acceptodds
Under review as a conference paper at ICLR 2027

Context2Action: An Agentic Framework for Real-World-Aware Time-Series Forecasting

Abstract

Time series are shaped not only by their historical dynamics but also by real-world contexts such as holidays, promotions, weather, news, and unexpected disruptions. Existing methods either focus primarily on numerical patterns or incorporate contextual information through end-to-end prediction, without explicitly determining how such contexts should influence forecasting decisions. We propose Context2Action, an agentic framework that translates real-world contexts into explicit and controllable forecasting actions. By analyzing the historical errors of numerical predictors, Context2Action learns how contextual factors are associated with both localized changes and broader temporal deviations. Retrieval-augmented LLM experts then adapt relevant historical experiences to the current forecasting state and generate structured forecast revisions, while a hierarchical decision mechanism filters unsupported actions. The framework is optimized through supervised fine-tuning and forecasting-oriented reinforcement learning, aligning contextual reasoning with numerical forecasting utility. Experiments on four context-rich datasets show that Context2Action outperforms strong numerical, LLM-based, and context-aware forecasting baselines. It also generalizes across heterogeneous numerical predictors without additional expert training, demonstrating its effectiveness and adaptability for real-world-context-aware time series forecasting.

Primary area:
learning on time series and dynamical systems
Keywords:
Time series prediction

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_QQcdOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Context2Action, an agentic framework for real-world-context-aware time-series forecasting. The method constructs local and global context memories based on historical forecasting residuals and employs two retrieval-augmented LLM experts to generate structured adjustments to the predictions of a frozen numerical forecasting model. The framework further incorporates a hierarchical decision mechanism and uses supervised fine-tuning and GRPO to optimize forecasting actions. Experiments on four datasets demonstrate improvements over numerical and context-aware forecasting baselines.

Critical strengths
  1. The paper presents a practical approach to incorporating real-world contextual information into time-series forecasting through explicit and interpretable forecast adjustments, rather than directly generating predictions.
  2. The proposed framework achieves improvements over numerical and context-aware baselines on four datasets. The ablation studies and transferability experiments further demonstrate the potential effectiveness and adaptability of the framework.
Critical weaknesses
  1. The framework mainly combines existing techniques, including residual correction, retrieval-augmented LLMs, and SFT/GRPO. The methodological novelty beyond their integration is not sufficiently established.
  2. Historical forecasting residuals do not necessarily reflect the effects of real-world contexts, as they may also arise from model errors or unobserved factors. The reliability of using residuals as context-action supervision remains unclear.
  3. The proposed local adjustments and global level/trend corrections rely on relatively simple predefined operations. It is unclear whether these actions can adequately capture complex, nonlinear, or interacting contextual effects.
  4. The framework assumes that relevant contextual information is available at prediction time. However, the availability of future weather, news, or unexpected events is not sufficiently clarified, raising concerns about practical applicability and evaluation validity.
  5. The largest improvement comes from M5, where contextual events and their effects are synthetically constructed. Furthermore, the same event mechanisms appear in both training and testing, making the claimed generalization to open-world contexts less convincing.
  6. The experiments lack strong lightweight baselines, such as context-conditioned residual regression or simpler retrieval-based correction methods. It is therefore difficult to determine whether the complex agentic framework is necessary.
  7. Although several ablations are provided, the individual contributions of the No-Action mechanism, contextual information quality, and structured action design are not sufficiently isolated. Additional controlled experiments are needed to establish the effectiveness of the core components.
Questions for authors
  1. How does the method distinguish context-induced deviations from ordinary forecasting errors or unobserved factors? What evidence supports the reliability of residuals as supervision for context-action learning?
  2. How is the availability of contextual information ensured at prediction time, particularly for future weather, news, and unexpected events? Are all contextual inputs strictly restricted to information available at the forecasting origin?
  3. Given that M5 uses synthetically constructed events and shares event mechanisms between training and testing, how well does the framework generalize to unseen event types or real-world contextual disruptions?
  4. What is the key methodological novelty beyond combining residual correction, retrieval-augmented LLMs, and SFT/GRPO? Why is an agentic framework necessary compared with simpler context-aware correction methods?
3 comments
AuthorsOpenReview ↗

Response to Reviewer QQcd (1/3)

Dear Reviewer QQcd,

Thank you for your careful review and constructive feedback. We appreciate your recognition of the practical, interpretable forecast-revision formulation and the consistent improvements across four datasets. We address all weaknesses and questions below with clarifications and additional controlled experiments.

1. Reliability of Residual-Grounded Supervision

Your Concern (W2 & Q1): Forecasting residuals may reflect model errors or unobserved factors rather than contextual effects. How does the method distinguish context-induced deviations, and what supports the reliability of residual-based supervision?

We agree that an individual residual should not be interpreted as a causal estimate of a contextual effect. Context2Action instead treats residuals as noisy empirical evidence and controls their use through the following design:

  1. Temporal and semantic grounding. Residuals are paired with timestamped contexts, and local supervision is derived from residual patterns aligned with the context interval.
  2. Multiple-case comparison. The expert retrieves several semantically related historical cases rather than relying on one residual observation.
  3. Forecasting-utility optimization. GRPO rewards actions according to their actual error reduction on training instances, suppressing unreliable context–residual associations.
  4. No-Action decision. A well-formed action is rejected when its timing, duration, or magnitude is insufficiently supported.

