acceptodds
Under review as a conference paper at ICLR 2027

Retrieval Is Density Estimation: Calibrated, Invariant, and Selective Retrieval-Augmented Forecasting

Abstract

Zero-shot forecasts from time-series foundation models such as Chronos and MOIRAI show severe performance drops when test distributions diverge from the pretraining data. While retrieval augmentation is intended to bridge this out-of-distribution generalization gap by supplying external memory, existing methods often fail. Because existing methods misalign unnormalized series via Euclidean matching, collapse predictive uncertainity through latent pooling, and unconditioned retrieval forces neighbor injection even when candidate trajectories diverge. Across standard benchmarks, unconditioned retrieval degrades baseline accuracy on 41–50% of test windows. We frame retrieval-augmented forecasting as non-parametric conditional density estimation over a scale-translation quotient manifold . By mapping lookback windows onto invariant shape orbits on , our approach contracts the covering space (, scaling exponent ) and thereby eliminates the need for post-hoc affine alignment. To aggregate forecast distributions, we use weighted 2-Wasserstein barycenters in quantile space. When local candidate scattering indicates negative expected utility, an instance-adaptive controller selectively abstains from retrieval. Under abrupt distribution shifts, a non-exchangeable split-conformal step expands prediction intervals to restore nominal coverage. Across four ETT datasets and three foundation models (Chronos-Bolt, Moirai-2.0, TimesFM-2.5), Calibrated Invariance-Aware Selective Retrieval (CAVIR) lowers CRPS by up to 2.15%, achieves 80.1% coverage under macroeconomic shocks, and stays within 0.85% of baseline error under strict deduplication checks.

Primary area:
learning on time series and dynamical systems
Keywords:
Time-series foundation models, retrieval-augmented forecasting, out-of-distribution generalization, invariant shape matching, Wasserstein barycenters, adaptive retrieval, conformal prediction

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

The paper proposes CAVIR, a retrieval augmentation framework for frozen time-series foundation models. It combines retrieval intended to be invariant to scale and translation, alignment of retrieved future trajectories to the query’s coordinates, aggregation through one-dimensional Wasserstein quantile barycenters, an adaptive controller that can abstain from retrieval, and conformal interval calibration. The submission develops nearest-neighbor bias–variance arguments and claims covering-number and distributional error guarantees. Evaluation includes four ETT datasets, additional forecasting domains, experiments with three foundation-model backbones, and ablation and deduplication analyses. The reported results include consistent but mostly modest ETT improvements, larger error reductions on Weather and Traffic, and approximately nominal coverage in selected shock experiments.

Critical strengths

The combination addresses a practically meaningful problem in using external memory with frozen forecasters. The paper recognizes that retrieval usefulness varies across queries and that similarity of past contexts does not necessarily imply agreement among future trajectories. Combining distribution-level forecast aggregation with an explicit option to decline retrieval is a useful design direction. This perspective is informative and could provide a practical way to improve frozen models without retraining their backbone. I find the attempt to connect neighborhood quality, forecast uncertainty, and retrieval decisions valuable, even though the proposed guarantees require substantial revision.

There is encouraging empirical evidence, together with useful attention to retrieval-specific failure modes. Table 1 reports improvements in both MSE and CRPS on all four ETT datasets; the ETTm1 reduction from in MSE is appreciable. The cumulative ablation and experiments examining target-memory exclusion and near-duplicate removal address relevant questions about where retrieval gains originate. The paper also reports failures when absolute scale carries predictive information, which is an informative limitation of the proposed invariance. These observations provide some support for the practical direction, although the reporting and protocol concerns below limit the conclusions that can currently be drawn.

Critical weaknesses

The central geometric and distributional guarantees are not established for the method as defined. The bounded affine transformation family in Eq. (1) is generally not a group: bounding scale and translation parameters does not preserve closure under composition or inversion. For example, allowing scales in [1/2,2][1/2,2] permits multiplication by two, but composing this transformation with itself produces multiplication by four. Consequently, the claimed orbit quotient and compact-group argument underlying Proposition 4 do not follow from the stated construction. Moreover, the alignment residual in Eq. (2) need not be symmetric because scaling is not a Euclidean isometry. It therefore cannot simply be treated as the claimed quotient metric. These issues undermine the advertised covering-number contraction, rather than only its exposition.

There is a separate gap in Proposition 5. Section 4.4 describes QiQ_i as an empirical quantile function derived from a retrieved trajectory, whereas Appendix F.6 assumes Qi(τ)=Q(τ∣xi)Q_i(\tau)=Q(\tau\mid x_i), the true conditional quantile function. The Jensen argument is valid for those oracle distributions, but does not account for estimation error in the empirical quantities supplied by the method. In particular, the stated bound becomes zero when all retrieved contexts coincide with the query, even though finitely many observed futures cannot generally determine the true conditional distribution exactly.

