time-series literature. The paper’s innovation lies in applying it to multi-resolution queries of frozen TSFMs. However, multi-resolution or multi-scale querying has already been explored extensively in TSFM inference. For example, SMR² enhances frozen TSFMs through multi-resolution residual correction, RGMR uses coarse-to-fine residual refinement, and TimeFound incorporates multi-resolution patching during pretraining.

SMR², published in March 2026, likewise targets inference-time enhancement of frozen TSFMs. AMRI differs technically by using known aggregation relationships for reconciliation instead of residual correction. Nevertheless, it substantially overlaps with SMR² in the core setting of multi-resolution querying with a frozen model. The paper mentions SMR² and RGMR in Related Work, but distinguishes them mainly through methodological details. It does not sufficiently explain why a new reconciliation approach is needed instead of reusing the residual-correction framework of SMR².

More fundamentally, the contribution claimed in the Introduction, “we study how multiple temporal resolutions can improve forecasts at the original resolution when the foundation model is fully frozen,” rests on an insight that is not itself entirely new: multi-resolution queries provide complementary information. Given existing work such as SMR² and RGMR, the paper needs to establish its specific technical contribution more convincingly, beyond presenting another variant of multi-resolution inference.

Weakness 2: The experimental evidence does not sufficiently support the central claim, and some results run counter to it.

The paper’s central claim is that AMRI improves the forecasting accuracy of frozen TSFMs through multi-resolution reconciliation. However, the experimental evidence has two problems.

(a) Performance gains are negative on some datasets and vary substantially. On electricity and traffic, AMRI has higher pooled MSE than the single-resolution baseline, as shown in Table 1. AMRI improves performance in 185 of the 224 cells, meaning that performance deteriorates or remains unchanged in 39 cells. Figure 4 shows that even among the improved cells, the gains are widely dispersed. The paper provides neither a systematic analysis of these failure cases nor an account of the conditions under which AMRI fails. This is an important gap for a method presented as broadly applicable.

(b) The comparison with long-context querying reveals a plausible alternative explanation. Section 5.5 reports that, on a 128-cell subset, a single long-context query achieves lower pooled error than AMRI, whose error is 1.7% higher. This suggests that AMRI’s gains may primarily come from extending the effective history window of each query, from 512 to 512r512r, rather than from multi-resolution reconciliation itself.

Under the default expand setting, a query at resolution rr uses up to 512r512r original observations. Thus, the r=8r=8 query uses 4,096 observations, effectively providing an implicit long context through multi-resolution querying. If long-context querying alone achieves comparable or better performance, the necessity of AMRI’s reconciliation framework is weakened.

The paper needs experiments that disentangle multi-resolution reconciliation from the expansion of the effective history window. For example, under the matched history setting, where all resolutions use the same 512 original observations, AMRI achieves a gain of only 4.4%, as shown in Table 2, substantially below the 9.9% gain under expand. This suggests that a substantial portion of the improvement may come from the expanded history window.

Questions for authors
  1. Positioning of the core contribution: Given that SMR² and RGMR have already explored multi-resolution querying of frozen TSFMs, what fundamentally distinguishes AMRI from these approaches? Beyond the technical detail of not using residual correction, could AMRI’s methodological contribution be more clearly positioned as introducing classical forecast reconciliation into TSFM inference? If so, why is this introduction nontrivial?

  2. Attribution of performance gains: How can multi-resolution reconciliation be disentangled from expansion of the effective history window? The gain is only 4.4% under matched history, compared with 9.9% under expand. Does this imply that the main source of AMRI’s improvement is the longer input history rather than multi-resolution reconciliation itself? Could an experiment directly compare a single long-context query using 512 observations with AMRI under the matched history setting?

  3. Practical relevance of the theoretical guarantee: Proposition 1 assumes access to the exact population precision matrix Σ−1\Sigma^{-1}. In practice, AMRI uses a diagonal estimate Λ^\hat{\Lambda}, and CAL contains as few as 15 forecast origins in some settings. Could the authors provide a finite-sample error bound for the diagonal approximation, or at least experimentally quantify the performance gap between estimated weights and oracle weights?

  4. Cases with negative gains: AMRI underperforms the baseline on electricity and traffic. Could the authors provide a diagnostic analysis identifying how these datasets differ from others, such as ETT and weather, in statistical properties including sequence length, autocorrelation structure, and signal-to-noise ratio? Under what conditions does AMRI fail?