The ablations provide supporting evidence. Random retrieval consistently underperforms relevant retrieval and even harms Bitcoin, while directly using the nearest case remains inferior to multi-candidate reasoning.

VariantM5ElectricityNYC TaxiBitcoin
Random Retrieval1.2451.8545.9973.175
Top-1 Retrieval1.1371.7485.9582.776
Context2Action1.0551.6335.9452.687

Thus, the method does not assume that every residual is context-induced. It learns which historical evidence is useful for forecast revision. We will clarify this distinction and avoid causal wording in the revision.

2. Availability of Contextual Information

Your Concern (W4 & Q2): It is unclear whether future weather, news, and unexpected events are available at prediction time.

Our protocol strictly restricts contextual inputs to information that would reasonably be available when the forecast is issued. No realized future context or post-origin information is used.

DatasetContext available at forecast issuance
M5Announced promotions, holidays, recalls, operational disruptions, and weather alerts/forecasts
ElectricityWeather forecasts, workday/holiday indicators, and previously published public news
NYC TaxiAnnounced venue events, scheduled times, event types, and available cancellation notices
BitcoinFinancial and cryptocurrency news published before the forecast origin

Unexpected events are included only when an announcement or alert is already available; otherwise, they are not provided to the model. Weather inputs are forecasts or alerts rather than realized observations. In Figure 3(d), all Bitcoin news items were published before the forecast origin, and the model estimates their possible carry-over effects.

We will add the following explicit temporal-cutoff statement:

For every instance, all contextual inputs must be available by the forecast origin. Information published or realized afterward, together with future observations and residuals, is strictly excluded.

This protocol is consistent with established context-aware forecasting settings, such as EventTSF [1] and the M5 [2] forecasting task, which use scheduled events and other covariates available when predictions are issued.

References

[1] Y. Ge, M. Jin, Y. Zhao, H. Li, B. Du, C. Xu, and S. Pan. “EventTSF: Event-Aware Non-Stationary Time Series Forecasting.” Proceedings of the Thirty-Fifth International Joint Conference on Artificial Intelligence, pp. 4313–4321, 2026.

[2] S. Makridakis, E. Spiliotis, and V. Assimakopoulos. “M5 Accuracy Competition: Results, Findings, and Conclusions.” International Journal of Forecasting, vol. 38, no. 4, pp. 1346–1364, 2022.

AuthorsOpenReview ↗

Response to Reviewer QQcd (2/3)

3. Generalization Beyond M5 Event Mechanisms

Your Concern (W5 & Q3): M5 uses synthetic events and shares event mechanisms between training and testing, making open-world generalization unclear.

M5 is designed as a controlled evaluation in which the underlying numerical series remains realistic while event effects can be precisely examined. Each event mechanism appears through different textual descriptions and independently sampled effects across training and testing. Therefore, the test instances cannot be solved by memorizing a fixed description or correction magnitude.

However, we agree that this setting evaluates generalization across new event occurrences, descriptions, and effect realizations, rather than entirely unseen event categories. We will narrow the corresponding claim accordingly.

Generalization beyond the controlled M5 setting is evaluated on three real-world datasets:

  • Electricity contains weather, calendar, and public-news contexts;
  • NYC Taxi contains real venue events and cancellations;
  • Bitcoin contains timestamped real-world financial news.

Context2Action improves the strongest baselines on all three datasets. These results provide complementary real-world evidence without relying on the M5 event-generation process.

4. Core Novelty and the Need for an Agentic Framework

Your Concern (W1, W6 & Q4): The framework combines residual correction, retrieval-augmented LLMs, and SFT/GRPO. What is the methodological novelty, and why is an agentic framework needed instead of a simpler correction method?

The novelty does not lie in introducing residual modeling, retrieval, or GRPO individually. It lies in a new context-to-action forecast-revision paradigm.

Traditional forecasting models primarily extrapolate numerical patterns, while real-world context is heterogeneous, open-form, and does not directly specify a numerical correction. Using it requires several coupled decisions:

  1. determine whether a context is relevant;
  2. retrieve comparable historical evidence;
  3. infer the affected interval, direction, and magnitude;
  4. coordinate local effects with overall level/trend calibration;
  5. reject an unsupported revision.

Context2Action formulates these decisions as explicit, inspectable actions instead of directly generating a new forecast. The numerical predictor remains frozen, while the experts translate contextual evidence into revisions that can transfer across forecasting backbones. The agentic formulation is therefore used for evidence comparison, contextual reasoning, structured action generation, and acceptance decisions, rather than merely as an alternative implementation of residual regression.

4.1 Direct Residual-Transfer Control

We additionally compare with lightweight retrieval-based corrections using the same base forecasts, contexts, retrieval pool, and test windows.

MethodMAE (×10³)MSE (×10⁶)RMSE (×10³)WMAPE (%)MAE Reduction
Base Predictor1.4995.8262.41411.09–
Top-1 Direct Transfer1.6917.1362.67112.51−12.82%
Top-3 Direct Transfer1.4895.5632.35911.020.66%
Top-5 Direct Transfer1.4435.3612.31510.673.76%
Context2Action1.0553.1331.7707.8129.59%

Directly transferring the nearest residual is harmful, while averaging multiple residuals yields only limited improvement. This result indicates that the gain primarily comes from adapting historical evidence to the current context and forecast state, rather than simply reusing numerical corrections.

