acceptodds
Under review as a conference paper at ICLR 2027

When Tomorrow Becomes Today: Self-Evolving Policies for Agentic Time Series Forecasting

Abstract

Agentic time series forecasting concerns systems whose underlying mechanisms evolve, making the relative effectiveness of numerical models, reasoning strategies, and intervention rules inherently time-varying. Consequently, a time series agent must adapt the forecasts it produces and the orchestration policy that determines which components to trust and how to coordinate them. The deployment process naturally provides supervision for this adaptation as forecast horizons elapse and realized targets reveal the effectiveness of earlier decisions. Committing all numerical expert forecasts and candidate agent paths before target observation allows each realized outcome to evaluate the entire alternative set, providing delayed feedback without additional annotation. However, existing time series agents primarily incorporate prior experience through forecast refinement, reflection, or retrieval, without systematically converting realized outcomes into persistent updates to the joint orchestration policy governing later origins. To exploit this delayed feedback systematically, we introduce TimEvlove, a frozen-backbone time series agent that converts each realized outcome into persistent joint updates of expert trust, agent path selection, and intervention strength. A temporally ordered predict, reveal, and update protocol applies this feedback to subsequent forecasts. Extensive experiments across eight Time-MMD domains spanning diverse applications demonstrate that TimEvlove achieves state-of-the-art performance. These results demonstrate the value of learning forecasting policies from the futures encountered during deployment. The code is available in the supplementary material.

Primary area:
learning on time series and dynamical systems
Keywords:
Time Series Forecasting, Agentic Time Series

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Va7fOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes TimeEvolve, a self-evolving framework for agentic time-series forecasting that adapts its orchestration policy using delayed feedback from realized forecasting outcomes while keeping the underlying language model and numerical forecasting models frozen. The framework consists of three components: EvolveTrust, which adaptively weights numerical forecasting experts based on their reliability; EvolveReason, which selects among LLM-generated candidate forecasting paths using a learned error predictor; and EvolveIntervene, which determines how strongly the selected candidate should modify the numerical ensemble forecast. Through a predict–reveal–update procedure, the framework uses observed outcomes to update expert reliability, candidate selection, and intervention strength for subsequent predictions, including feedback from unselected alternatives. Experiments on eight Time-MMD domains compare TimeEvolve against 14 baselines and report the best average MSE and MAE rankings, while ablation studies demonstrate improvements over a counterpart with a frozen orchestration policy.

Critical strengths

S1. The paper introduces a structured predict–reveal–update framework that leverages realized forecasting outcomes to update expert weighting, candidate path selection, and intervention strength while keeping the underlying forecasting models and LLM frozen. A notable strength is that the framework can evaluate both selected and unselected forecasting alternatives against the same observed outcomes, providing additional supervision without requiring manual annotations. This offers a practical and computationally lightweight approach to adapting agentic forecasting systems.

S2. TimeEvolve achieves competitive results across eight Time-MMD domains, achieving the best average MSE and MAE rankings among 15 methods. The comparison against TimeEvolve-Stack, which uses the same forecasting models and pipeline without online policy updates, also shows improvements in both metrics across all eight domains. The component-wise and feedback-coverage ablations further support the effectiveness of the proposed adaptation mechanism.

Critical weaknesses

W1. A key motivation behind this paper is the idea that the effectiveness of forecasting models and agent policies can vary over time, making it essential to continually evolve these policies. However, the experiments focus mainly on reporting aggregated forecasting errors for the entire test period. The paper falls short in that it doesn't provide in-depth analyses of how each expert's reliability shifts over time, or how the learned weights respond to these changes. It also misses out on showing how the performance advantage develops as more feedback is gathered. Moreover, the proposed progressive weighting schedule hasn't been tested independently against other scheduling methods. So, while the findings indicate that outcome-driven updates can enhance overall accuracy, they don't convincingly demonstrate that the method can adapt to changing conditions over time, which is a central goal of this research.