  5. Fairness of the experimental comparison: In Table 3, the gradient head covers only 64 cells, whereas AMRI covers 224. Could the authors report results on the same 64-cell subset to ensure that pooled MSE is compared over an identical set of cells?

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

This paper investigates zero-shot forecasting with fully frozen time-series foundation models (TSFMs) under query and input-length constraints, without parameter updates or additional learned modules. To exploit predictive patterns across different temporal scales, the authors propose Adaptive Multi-Resolution Inference (AMRI), which downsamples the observed history into non-overlapping block averages across multiple predefined resolutions R={1,2,4,8}\mathcal{R}=\{1, 2, 4, 8\}, queries the same frozen model at each resolution to predict corresponding future block averages, and reconciles these outputs back into an original-resolution forecast via closed-form weighted least squares (WLS). The reconciliation weights are calculated as regularized empirical inverse forecast-error variances on a pre-test calibration split (CAL) and kept fixed during testing. The method is empirically evaluated across 7 frozen TSFM backbones, 8 benchmark datasets, and 4 forecasting horizons, totaling 224 evaluation cells.

Critical strengths
  1. The empirical evaluation is extensive and systematic, spanning 7 recent foundation models across 8 benchmarks and 4 horizons, and is accompanied by detailed ablation studies covering query budgets, resolution selection strategies, and online weight updates.

  2. The framework is entirely training-free and computationally lightweight during inference, requiring no gradient backpropagation or parameter adaptation while performing forecast reconciliation via a single closed-form matrix-vector product.

  3. The authors provide transparent disclosure, explicitly documenting extensive generative AI assistance for coding, debugging, literature review, and drafting in an AI Use Statement, while candidly reporting negative findings such as a standard single long-context query outperforming the proposed multi-query reconciliation.

Critical weaknesses
  1. The method combines block-average inputs with existing hierarchical forecast reconciliation (WLS/MinT) without introducing new model architectures or foundation-model-specific mechanisms. Furthermore, Proposition 1 applies the Gauss-Markov theorem under the assumptions of unbiased forecast errors and known population covariance, which represents standard generalized least squares properties rather than a new theoretical framework for foundation models.

  2. The reported gains are primarily driven by expanded historical coverage rather than true multi-resolution synergy: when restricted to a matched history budget (512 raw observations across all scales), the MSE reduction plummets from 9.9% to 4.4%, and more critically, Appendix D.1 reveals that a single original-resolution query with long context (e.g., 4096 steps) achieves lower pooled MSE (0.4995) than AMRI (0.5081), leaving AMRI with 1.7% higher error despite consuming four times as many model queries.

Questions for authors

This paper seems to rely too heavily on auto research.

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

effective use of coarser (block averaged) versions of a given time series to provide additional forecasts that can be used to improve the fine-resolution time series forecast using extensions of classical aggregation reconciliation methods from stats. Idea is conceptually simple, well expressed and investigated; the gains are modest but significant, with very little extra work involved, and so this is a very useful method. Expts are quite comprehensive across a range of TSFMs.

Critical strengths

simple setup; mathematically sound reconciliation method, building on prior work in the classical time series forecasting community, detailed expts and appendix.

Critical weaknesses

not really, but see the questions below.

Questions for authors
  1. Since the original time series is "reused" multiple times at coarser resolutions, it may be that a split other than 30/40/30 may be more appropriate (fig 3). The authors should do some expts to see if there is an alternate split that tends to be superior.
  2. In terms of future work: a) beyond the classical aggregation work (Hyndman etc), there have been more recent works (e.g. Xing Han's "Simultaneously reconciled quantile forecasting of hierarchically related time series." In International Conference on Artificial Intelligence and Statistics, pp. 190-198. PMLR, 2021 and associated papers, also see recent review: Athanasopoulos, George, et al. "Forecast reconciliation: A review." International Journal of Forecasting 40.2 (2024): 430-456.) reconciliation in hierarchical TS in broader settings, that can point to avenues for future work. b) The coarser TS augmentations may have additional benefits if the original TS has outliers or additional noise. The authors can inject such aberrations into a TS and then check if the proposed design is more tolerant than other baselines.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.