AuthorsOpenReview ↗

Response to Reviewer QQcd (3/3)

5. Expressiveness of Structured Actions

Your Concern (W3): The predefined local and global operations may not capture complex, nonlinear, or interacting contextual effects.

The constrained action space is intentional: it preserves the temporal structure learned by the numerical predictor and makes every revision inspectable.

  • Local actions independently specify the start time, duration, and signed magnitude of each contextual effect.
  • Multiple overlapping contexts are considered jointly, allowing their interactions to affect the final timing and magnitude.
  • The Global Expert further adjusts systematic level and trend deviations after local revision.

Therefore, a forecasting window can contain multiple localized actions together with a global calibration, producing a piecewise and context-dependent revision rather than a single uniform correction.

We nevertheless agree that the current action space is not intended to represent every possible nonlinear transformation. This is a deliberate interpretability–expressiveness trade-off, and richer nonlinear action bases are a valuable direction for future work. We will clarify this scope rather than claiming unrestricted expressiveness.

6. Isolating No-Action and Component Contributions

Your Concern (W7): The individual contributions of No-Action, contextual-information quality, and structured action design are not sufficiently isolated.

6.1 No-Action Ablation

We remove only the acceptance decision while always applying the same valid candidate actions.

StageAlways ApplyWith No-ActionMAE ReductionRejection Rate
Local Expert1.1921.1374.61%6.25%
Global Expert1.0771.0552.04%9.38%

This confirms that No-Action rejects a small but consequential set of harmful revisions.

For example, one candidate proposed a constant +27.1% adjustment over five days, although the retrieved evidence declined from +28.0% and +29.4% to +6.1%, +4.2%, and −1.8%. The expert selected NO_ACTION. Evaluation performed afterward showed that this decision reduced MAE by 12.12% relative to applying the candidate.

6.2 Context Quality and Structured Revision

  • Random Retrieval vs. Top-1 vs. Context2Action isolates the importance of contextual relevance and multi-case evidence selection.
  • Direct residual transfer vs. Context2Action tests whether retrieved numerical effects alone explain the improvement.
  • Always Apply vs. No-Action isolates the acceptance mechanism.
  • Local Only, Global Only, and the complete framework distinguish localized contextual revision from level/trend calibration.

Together, these controls show that the improvement arises from the complete context-to-action process—relevant evidence retrieval, structured revision, and utility-aware acceptance—rather than from residual correction or output-format constraints alone.

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

This paper proposes Context2Action, an agentic framework for incorporating real-world contexts into time-series forecasting through structured forecast revisions. The framework first generates base predictions using a frozen numerical forecasting model and constructs local and global context memories from historical forecasting residuals. Then two retrieval-augmented LLM experts generate local and global forecast adjustments. Experiments on four datasets show improvements over various baselines.

Critical strengths
  1. The paper introduced a new way to improve time series forecasting to connect real-world contextual information with forecast revisions rather than directly generating new predictions.

  2. The proposed framework is reasonably designed, combining historical residual-based memories, local and global forecasting adjustments. These components provide a structured and potentially interpretable way to modify existing numerical predictions.

  3. The paper includes experiments on four datasets covering different domains and compares the proposed approach with numerical, LLM-based, and context-aware forecasting baselines. Different LLM-backbones were also assessed on M5.

Critical weaknesses
  1. The paper does not directly examine whether the improvements depend on the semantic content of the contextual information. Since the framework also learns forecast corrections from historical residuals, it is unclear how much of the improvement comes from understanding textual contexts rather than learning numerical correction patterns.

  2. Context2Action achieves substantial improvements on M5 and Electricity. But the reported MAE improvements over the strongest baselines are only 0.75% on NYC Taxi and 1.54% on Bitcoin. This shows inconsistency of the method's effectiveness across different datasets.

  3. While the paper explicitly prevents information leakage through historical retrieval and residual access, it doesn't state whether external contexts, such as news, weather, and event cancellations, are available at the forecasting origin.

Questions for authors
  1. How do the authors ensure that if the forecasting improvement really comes from the semantic content of the contextual information?

  2. Is all contextual information, particularly news, weather, and event cancellations, available at the forecasting origin rather than after the prediction period has started? (Data-leakage problem)

  3. Could the authors further isolate the contributions of the No-Action decision mechanism and the different GRPO reward components?

4 comments
AuthorsOpenReview ↗

Response to Reviewer Rvmi [1/2]

Dear Reviewer Rvmi,

Thank you for your careful review and constructive feedback. We appreciate your recognition of the soundness of our paper, the context-to-revision formulation, and the structured and interpretable design of Context2Action. We address all weaknesses and questions below with additional controlled experiments and analyses.

1. Contribution of Contextual Semantics

Your Concern (W1 & Q1): The improvement may arise from numerical correction patterns in historical residuals rather than understanding the semantic content of contextual information. How can the contribution of contextual semantics be verified?

Thank you for raising this important question. To isolate numerical residual reuse, we introduce direct residual-transfer baselines under the same base forecast and retrieval pool.