W2. The performance boost seen with the new method is hard to pin down to just the proposed policy evolution. In six out of eight areas tested, the frozen-policy version alone shows an impressive improvement, accounting for about 70% to 97% of the MSE reduction compared to KairosAgent. For instance, in the Climate domain, KairosAgent has an MSE of 0.863, while the TimeEvolve-Stack drops that to 0.342. The full method only slightly improves upon this with an MSE of 0.314. This means that an impressive 94.9% of the improvement occurs without any online policy evolution. In the Traffic domain, the MSE goes from 0.151 to 0.089 with the frozen policy, while the full approach achieves 0.087, reflecting 96.9% of the overall improvement. Moreover, the paper doesn’t compare the proposed method to standard online ensemble-learning techniques like gradient-based weight optimization. This omission leaves us unsure whether EvolveTrust really offers any benefits beyond what established methods can provide. Even though the framework includes paths generated by LLMs and external text evidence, there are no tests removing these elements to see how much they contribute. Without these comparisons, it’s hard to determine whether the presented systematic approach is essential or more effective than simpler numerical forecasting and online ensemble methods.

Questions for authors

Q1. The paper motivates TimeEvolve by arguing that expert reliability changes over time, yet the experiments primarily report aggregate test-set performance. Can the authors provide evidence demonstrating that EvolveTrust tracks changes in expert reliability and that the benefits of policy evolution persist or increase as more feedback becomes available?

Q2. In six of eight domains, TimeEvolve-Stack already accounts for about 70%–97% of the MSE improvement over KairosAgent, despite having no online policy updates. Can the authors clarify how much of the improvement comes from the underlying numerical expert pool and initial forecasting pipeline rather than the proposed policy evolution? Are results available under a matched numerical expert pool?

Q3. The numerical forecasting models primarily operate on time-series inputs, while external textual evidence is incorporated indirectly through LLM-generated candidate paths. Can the authors provide numerical-only and LLM-removal ablations to quantify the contribution of textual evidence and LLM-generated paths? Additionally, how does EvolveTrust compare with standard online ensemble-weighting approaches, such as gradient-based optimization or exponentiated-gradient updates, using the same numerical experts?

2 comments

Thank you for the detailed review and constructive suggestions. We appreciate your recognition of the feedback framework and your emphasis on explaining how adaptation develops over time and contributes to forecasting accuracy.

Q1. Does EvolveTrust respond to changing expert reliability and accumulating feedback?

R1. EvolveTrust updates expert weights using errors from completed forecast windows. Its progressive schedule increases the contribution of deployment feedback as more outcomes become available (Section 4.3). Table 3 provides complementary evidence that broader feedback improves forecasting: supervising all precommitted alternatives reduces MSE/MAE by 5.0%/6.3% relative to selected-only feedback. To connect this aggregate result to temporal behavior, we will add chronological plots of expert errors, issued weights, and cumulative loss differences from the evaluation records. These plots will show when allocation changes and whether subsequent predictions benefit.

Q2. How much improvement comes from the initial forecasting pipeline?

R2. The initial pipeline contributes substantially to overall performance. The adaptation increment is measured directly against Stack, which shares the expert pool and forecast-construction pipeline. In Climate, MSE decreases from 0.342 to 0.314, an 8.19% reduction. In Traffic, it decreases from 0.089 to 0.087, a 2.25% reduction. Across all eight domains, the average relative reductions are 6.3% MSE and 7.6% MAE. We will present this matched comparison as the evidence for policy adaptation and use comparisons with KairosAgent to characterize overall system performance.

Q3. What do the numerical policy, LLM paths, and textual evidence contribute?

R3. Table 3 measures the contribution of path supervision while retaining full expert feedback and intervention learning. Expanding supervision from the selected path to all precommitted paths improves MSE and MAE in every domain, averaging 1.6% and 2.2% relative reductions. In Economy, MSE decreases from 0.187 to 0.181. To isolate candidate generation and textual evidence, we will add numerical-only and no-text controls alongside a same-pool online ensemble. Comparing the final forecast with its own recorded prior will additionally quantify the net contribution of path adoption under the same trust state.

We appreciate your suggestions for strengthening the temporal analysis and contribution attribution. We will incorporate these clarifications and targeted comparisons to make the relationship between the proposed mechanism and the observed gains easier to assess.