These problems affect headline theoretical contributions and require a corrected formulation and analysis. The insufficient specification of how each trajectory yields horizon-specific quantiles also makes it difficult to determine precisely which estimator the theory should analyze.

The calibration and evaluation evidence does not substantiate the claimed reliability of the complete system. The conformal construction is insufficiently specified, and its guarantees are overstated. The score and endpoint correction in Section 4.4 follow conformalized quantile regression, which should be attributed to Romano, Patterson, and Candès (2019). The cited covariate-shift result requires an unchanged conditional outcome distribution and appropriate likelihood-ratio weighting; the cited result beyond exchangeability instead permits a nonzero coverage gap. Neither directly establishes nominal coverage under arbitrary abrupt regime changes. The submission needs to identify the applicable assumptions and the actual weighted calibration procedure.

The empirical evidence also needs reconciliation. In Table 1, Cov@80 worsens from 78.83% to 75.38% on ETTh2 and from 79.58% to 76.38% on ETTm2. Table 6 reports substantially different coverage from the main tables. For example, 37.66% versus 80.12% on ETTh1. A consistent affine change of output units alone cannot change interval coverage; if the target channels, evaluation windows, or forecasting configurations differ, those differences must be stated. Furthermore, removing the selective gate leaves every reported Table 6 metric unchanged across all four datasets, so that ablation does not demonstrate the gate’s contribution in that setting.

Finally, the approximately 80% shock coverage is reported without interval-width or interval-score comparisons and without a backbone baseline using the identical conformal calibration procedure. These comparisons are necessary to establish useful uncertainty improvement and attribute it to the proposed retrieval framework. Taken together, these issues prevent the current experiments from establishing the paper’s central calibration and safety claims.

Minor: The presentation makes the method unnecessarily difficult to follow. The manuscript would benefit from careful proofreading: for example, the abstract contains a sentence fragment beginning “Because existing methods…” and the typo “uncertainity.” Beyond language, the ordering of the exposition needs improvement. Section 4.5 discusses Theorem 1 and Proposition 2 before their statements while introducing several controller quantities, making it difficult to distinguish the implemented procedure from its theoretical motivation. The relationship between the embedding distance d(eq,ei)d(e_q,e_i) in Eq. (4) and the proposed quotient distance in Eq. (2) also needs to be explained explicitly. Figure 1 identifies the main modules, but does not sufficiently clarify how retrieved trajectories become quantile forecasts or how the controller determines retrieval and fusion decisions. A self-contained algorithm description, consistent notation, and a clearer account of the data flowing between modules would substantially improve readability.

Questions for authors
  • How can the transformation family in Eq. (1) satisfy the group axioms, and how can Eq. (2) satisfy symmetry? Please provide a corrected construction and identify which theoretical conclusions, especially Proposition 4, remain valid.
  • How is Qi(τ)Q_i(\tau) constructed from a single retrieved future trajectory, separately for each forecast horizon? Is it based on additional model forecasts, values pooled across time, or an empirical distribution across neighbors? How does Proposition 5 account for the estimation error of the quantities actually used?
  • Which precise coverage theorem applies to the implemented conformal procedure? Please specify the likelihood-ratio estimation and weighted quantile rule, the assumptions about temporal dependence and distribution shift, and which residuals are available when an interval is issued. In the “strict zero-shot” experiments, are target-domain labels used for controller training or calibration?
  • What differences in target channels, evaluation windows, preprocessing, or model configuration explain Tables 1/4 versus Table 6? How do the raw ETTh1 MSE values cited in the main text relate to Table 6, and why are the full and ungated configurations identical throughout that table?
  • On matched evaluation windows, how does CAVIR compare with the backbone using the same conformal calibrator, and with a straightforward normalized-neighbor baseline using the same fusion and calibration? Please include interval width or interval score alongside coverage. These comparisons would materially affect my assessment of the additional value of the proposed framework.
Reviewer_Cwo3OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces a retrieval-augmented forecasting framework (CAVIR) designed to improve the robustness of frozen time-series foundation models under distribution shifts. The proposed method combines affine-invariant nearest-neighbor retrieval, Wasserstein barycenter aggregation of predictive quantiles, an adaptive mechanism that selectively disables retrieval, and conformal prediction for uncertainty calibration. The paper also provides theoretical results concerning retrieval bias–variance trade-offs, neighborhood selection, and distributional forecasting error. The framework is evaluated three different foundation models, namely Chronos-Bolt, Moirai-2.0, and TimesFM-2.5 and four benchmark data sets. The results reported show modest forecasting improvements on the benchmarks and improved prediction interval coverage under selected distribution shifts.

Critical strengths
  • The paper addresses the limitations of retrieval-augmented forecasting under distribution shifts; in particular, the possibility that retrieving irrelevant historical observations can degrade predictions. Further, the option to allow the method to abstain from retrieval when historical neighbors are uninformative can lead to robustness improvements.