1.1 Experiment Setup

For each M5 instance, we retrieve historical cases using the same context-retrieval mechanism and directly apply their residuals:

  • Top-1 Transfer: applies the residual of the most similar case.
  • Top-3/5 Average: averages the residuals of the three or five most similar cases.
  • Context2Action: interprets the context and retrieved evidence to generate a case-specific action.

1.2 Results

MethodMAE ↓MSE ↓RMSE ↓WMAPE ↓MAE Improvement
Base Forecast1.4995.8262.41411.09–
Top-1 Transfer1.6917.1362.67112.51−12.82%
Top-3 Average1.4895.5632.35911.02+0.66%
Top-5 Average1.4435.3612.31510.67+3.76%
Context2Action1.0553.1331.7707.81+29.59%

Direct residual reuse is unreliable, while averaging several residuals provides only limited gains. Moreover, the expert selects a lower-ranked historical case in 37.57% of non-SNAP instances rather than mechanically using the top retrieval result.

These results show that numerical correction patterns alone cannot explain the improvement. The substantial gap supports the contribution of interpreting contextual semantics and adapting historical evidence to the current forecasting state.

2. Performance Variation across Datasets

Your Concern (W2): Context2Action achieves substantial improvements on M5 and Electricity, but its overall MAE improvements over the strongest baselines are only 0.75% on NYC Taxi and 1.54% on Bitcoin, suggesting inconsistent effectiveness.

The overall metrics average all prediction points, whereas contextual events often affect only part of the forecasting horizon. Consequently, improvements can be diluted by unaffected points, particularly on NYC Taxi, where relevant events are sparse.

To examine performance where the model determines that context has an effect, we additionally evaluate all methods on the same points receiving non-zero Local actions.

2.1 Results on Context-Affected Points

DatasetEvaluated PointsStrongest BaselineBaseline MAEContext2Action MAEMAE Reduction
NYC Taxi9 points in 7/18 windowsTime-LLM21.89618.66214.77%
Bitcoin640 points in 122/127 windowsAlphaCast3.1732.8859.08%

This diagnostic complements rather than replaces the full-horizon evaluation. It shows that Context2Action produces clearer gains at context-affected points while preserving the numerical forecast elsewhere, explaining why the overall improvement can appear modest when contextual effects are localized.

AuthorsOpenReview ↗

Response to Reviewer Rvmi (2/2)

3. Availability of Contextual Information

Your Concern (W3 & Q2): The paper does not clearly state whether news, weather, and event cancellations are available at the forecasting origin, raising a potential data-leakage concern.

Our protocol restricts the model to contextual information that would reasonably be available when the forecast is issued. Information observed or released during the prediction window is excluded.

3.1 Context Used in Each Dataset

DatasetContext available at forecast issuance
M5Announced promotions, holidays, calendar events, and retail operational information
ElectricityCalendar information, weather forecasts, and relevant public information
NYC TaxiPreviously announced venue events, including event type, scheduled time, and status updates
BitcoinNews and public announcements published before the forecast origin

Weather inputs refer to forecasts available at the origin rather than subsequently observed weather. Future targets, realized conditions, and post-event reports are never provided to the model.

This protocol follows established forecasting practice. EventTSF [1] considers previously announced holidays and promotions, while the M5 competition [2] provides promotion, price, calendar, and special-day information. Such inputs represent the information available to a real forecaster and do not reveal the prediction outcomes.

We will explicitly clarify:

At each forecast origin, Context2Action uses only contextual information that would be available when the forecast is issued. No target values, realized conditions, or information released during the prediction window are used.

4. No-Action and GRPO Reward Components

Your Concern (Q3): Could the authors further isolate the contributions of the No-Action decision mechanism and the different GRPO reward components?

4.1 No-Action Mechanism

No-Action is not the rejection of malformed outputs. After generating a structurally valid candidate action, each Context Expert explicitly outputs:

  • APPLY: execute the candidate action.
  • NO_ACTION: retain the current prediction.

We compare this mechanism with always applying the same valid candidate actions.

StageAlways ApplyWith No-ActionMAE ReductionRejection Rate
Local Context Expert1.1921.1374.61%6.25%
Global Context Expert1.0771.0552.04%9.38%

No-Action improves both stages while rejecting only a small fraction of candidates, demonstrating selective filtering rather than conservative suppression.

In one M5 case, the Local Expert proposed a valid 27.1% five-day increase, although the retrieved effect weakened from +28.0% and +29.4% in the first two days to +6.1%, +4.2%, and −1.8% afterward. The expert selected NO_ACTION. After the true sales became available solely for evaluation, this decision reduced MAE by 12.12% compared with applying the candidate.

4.2 GRPO Reward Components

The Local Expert combines format, context-coverage, retrieval, and forecasting rewards. The Global Expert uses format, retrieval, and forecasting rewards because it performs window-level calibration rather than event-level reasoning.

Reward curves figure

The reward curves show that format and coverage rewards stabilize early, establishing valid and complete outputs. Retrieval and forecasting rewards continue improving afterward, progressively aligning evidence selection and generated actions with forecasting utility. We will add this visualization and the corresponding reward definitions to the revised appendix.