Thank you to the authors for the rebuttal, which clarifies the benefits of policy evolution and provides an illustrative case study. However, my main concerns remain unresolved.

First, the manuscript still lacks systematic temporal evidence that EvolveTrust tracks changing expert reliability or that its benefits persist or increase as feedback accumulates, while Table 3 reports aggregate results and the progressive weighting schedule remains untested against alternatives.

Second, although the Stack comparison isolates the incremental benefit of policy updates, no matched-expert-pool comparison with KairosAgent is provided, and Section 5.2 continues to use improvements over agentic baselines to support the orchestration mechanism without isolating its contribution.

Third, the comparisons with standard online ensemble methods using the same numerical experts, along with numerical-only and no-text ablations, are not provided.

As these concerns remain insufficiently addressed, I maintain my original rating.

Reviewer_Ji8uOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes TimEvolve, a time-series forecasting agent that learns from delayed feedback without updating its underlying models. The framework combines three components: EvolveTrust adjusts the weights of numerical forecasting experts, EvolveReason selects among candidate reasoning paths, and EvolveIntervene controls how strongly the selected path changes the final forecast. Experiments on eight Time-MMD domains report strong results against fourteen baselines.

Critical strengths

The paper addresses a meaningful problem by turning observed forecasting outcomes into feedback for future decisions. The predict–reveal–update protocol provides a clear way to avoid using future targets during prediction.

The experiments cover diverse forecasting domains, and the component ablations and feedback-coverage comparisons help show the benefits of online policy updates.

Critical weaknesses

(1) The claimed joint policy learning is not fully supported by the method, since expert trust, path selection, and intervention strength are updated through separate learning rules without directly optimizing their interactions.

(2) The comparison with existing forecasting methods is not sufficiently controlled, since the proposed system uses multiple pretrained forecasters and an adaptive ensemble, making it difficult to separate the benefits of policy evolution from those of stronger expert combinations.

Questions for authors

(1) How does TimEvolve compare with simple online ensemble methods, such as exponential weighting or rolling-window expert selection, when using exactly the same numerical forecasters and feedback?

(2) What evidence demonstrates that the three components benefit from joint adaptation rather than independent online updates, and can the authors measure their interactions through more targeted ablations?

2 comments

Thank you for the thoughtful review and for recognizing the value of learning from delayed forecasting outcomes. Your questions about matched comparisons and component interactions help us sharpen the presentation of the method and its evidence.

Q1. How does TimEvolve compare under the same numerical experts and feedback?

R1. Table 2 already matches the numerical experts, initialization, evidence pipeline, forecast origins, and reveal schedule between TimEvolve and Stack. Policy updates reduce MSE and MAE in all eight domains, averaging 6.3% and 7.6% relative improvements. For example, Economy MSE decreases from 0.220 to 0.181. We will extend this control with Hedge, rolling-window expert selection, and the numerical-only EvolveTrust policy, using identical expert forecasts, delayed outcomes, and chronological validation.

Q2. What evidence supports interactions among the three updates?

R2. Trust changes both the prior and candidate mixtures, while the selected path determines the correction calibrated by the gate (Eqs. 5–13). Table 2 also provides evidence of non-additive effects on the observed test set. In Climate, freezing Trust, Reason, or Intervene increases MSE by 0.020, 0.014, or 0.010. Their conditional costs sum to 0.0440.044, whereas freezing all three increases MSE by 0.342−0.314=0.0280.342-0.314=0.028. An additive model would predict equality. This difference supports dependencies among the updates within the evaluated pipeline. We will describe the method as “coordinated adaptation of coupled decisions,” with the separate learning objectives stated explicitly.

We appreciate these focused suggestions. We will incorporate the matched comparisons and clarify the component dependencies so that the revised manuscript communicates the source of the improvements more directly.

These analyses address important aspects of my concerns. I will maintain my current positive score.

Reviewer_bxiCOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper proposes TimeEvolve: a time-series forecasting agent that learns from its own past forecasting errors during deployment. It keeps underlying model frozen frozen and uses newly observed outcomes to continually adjust which numerical experts to trust, which LLM-generated forecasting path to choose, and how strongly that path should modify the numerical forecast. The core claim is that this persistent, outcome-driven adaptation produces better forecasts than a fixed decision policy.

