acceptodds
Under review as a conference paper at ICLR 2027

What a Context-Blind Proposal Costs: Structured Operators versus Selection in Event-Announcement Forecasting

Abstract

Text-conditioned forecasting predicts a series from its numerical history together with natural-language context describing events the history cannot reveal. The dominant training-free recipe keeps a time-series foundation model as the numerical engine and lets a language model select among sampled trajectories. We show this carries a cost whenever the proposal is context-blind: a selector can only return a trajectory the proposal happened to draw, and the budget needed before an announced deviation appears grows exponentially in how surprising that deviation is—a statement about sample complexity, not impossibility. On the Context-is-Key subset whose context announces a dated, signed, magnitude-specified deviation ( families, instances), an oracle selector over candidates reaches only of the context-blind error, while on the same samples a five-parameter operator reaches , a richer edit schema , and a frozen B model that writes that schema zero-shot—never touching the ground truth—reaches , beating the selection oracle on instances. Two oracle-guided sequential proposals do no better (, ). The regime is narrow: across all families the operator ceiling no longer dominates ( against ), and a surprise statistic computed without consulting either branch classifies Time-MMD as out of scope, where the ordering reverses. We have no forecasting-time scope detector; three attempts fail. The contribution is an action-space analysis with a characterised regime, not a general text-conditioned forecaster.

Primary area:
learning on time series and dynamical systems
Keywords:
text-conditioned forecasting, time-series foundation models, structured operators, test-time selection

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

This paper studies an interesting problem in text-conditioned time-series forecasting models, when language context announces events that are not reflected in the numerical data. The paper claims that the selection-based approaches have limitations when they cannot recover events that are not in the candidate pool, and the required sampling budget can grow exponentially with the event’s surprise. Therefore, the paper applies structured, context-aware operators like scaling, shifting, adding, or setting values over specified time intervals, to the model’s sampled trajectories, with an LLM extracting the operator parameters directly from text. Extensive experiments on event-announcement tasks show that these operators outperform even oracle-assisted selection, while zero-shot LLM-generated edits outperform selection on most in-scope cases and transfer across different forecasting backbones. However, the paper and rebuttal shows that the proposed approach works mainly for clearly specified, localized events; they do not determine when editing will actually improve the forecast or make it worse.

Critical strengths
  1. I think the paper identifies a fundamental limitation of context-blind proposal-and-selection methods where a selector cannot recover a context-induced event that does not appear in the sampled candidate pool. The action-space perspective is an interesting insight.

  2. Extensive experiment results provide credible evidence that structured, context-aware modifications can substantially outperform selection in the targeted event-announcement regime.

Critical weaknesses
  1. One major limitation is that the proposed operators are designed for CiK tasks that explicitly specify event timing, direction, and magnitude, while they perform poorly on out-of-scope families and Time-MMD. Since the operator schema closely matches the benchmark’s event structure, it remains unclear whether the reported advantage generalizes to independently designed text-conditioned forecasting tasks.

  2. It is not clear when a proposed operator should be applied. The strongest comparisons rely on oracle-fitted operators, while the zero-shot extractor can be harmful outside the intended regime. Moreover, the proposed gating strategies fail to reliably distinguish cases where editing improves the forecast from those where it degrades the original prediction. This substantially limits the practical applicability and general significance of the approach.

Questions for authors
  1. How to determine the operator schema, and is the schema fixed before examining the CiK tasks?

  2. How should a deployed system determine whether to apply an operator or preserve the original forecast?

  3. How sensitive are the conclusions to the choice of selection baseline?

Reviewer_r8tNOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies a forecasting environment which contains both numerical and textual information. One approach is to use a time series foundation model (TSFM) to propose candidate trajectories and have a selector choose among these proposals. The paper points out that in some situations (for example, something like “a heat wave will lead electricity consumption to be four times the usual”) the TSFM may not generate good candidate trajectories. So they propose to edit the TSFM proposals through transformations like scaling and shifting proposed by an LLM after it reads the textual information. Using the Context-is-Key dataset, they (1) have the TSFM generate 256 candidates from the numerical history and pick 25 that increase performance given future information, and (2) have the TSFM generate 25 candidates and have an LLM transform the candidates. The main result is that in a favorable setting, the second method outperforms the first. There is also some formalism on the argument that selection requires many samples to get good performance on surprising events, and results using a second TSFM.

Critical strengths

The paper includes experiments using multiple TSFMs, which adds some robustness.

Critical weaknesses

The evaluation is selected to be quite favorable to the method. In the main evaluation, the paper considers 13 of the 71 Context-is-Key families “whose context announces a dated, signed, magnitude-specified deviation inside the forecast window.” These are also the characteristics of the transformations they allow the LLM to perform on the forecasts. For example, if there is some textual information that future values will be multiplied by four, but the selector can only use forecasts drawn from historical data, it is perhaps not surprising the selector's forecast will be worse. So I am not sure what we have learned here.

The methodological novelty is limited, as the paper’s operator is quite similar to the idea and transformations in Liao et al. (2026).

Questions for authors

What exactly was the procedure for selecting the subset of 13 families from the dataset?

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

The paper studies the sampling cost of selecting forecasts from a frozen, context-blind time-series model when text announces a dated, signed, magnitude-specified event. It combines sample-complexity analysis with oracle-assisted selection and operator comparisons, and uses a frozen LLM to translate text into structured additive, multiplicative, or absolute-value edits applied to forecast samples. On 13 Context-is-Key event families (195 instances), the zero-shot Phi-4 operator achieves a median relative RCRPS of 0.098 versus 0.864 for ground-truth-assisted selection from 256 candidates. The advantage is regime-dependent: broader benchmark results and Time-MMD expose harmful edits and limited generality, and a reliable forecasting-time gate remains unresolved.

Critical strengths
  • The experimental design helps distinguish limitations of candidate generation from those of selection and text-to-parameter extraction. Ground-truth-assisted selection, oracle-fitted operators, and oracle-free text-derived edits provide complementary comparisons, while stronger selection searches, larger sampling budgets, and a second TSFM backbone offer additional support for the identified candidate-generation bottleneck in the event-announcement regime.

  • The paper transparently characterizes its scope and failure modes. Beyond the favorable event-announcement subset, it reports results on the full Context-is-Key benchmark and Time-MMD, as well as negative or mixed findings for gating and supervised fine-tuning. Reporting both median and pooled error ratios also exposes the impact of severe failures that a single aggregate could obscure.

Critical weaknesses
  • The methodological novelty is limited: the strongest gains occur when the text already specifies the timing and magnitude of the required edit. The proposed pipeline largely amounts to extracting these parameters and applying predefined transformations, and the evaluation does not establish an additional advantage over a simple parameter-extraction and rule-based editing baseline or a context-aware proposal.

  • The paper lacks explicit mechanisms and controlled experiments for ambiguous text or conflicting information from multiple sources. Applying a single deterministic set of edits to all forecast samples does not explicitly represent uncertainty over textual interpretations. Applicability to settings such as stock forecasting, where sources may imply opposing effects and uncertain magnitudes or timing, therefore remains unclear.

Questions for authors
  1. What substantive methodological or analytical contribution goes beyond extracting explicitly stated event parameters and applying rule-based edits? Please clarify what existing evidence distinguishes this contribution from straightforward parameter extraction and editing, or establishes an advantage relative to a context-aware proposal.

  2. How does the method handle ambiguous text or conflicting directions from multiple sources, and represent uncertainty over the resulting edits? Please identify any existing evaluation of these cases, or clarify the limitations they impose on the method’s applicability.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.