References

[1] Y. Ge, M. Jin, Y. Zhao, H. Li, B. Du, C. Xu, and S. Pan. “EventTSF: Event-Aware Non-Stationary Time Series Forecasting.” Proceedings of the Thirty-Fifth International Joint Conference on Artificial Intelligence, pp. 4313–4321, 2026.

[2] S. Makridakis, E. Spiliotis, and V. Assimakopoulos. “M5 Accuracy Competition: Results, Findings, and Conclusions.” International Journal of Forecasting, vol. 38, no. 4, pp. 1346–1364, 2022.

The clarification resolves my concern about the availability of context at forecast time, and the always-apply comparison addresses my question about No-Action. My main concern remains the contribution of textual semantics. The M5 results show that Context2Action outperforms copying or averaging retrieved residuals. However, the trained experts can also adjust the timing and magnitude of a correction using numerical patterns in those residuals. Hence the contribution of textual semantics is still not isolated. The gains on NYC Taxi and Bitcoin remain modest overall. The larger gains are reported on points selected by the model itself, including only nine points on NYC Taxi. The GRPO reward curves also leave each reward component’s contribution to forecast accuracy unclear. For now, I will maintain my rating of 2.

AuthorsOpenReview ↗

Thanks for your feedback!

Dear Reviewer Rvmi,

Thank you for your constructive follow-up. We are glad that our rebuttal addressed your concerns about context availability and No-Action. Your question about the contribution of textual semantics to the model’s reasoning prompted us to test it directly. In Context2Action, retrieval supplies candidate historical cases. Our context experts interpret the current event to decide which case is most relevant and how its evidence should inform the forecast revision.

1. Experimental strategy

On M5, we changed only one event description at a time, keeping the base forecast, observed history, event dates, selected historical case, and its residuals fixed. We replaced sales-increasing events with sales-decreasing events and vice versa. Testing both directions checks whether the expert can revise its forecast upward or downward as the meaning of the event changes. We used the same expert and prompt template.

2. Results

The table shows how often the expert reversed the direction of its proposed adjustment when the expected effect of the event reversed, given the same historical evidence. Across 190 event–window pairs from 93 M5 forecast windows, the proposed adjustment reversed sign in 176 pairs (92.6%).

Change in expected sales impactProposed adjustments that reversed sign
Positive to negative99/99 (100.0%)
Negative to positive77/91 (84.6%)
Overall176/190 (92.6%)

3. Findings

The two exampels below illustrate how the expert responds when the current event description and the retrieved residual point in opposite directions.

Original event → Replacement eventRetrieved event-day residualProposed adjustment before → after
Household promotion → Product withdrawal with checkout restrictions+37.5%+31.3% → −18.3%
Payment outage → Anniversary promotion with checkout discounts and extended hours−33.2%−18.9% → +12.5%

In both cases, the expert kept the proposed date and one-day duration and reversed the adjustment in the direction implied by the new event. This provides evidence that the expert uses event semantics to determine the direction of a revision. The replacement events are hypothetical, so the experiment tests the reasoning behind the proposed actions.

Thank you again for your thoughtful review and constructive discussion. We would welcome your perspective on whether the additional evidence informs your overall assessment.

Best regards,
The Authors of Submission #46235

Reviewer_28DCOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper uses historical forecast errors to train two LLM experts: one adjusts forecasts around local events, and the other corrects overall level and trend. It turns contextual information into explicit corrections that can be reused across numerical forecasters. The evaluation includes forecasting comparisons, retrieval and training ablations, and transfer of fixed experts to other predictors on a controlled M5 dataset.

Critical strengths
  • The local/global action design makes forecast revisions easy to inspect. The ablations provide useful evidence that relevant retrieval and forecasting-reward training improve this design (Table 2).
  • Transferring fixed experts to other predictors on controlled M5 tests whether the corrections depend on one base model. This supports the idea of a reusable correction module, within the setting studied (Table 3).
Critical weaknesses
  • The paper restricts historical retrieval, but does not clearly establish when current news, weather, and cancellation information becomes available. This leaves uncertainty about whether all inputs would be available when the forecast is made (A.3; C.1).
  • The experiments do not isolate what the reasoning and action-rejection mechanisms add. M5 reuses injected event mechanisms across training and testing, and there is no direct residual-transfer control. The schemas and ablations also leave unclear how unsupported actions are rejected, beyond rejecting malformed outputs (Table 2; C.1; pp. 14–18).
Questions for authors
  1. What contextual information is available at each forecast origin? Please clarify publication times for Figure 3(d) and whether weather inputs are forecasts or realized observations. If any inputs arrive after the forecast origin, how does excluding them affect the results?
  2. How does the method compare with directly transferring retrieved residual effects, with scaling fitted on validation data? A comparison using the same base forecast, context, retrieval pool, and training examples on M5 and Electricity would help establish what the learned reasoning adds.
  3. How is No-Action represented? Compared with always applying the same valid candidate actions, does the learned acceptance decision reduce harmful revisions?
2 comments
AuthorsOpenReview ↗

Response to Reviewer 28DC [1/2]

Dear Reviewer 28DC,

Thank you for your careful review and constructive feedback. We appreciate your recognition of the inspectable local/global action design and the transferability of the learned revision module. We address all concerns below through protocol clarification and additional controlled experiments.

1. Availability of Contextual Information

Your Concern (W1 & Q1): It is unclear whether news, weather, and cancellation information is available when the forecast is issued, particularly for Figure 3(d).