Critical strengths

The paper’s main strength is a clean and intuitive learning setup that once outcomes start arriving, the forecaster can use them to improve future model weighting, path selection, and intervention without retraining the underlying models. The method is modular and interpretable, and the empirical gains from adopting this way of thinking are real.

Critical weaknesses

The main weakness is that the paper is rebranding fairly standard online ensembling as “agentic policy evolution”. The three “joint” updates are separate rules with separate targets, saved together as one checkpoint, with no shared objective.

The experiments also do not isolate the claimed agentic contribution. Their method gets continual test-time feedback and access to a strong pool of forecasting models, but is not compared against a simple online ensemble using the same pool and the same feedback with no LLM reasoning. That is especially important because the ablations suggest that ordinary expert reweighting provides most of the benefit, while the LLM-selected path can add little or even hurt. The frozen-policy variant scores worse than Chronos-2, one of its own experts, in Agriculture and Climate. Freezing trust causes the largest average drop. In the case study, the learned gate nearly switches off the selected path, and the prior alone scores slightly better than the final forecast. That case also was chosen after seeing results and uses a simpler expert pool than the main experiments. So the results convincingly show that online adaptation helps, but not that the proposed agentic orchestration is what makes it work.

The adaptation rules themselves are hand-designed, not learned. Only the coefficients inside them update. The settings governing them, such as cold-start speed, window lengths, penalty strengths, features and the trust-view mixture, are described as fixed calibration choices. The paper gives neither their values nor how they were chosen. This entire meta-policy should also be learned from data, not set by hand.

Questions for authors

First and foremost, you have to isolate what the language model contributes. Add a baseline that combines the same expert pool online without the language model, using simple inverse-error weighting or a standard expert-advice method like Hedge or some type of no-regret updates. Report every pool model on its own as well.

Make the comparison with other adaptive methods fair and transparent. State how each baseline’s numbers were obtained and the information set available to each method.. At a minimum, give MemCast, the same stream of outcomes TimEvolve gets. Report how often the learned gate actually uses the selected path across all test forecasts, not just the hand-picked case. Choose case studies by a rule fixed in advance, don't hand-pick them ex-post.

In order for this to be interesting, you have to learn the adaptation settings from data instead of setting them by hand. Settings like the trust-view mixture, window lengths and ridge penalties can be learned, not set. Hedge-style weighting can replace the fixed view schedule, several window lengths can run in parallel with learned weights, and leave-one-out cross-validation can choose the penalties. Anything left over should be tuned by replaying the training and validation periods as simulated deployment, never touching the test period. You have to study/report the learning dynamics in this meta domain also.

You should broaden the data to support your claims. Time-MMD gives only eight single-variable series, mostly sampled monthly, with many series holding only a few hundred observations. Security, for example, has 297 monthly samples, which leaves about 60 test months and only a handful of feedback cycles. Its web-retrieved text also risks pretraining contamination and leaks of future information. Better candidates are TimesX, with 190 real variables across 19 domains and strict time isolation from model training cutoffs, and Impermanent, which scores forecasts sequentially on continuously updated, non-stationary data streams, matching the paper’s own protocol. More standard datasets, like M4/M5, Monash forecasting datasets, Electricity/Traffic/ETT could also make an appearance in the Appendix somewhere.

Smaller issues: name the language model and test for knowledge leakage, for example by evaluating on data after its training cutoff; report variance across repeated runs and use significance tests where margins are small; position the work plainly within the prediction-with-expert-advice literature; either drop the word “joint” or train the three components against a shared objective.

2 comments

Thank you for the careful review and constructive suggestions. We appreciate your emphasis on identifying the source of the gains, controlling feedback access, and clarifying the learned policy. We provide additional quantitative analysis of the existing experiments below.

Q1. What does TimEvolve contribute beyond online ensembling, and what is the contribution of its agentic components?