  • The framework also is equipped with uncertainty quantification, rather than only point forecasts.

  • Carefully designed ablation studies that provide insights into the role and effects of invariant retrieval, adaptive neighborhood selection, selective gating, and conformal calibration. The paper also pays attention to data leakage issues.

Critical weaknesses
  • The proposed framework primarily combines established techniques from nonparametric estimation, invariant representations, Wasserstein barycenters, selective prediction, and conformal inference. Hence, it remains rather unclear what fundamentally new methodological contribution is introduced beyond their integration into retrieval-augmented forecasting.

  • Section 4.3 defines a bounded family of affine transformations and considers it as a compact group. This is not correct, since it is not closed under composition or inversion. This group structure claim is then leveraged in the quotient manifold and covering number results, which consequently are also not correct. I believe that the authors have the right geometric intuition about the bound they are seeking; for example, by defining G as a compact subset of the full, unrestricted affine Lie group and further by considering not a quotient manifold, but some appropriate equivalence relationship of projection, they can potentially re-establish the desired bound that involves the volume of G.

  • The proposed bias–variance, adaptive-neighborhood, and Wasserstein risk results largely follow from classical nonparametric estimation and standard inequalities. Further, their applicability to the actual retrieval procedure is not clearly established.

  • The established performance gains are relatively small, while inference time is significant. Hence, there is doubt about the practical benefits of the proposed framework, compared to the additional complexity introduced.

  • The paper makes claims about finite sample coverage under distribution shifts. However, such results depend on assumptions that are not posited, especially for abrupt regime changes. A mere reference to background papers, without clearly arguing that the required conditions are satisfied is not adequate.

Questions for authors
  • Check mathematical validity of the quotient formulation.

  • Validity of the conformal guarantees.

  • What is the practical significance of the proposed framework, given increased cost and rather marginal performance improvements?

  • How does the framework determine when affine invariance is appropriate?

  • How do simple competing methods such as NN forecasting and simple quantile-based forecast combinations perform using the same retrieval memory and calibration data?

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

The paper proposes Calibrated Invariance-Aware Selective Retrieval (CAVIR) for zero-shot time series forecasting. The goal is to predict the future yqy_q of the query series xqx_q. CAVIR works as follows:

  1. It maintains a memory bank consisting of historical context and future.
  2. CAVIR uses a backbone foundation model to make an initial prediction
  3. CAVIR looks for the context sequences in the memory that are similar to the query series in shape and extracts their future series
  4. Perform the weighted average of the initial prediction and the extracted series from the bank in the quantile space
  5. A learned gate is used to check whether the extracted series are used. If the gate is 0, only the backbone prediction is given
  6. Finally, conformalize the 80% interval for better calibration

Experimental results show that the CAVIR has better point predictions compared to the backbone foundation model. For probabilistic predictions, the improvements are small.

Critical strengths
  1. Selective retrieval in CAVIR is a simple and practical idea. It can be used on any backbone model that provides quantiles. Averaging in quantile space retains the spread of each forecast.
  2. Experiments show that CAVIR performs better than the baseline foundation model in terms of MSE and slightly in terms of CRPS. Experiments also show the calibration of CAVIR via coverage, and CAVIR mostly improves the coverage of the backbone foundation model, especially on shock windows.
Critical weaknesses
  1. The zero-shot claim of the paper is not clear. In the experiments, the paper says the data is split 60:20:20 for training, validation, and testing, respectively. Yet, in Section 1, the paper mentions no target sequences in the memory. Appendix C states that the same series is used to train the controller and for conformal calibration.
  2. The comparison looks unfair. CAVIR has additional information, while the baseline model does not. Since the model already accesses additional data, a simple baseline trained on the additional data and evaluated on the target data should be seen as the floor. Furthermore, paper states that it follows the TS-RAG setup, yet there is no result on TS-RAG.
  3. Results in Table 3 are off. Chronos-Bolt has a very high MSE compared to CAVIR (almost 4 times in the weather and traffic datasets). This looks like a pipeline issue. To clarify this, one could run the experiment on the splits of the backbone.
  4. Authors argued that their model is generalized to any foundation model that provides quantiles. Yet, there is no comparison with models like Moirai and TimesFM for the main metrics such as MSE, MAE, and CRPS.
  5. Writing can be improved. The method is relatively simple in a positive manner, yet the heavy terminology obscures its advantage. Furthermore, Figure 1 is hard to understand.
Questions for authors
  1. Can you explain how the shocks are defined? The authors conducted an experiment on a dataset with shock windows, but there is no clear definition of how those windows are selected.
  2. What happens if we use conformal prediction for the backbone without retrieval? This separates the effect of retrieval from the effect of calibration. It is not clear from Table 4.
  3. Can you explain the architecture of the controller?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.