We apologize that the temporal cutoff was not stated clearly. Our protocol only uses contextual information that would reasonably be available when the forecast is issued. No realized outcome or context published after the forecast origin is used.

1.1 Contexts Used in Each Dataset

DatasetContext available at forecast issuance
M5Announced promotions, holidays, product recalls, operational disruptions, and weather alerts/forecasts
ElectricityWeather forecasts, workday/holiday indicators, and previously published public news
NYC TaxiAnnounced venue events, scheduled times, event types, and cancellation notices available before issuance
BitcoinFinancial and cryptocurrency news published before the forecast origin

For Figure 3(d), all displayed Bitcoin news items were published before the forecast origin. Their timestamps indicate publication time, and the model estimates their potential carry-over effects on the subsequent forecasting window. Similarly, weather inputs are forecasts or alerts available at issuance, rather than realized weather observations.

Therefore, excluding post-origin information does not change the reported results because such information is not included in the first place. We will explicitly add the following temporal-cutoff statement:

For every forecasting instance, contextual inputs must be available by the forecast origin. Contexts published or realized afterward, as well as future numerical outcomes, are strictly excluded.

This protocol is also consistent with established context-aware forecasting settings, including EventTSF [1] and the M5 [2] forecasting task, where scheduled events and calendar-related covariates available at prediction time are used to forecast subsequent observations.

2. Reasoning Beyond Direct Residual Transfer

Your Concern (W2 & Q2): The experiments do not isolate whether the gains come from learned reasoning or direct reuse of retrieved residual effects. M5 also reuses event mechanisms across training and testing.

Following this suggestion, we added a direct residual-transfer control on M5 using the same base forecasts, contexts, retrieval pool, and evaluation windows as Context2Action.

2.1 Experiment Setup

For each M5 instance, we retrieve historical cases using the same context-retrieval mechanism and directly apply their residuals:

  • Top-1 Transfer: applies the residual of the most similar case.
  • Top-3/5 Average: averages the residuals of the three or five most similar cases.
  • Context2Action: interprets the context and retrieved evidence to generate a case-specific action.

2.2 Results

MethodMAE ↓MSE ↓RMSE ↓WMAPE ↓MAE Improvement
Base Forecast1.4995.8262.41411.09–
Top-1 Transfer1.6917.1362.67112.51−12.82%
Top-3 Average1.4895.5632.35911.02+0.66%
Top-5 Average1.4435.3612.31510.67+3.76%
Context2Action1.0553.1331.7707.81+29.59%

2.3 Findings

Direct residual reuse is unreliable, while averaging several residuals provides only limited gains. Moreover, the expert selects a lower-ranked historical case in 37.57% of non-SNAP instances rather than mechanically using the top retrieval result.

These results show that numerical correction patterns alone cannot explain the improvement. The substantial gap supports the contribution of interpreting contextual semantics and adapting historical evidence to the current forecasting state.

AuthorsOpenReview ↗

Response to Reviewer 28DC [2/2]

3. Representation and Effect of No-Action

Your Concern (W2 & Q3): It is unclear how unsupported but well-formed actions are rejected and whether No-Action reduces harmful revisions.

No-Action is not limited to malformed-output rejection. After a Local or Global Expert proposes a valid candidate action, the corresponding acceptance decision evaluates whether its timing, duration, direction, and magnitude are sufficiently supported by the retrieved historical evidence.

The decision is represented as either APPLY or NO_ACTION:

  • Local NO_ACTION: retain the base forecast;
  • Global NO_ACTION: retain the Local Expert output.

3.1 Always-Apply Comparison

StageAlways ApplyWith No-ActionMAE ReductionRejection Rate
Local Expert1.1921.1374.61%6.25%
Global Expert1.0771.0552.04%9.38%

The comparison applies exactly the same valid candidate actions and only removes the acceptance decision. The improvements therefore directly measure the value of rejecting harmful revisions.

3.2 Representative Case

In one M5 HOUSEHOLD window, the candidate proposed a constant +27.1% adjustment for five days. However, the retrieved residual evidence decreased from +28.0% and +29.4% on the first two days to +6.1%, +4.2%, and −1.8% afterward. The acceptance decision consequently selected NO_ACTION because the proposed duration and magnitude were insufficiently supported.

After the decision was made, evaluation against the held-out observations showed that rejecting this candidate reduced MAE by 12.12% relative to applying it. Importantly, the decision itself had no access to these future observations.

These results clarify that No-Action serves as an evidence-based safeguard against validly formatted but unsupported forecast revisions.

References

[1] Y. Ge, M. Jin, Y. Zhao, H. Li, B. Du, C. Xu, and S. Pan. “EventTSF: Event-Aware Non-Stationary Time Series Forecasting.” Proceedings of the Thirty-Fifth International Joint Conference on Artificial Intelligence, pp. 4313–4321, 2026.

[2] S. Makridakis, E. Spiliotis, and V. Assimakopoulos. “M5 Accuracy Competition: Results, Findings, and Conclusions.” International Journal of Forecasting, vol. 38, no. 4, pp. 1346–1364, 2022.