R1. Tables 2 and 3 provide controlled comparisons using the same numerical experts, evidence pipeline, candidate-generation configuration, initialization, and forecast origins. Recomputing the domain-wise relative reductions gives the following improvements of TimEvolve over each control:

ControlMSE reductionMAE reduction
All policy updates frozen (Stack)6.3%7.6%
Expert-trust updates frozen4.2%4.8%
Path-selection updates frozen2.8%3.2%
Intervention updates frozen1.9%2.2%
Path feedback restricted to the selected candidate1.6%2.2%

Each percentage averages the relative reduction across eight domains. Each single-component freezing comparison keeps the other two updates active. The path-feedback comparison retains full expert feedback and intervention learning.

These results quantify contributions from adaptive path selection and intervention within the same expert pool. In Economy, for example, freezing path selection increases MSE from 0.181 to 0.196, while freezing intervention increases it to 0.190. Expanding supervision from the selected path to all precommitted paths reduces MSE from 0.187 to 0.181. Thus, learning how to select and apply candidate forecasts improves the use of the numerical experts.

To measure the contribution of LLM generation and external text directly, we will add a compact comparison with Hedge, numerical-only EvolveTrust, and the full system without text. These controls will share expert forecasts, outcome-release times, and chronological validation. We will also report every pool member individually.

Q2. What makes the updates coordinated, and which quantities are learned?

R2. Trust weights determine both the numerical prior and candidate mixtures. Selecting a path determines the correction that the intervention model calibrates (Eqs. 5–13). These dependencies connect the three decisions during forecasting.

The existing Climate results also show non-additive effects on the evaluated test set. Separately freezing Trust, Reason, and Intervene increases MSE by 0.020, 0.014, and 0.010, respectively. Their conditional costs sum to 0.0440.044. Freezing all three increases MSE by 0.342−0.314=0.0280.342-0.314=0.028. An additive model of the three updates would predict equality. The observed difference of 0.0160.016 illustrates how their contributions depend on the other active updates.

The learned quantities are expert reliability estimates and the selector and gate coefficients. Each is updated from previously revealed outcomes. We will describe this mechanism as “coordinated adaptation” and provide the complete calibration settings and selection procedures. Sections 1 and 2.2 establish the connection to prediction with expert advice.

Q3. Are adaptive baselines treated fairly, and how should the gate example be interpreted?

R3. We will document each baseline’s result source, information set, initialization, and update timing. The MemCast comparison will retain its native memory mechanism while matching TimEvolve’s forecast origins and delayed-outcome stream.

The Climate example illustrates the gate’s control of candidate influence. The selected path has MSE 0.081219, while the updated numerical prior has MSE 0.021866. Assigning the path a weight of 0.009 yields a final MSE of 0.022173, keeping the forecast close to the stronger prior. Across the full evaluation, Table 2 shows that learning intervention improves MSE/MAE by 1.9%/2.2% relative to freezing its parameters. We will complement this result with all-origin gate distributions and paired prior-to-final error differences. Additional examples will use the main expert pool and fixed chronological positions.

Q4. How are temporal coverage and information leakage handled?

R4. Forecasts use stride one, and a prediction issued at ss becomes eligible for feedback at s+Hs+H. Successive overlapping windows can therefore produce consecutive updates. We will report eligible reveal counts alongside nonoverlapping target-block counts.

Appendix B specifies chronological splits, training-only scaling, and text filtering by recorded availability, excluding target-overlapping items. The evaluation includes monthly series, weekly Climate and Energy, and daily Environment. Our empirical conclusions concern these eight Time-MMD domains. We will specify the exact LLM version and available training-cutoff documentation to make the distinction between supplied-context timing and pretraining exposure explicit.

Q5. How will uncertainty in small performance differences be assessed?

R5. We will report paired temporal-block uncertainty estimates using unrounded origin-level losses, together with variability across repeated generations. Temporal blocks account for dependence between overlapping forecast windows. This analysis will evaluate the reported margins under the same forecast origins and scoring rules.

We appreciate the detailed feedback. These clarifications and targeted comparisons will make the contribution of each learned decision and the evaluation conditions more transparent in the revised manuscript.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.