Reviewer_GtdeOpenReview ↗
Rating
3
Confidence
4
Informativeness
3(+1)
Presentation
2
Soundness
3
Read review
Paper summary

This paper presents Context2Action, an agentic framework that reformulates real-world-context-aware time-series forecasting as a post-hoc forecast revision problem. It first obtains a base forecast from a frozen numerical predictor and then employs a Local Context Expert and a Global Context Expert to address event-specific deviations and window-level biases, respectively. The framework constructs a dual context memory grounded in historical forecasting residuals and combines retrieval augmentation, supervised fine-tuning, and GRPO to generate explicit actions specifying when, where, and by how much a forecast should be revised. Experiments cover retail, electricity, transportation, and financial forecasting, with additional analyses of component ablations, predictor transfer, LLM backbones, and representative cases.

Critical strengths
  • The most interesting contribution is the reformulation of context-aware forecasting from directly integrating contexts into a predictive model to determining whether and how an existing forecast should be revised. This separation between numerical prediction and contextual reasoning produces actions with explicit timing, duration, direction, and magnitude, making the resulting revisions easier to inspect.
  • The residual-grounded dual memory aligns naturally with the two-stage revision process. The Local Expert handles localized deviations associated with individual events, while the Global Expert adjusts broader level and trend biases. Compared with asking an LLM to generate a complete forecast, this design preserves the numerical predictor and constrains the role of language-based reasoning.
  • The inclusion of No-Action is practically meaningful. Contextual information is not uniformly useful, and retaining the numerical forecast when the available evidence is insufficient is preferable to enforcing a potentially harmful correction.
  • The empirical evaluation goes beyond comparisons with numerical, LLM-based, and context-aware baselines. The ablations, LLM-backbone analysis, and predictor-transfer experiments provide useful evidence for the modularity of the framework. In particular, directly transferring PatchTST-trained experts to other numerical predictors supports the claimed role of Context2Action as a reusable correction module.
Critical weaknesses
  • The current action space mainly consists of interval-wise local scaling, global level shifts, and linear trend corrections. This design makes forecast revisions controllable, but its ability to represent delayed, nonlinear, or interacting contextual effects is not sufficiently discussed.
  • The method relies on historical residual memories and retrieved cases to generate forecasting actions. Although the paper includes ablations on retrieval strategies, it remains unclear how the experts reconcile retrieved cases that are weakly related or exhibit conflicting residual patterns.
Questions for authors
  • Can the current structured actions represent delayed, nonlinear, or interacting contextual effects?
  • Could the authors separately report the proportions and forecasting performance of accepted actions and No-Action decisions to clarify the performance differences across datasets?
3 comments
AuthorsOpenReview ↗

Response to Reviewe Gtde

Dear Reviewer Gtde,

Thank you for your careful and insightful review. We especially appreciate your recognition of the context-to-revision formulation, the interpretable dual-expert design, the practical role of No-Action, and the framework’s modularity. We address your concerns below.

1. Expressiveness of Structured Actions

Your Concern (W1 & Q1): Can the current action space represent delayed, nonlinear, or interacting contextual effects?

The current action space supports these effects to different extents:

  1. Delayed effects. The Local Expert independently predicts the action start time and duration. Therefore, an action may begin after the contextual event rather than at its timestamp.
  2. Interacting effects. All contexts within a window are jointly considered when generating local actions. When their affected intervals overlap, the expert reasons over their combined evidence rather than treating each context independently.
  3. Nonlinear effects. Multiple local actions combined with global level and trend calibration can produce piecewise, context-dependent revisions. However, a single local action applies one magnitude within its interval and therefore cannot represent an arbitrary nonlinear response.

This constrained design is intentional: it preserves the temporal structure learned by the numerical predictor and keeps each revision inspectable. We agree that it trades unrestricted expressiveness for controllability. We will clarify this scope and discuss richer nonlinear action bases as future work.

2. Reconciling Weak or Conflicting Retrieved Cases

Your Concern (W2): It is unclear how the experts handle retrieved cases that are weakly related or exhibit conflicting residual patterns.

The experts do not directly average all retrieved residuals. Instead, they:

  1. retrieve multiple semantically relevant historical cases;
  2. compare their context, timing, residual pattern, and compatibility with the current forecast;
  3. select the most informative evidence and generate a new action conditioned on the current state;
  4. choose NO_ACTION when the evidence does not sufficiently support the proposed timing or magnitude.

GRPO further optimizes evidence selection according to forecasting utility. Among nontrivial M5 actions, the expert selects a lower-ranked case rather than the top-1 retrieval in 37.57% of instances. Moreover, Top-1 Retrieval remains inferior to the complete model on all four datasets, indicating that the experts do not simply follow the nearest case.

A representative rejection case further illustrates conflict handling. A candidate proposed a constant +27.1% adjustment for five days, whereas the retrieved residual evidence decreased from +28.0% and +29.4% to +6.1%, +4.2%, and −1.8%. Because the duration and magnitude were not consistently supported, the expert selected NO_ACTION.

3. Accepted Actions and No-Action Decisions

Your Concern (Q2): Please report the proportions and forecasting performance of accepted actions and No-Action decisions.

We added an always-apply control on M5. It generates exactly the same valid candidate actions but removes the acceptance decision.

StageAcceptedNo-ActionAlways ApplyWith No-ActionMAE Reduction
Local Expert93.75%6.25%1.1921.1374.61%
Global Expert90.62%9.38%1.0771.0552.04%

No-Action is used selectively rather than conservatively suppressing revisions. Although only 6.25%–9.38% of candidates are rejected, retaining the current forecast for these cases consistently reduces MAE.

We use the always-apply comparison because directly comparing conditional MAE between accepted and rejected subsets would be selection-biased: No-Action is naturally assigned to more ambiguous cases, and the two subsets are not equally difficult. The controlled comparison above holds the candidate actions and instances fixed and therefore isolates the contribution of the acceptance decision.

For the representative case described above, evaluation after the decision showed that rejecting the unsupported action reduced MAE by 12.12% relative to applying it. The expert had no access to the future observations when making this decision. We will add the decision proportions, controlled performance comparison, and representative case to the revised manuscript.

Thank you for the rebuttal. The clarification of the action space and the always-apply comparison address my main questions. I also read the other reviews and noted that several reviewers recognized merits in the approach and raised questions about the experimental evidence. For me, a distinctive contribution is that the framework makes individual forecast revisions open to scrutiny. Even when two methods achieve similar aggregate accuracy, explicit actions let us examine whether the available evidence justifies the timing and magnitude of a revision. I think it is a useful contribution to context-aware forecasting and remain supportive of acceptance.

AuthorsOpenReview ↗

Thanks for your positive response!

Dear Reviewer Gtde,

Thank you for your thoughtful follow-up and for supporting our paper. We’re glad to hear that our rebuttal addressed your main questions, and we appreciate your highlighting the value of being able to inspect individual forecast revisions.

Best regards, The Authors of Submission #46235

Comments on the paper

1 comment
AuthorsOpenReview ↗

Final Remarks by Authors

Dear Area Chair, Senior Area Chair, Program Chairs, and Reviewers,

Thank you for your careful reviews and constructive feedback. When a retailer announces a promotion, the sales history may still look the same, but the forecast should account for the new information. Context2Action proposes an agent-based solution to turn that information into a concrete decision about whether to revise the forecast, over which interval, and by how much. We evaluate the framework across four datasets and use controlled experiments and case studies to examine how it makes those decisions.


Key strengths noted by the reviewers

  • S1. Forecast revision as an explicit decision. Context determines whether and how to revise an existing forecast. (Gtde, Rvmi, QQcd)
  • S2. Inspectable local and global actions. The framework distinguishes event-specific adjustments from changes to the overall level and trend. (Gtde, Rvmi, 28DC)
  • S3. Reuse across numerical forecasters. The same PatchTST-trained experts improve three other forecasters on M5 without retraining (Table 3). (Gtde, 28DC)
  • S4. An explicit option to retain the forecast. No-Action lets the framework preserve the current prediction when the evidence for revision is weak. (Gtde)
  • S5. Evaluation across four domains. The experiments cover retail, electricity, transportation, and finance and include comparisons with numerical and context-aware baselines. (Rvmi, QQcd)

Key clarifications and additional evidence

We responded point by point to questions about the forecasting protocol, the revision mechanism, and the interpretation of the experimental results.

  • What information is available when the forecast is issued? (Rvmi, 28DC, QQcd) We clarified the cutoff for contextual inputs and distinguished when information becomes available from when the event occurs. Rvmi confirmed that this resolved the concern.
  • What can the structured actions represent, and when does No-Action help? (Gtde, Rvmi, 28DC, QQcd) We explained how actions cover delayed and overlapping effects and clarified their limits for nonlinear responses. With the same candidate actions on M5, No-Action reduced MAE by 4.61% at the local stage and 2.04% at the global stage relative to always applying them. Rvmi confirmed that the comparison answered the No-Action question, and Gtde confirmed that the clarification and comparison addressed the main questions.
  • How much does direct residual transfer help, and does event wording change the proposed action? (Rvmi, 28DC, QQcd) On M5, the best direct-transfer control reduced MAE by 3.76% relative to the base forecast, compared with 29.59% for Context2Action. In a separate test with fixed numerical evidence, replacing an event description with one implying the opposite sales effect reversed the proposed action in 176 of 190 pairs (92.6%). This tests the role of event semantics in action generation.
  • Does Context2Action help beyond the controlled M5 setting?(Rvmi, 28DC, QQcd) M5 tests new occurrences of event types seen during training. The real-world evaluations complement this test, with full-test-set MAE reductions of 0.75% on NYC Taxi and 1.54% on Bitcoin relative to the strongest reported baselines.

Review assessments. Gtde and Rvmi each gave 3 for soundness, indicating no substantive concerns about the paper’s soundness. Gtde also gave 3 for informativeness, reporting a substantially deeper understanding of the problem, and reaffirmed support for acceptance after the discussion. 28DC confirmed that the rebuttal answered the original questions and made the paper more self-contained.

In reaching a decision, we hope the Area Chair will consider the practical importance of responding to real-world events in forecasting and our core technical contribution of turning contextual evidence into explicit decisions about whether and how to revise a forecast. The remaining empirical questions concern what drives the reported gains. As we have yet to receive reviewer feedback on the additional controlled experiment, we would be grateful for the Area Chair’s assessment of what this evidence supports.

Thank you again for your consideration and for coordinating the review.

Sincerely,
The Authors of Submission #46235

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.