acceptodds
Withdrawn from ICLR 2027

News-Aware Time Series Forecasting Agent Enhanced by Offline Reinforcement Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Accurate time-series forecasting in event-sensitive environments can benefit from external news, but semantically relevant news is not necessarily useful for forecasting. We study news use as a forecasting-feedback-driven selection problem. Our framework combines an LLM-based forecaster with a sequential selector that decides whether each candidate article should be kept, dropped, or summarized, conditioned on the numerical history, structured context, and information already retained. The selector is learned from fixed historical trajectories using offline reinforcement learning, with terminal reward defined by improvement in downstream forecasting loss relative to a no-news forecast. We further construct a forecasting-error reflection memory that retains only counterfactual selection corrections that reduce historical forecasting loss. We evaluate the framework on electricity demand, AUD/USD, and Bitcoin forecasting. End-to-end results show that the usefulness of external news and reflection varies across tasks: electricity favors highly conservative information use, whereas selected news provides larger benefits in the financial tasks. In a controlled fixed-forecaster evaluation, changing only the supplied news context materially affects forecasting accuracy, and the best forecasting-feedback-selected context achieves lower RMSE than both no-news and LLM-selected contexts across all three domains. These results show that forecasting utility should be learned from downstream forecasting feedback rather than inferred from semantic relevance alone.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Time-Series Forecasting, Large Language Models, Offline Reinforcement Learning, News Selection, Exogenous Information, Forecasting with External Context

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_vK2pOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies how to select news for time-series forecasting based on predictive usefulness rather than semantic relevance alone. An offline reinforcement learning selector chooses whether to keep, drop, or summarize each article, using forecasting-loss improvement as feedback. A reflection mechanism stores verified action corrections from historical examples and retrieves them to inform later selection. The approach is evaluated on electricity demand, AUD/USD, and Bitcoin, with comparisons among offline RL algorithms, fixed-forecaster context interventions, and a larger electricity experiment. The theoretical appendix reviews conservative policy-evaluation guarantees under idealized CQL assumptions.

Critical strengths

The distinction between relevant news and useful news is worth studying. Learning from downstream forecasting feedback gives the selection problem a concrete objective, and allowing the selector to discard news is sensible when external context adds noise.

Table 3 is an important experiment. Holding the forecaster fixed while changing the supplied context addresses a real confound in the end-to-end comparison. It provides evidence that selection itself can affect forecast quality. I also appreciate that the appendix reports cases where reflection does not help, rather than presenting only favorable examples.

Critical weaknesses

The experiments do not yet establish the benefit of offline RL over a simpler forecasting-utility selector. Comparing DQN, BCQ, and CQL mostly distinguishes RL algorithms; it does not show why supervised utility scoring or a budget-matched greedy selector is insufficient. Appendix B explicitly analyzes idealized CQL policy evaluation, so its guarantees do not resolve this empirical question. The closely related work "From Long News to Accurate Forecast: Importance-Aware Fusion and PRM-Guided Reflection for Time Series Forecasting" [1] also deserves discussion: it learns news compression and selection from forecasting feedback, which makes the incremental role of CQL and correction memory especially important to establish.

The evidence for reflection and the source of the end-to-end gains remain mixed. Table 8 reports zero selected news for the electricity CQL configuration, although Table 2 shows an improvement over No News; this cannot be attributed to supplying more useful news without accounting for differences in forecaster training. In the fixed-forecaster Bitcoin comparison in Table 3, CQL obtains RMSE 5.123, versus 5.233 and 5.230 for the reflection variants, in the table's units. Neither reflection variant improves the larger electricity result in Table 10. These outcomes need not invalidate the approach, but they limit the evidence for reflection as a reliable additional contribution. The financial test sets contain only 50 forecast origins each (Table 4), with no uncertainty estimates to assess the smaller differences.

[1] Liu, M., Kang, Q., Wang, Y., Kai, S., Liang, K., Zhen, H. L., ... & Song, L. (2026). From Long News to Accurate Forecast: Importance-Aware Fusion and PRM-Guided Reflection for Time Series Forecasting. arXiv preprint arXiv:2606.03097.

Questions for authors
  1. Could you clarify exactly how the CQL and No News forecasters differ in training, particularly for the zero-news electricity result? Which improvements can be attributed to context selection, and which depend on adapting the forecaster?

  2. What evidence distinguishes the proposed selector from a simpler utility-trained ranking or greedy selection method under the same news and forecasting budgets? Please also clarify the methodological difference from the related forecast-utility and PRM-based work noted above.

  3. Can you quantify uncertainty in the paired financial results, accounting for temporal dependence, and explain when reflection helps beyond the base selector? The fixed-forecaster and larger-electricity results seem especially useful for answering the latter question.

Reviewer_9HgtOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a news-aware time-series forecasting framework that uses offline reinforcement learning to select useful news for an LLM-based forecaster. The selector makes sequential KEEP, DROP, or SUMMARIZE decisions based on forecasting feedback, and a reflection memory stores historical action corrections that reduce prediction errors. Experiments on electricity demand, exchange rates, and Bitcoin prices show improvements over several news-selection baselines, although the benefits vary across datasets.

Critical strengths

The paper addresses an important problem by distinguishing news relevance from actual forecasting value. Using downstream prediction error to guide news selection provides a clear training objective.

The fixed-forecaster experiment is a useful design choice because it helps separate the effect of selected news from changes in the forecasting model.

Critical weaknesses

The main contribution is not sufficiently validated because the experiments do not compare offline RL against simpler supervised news-ranking methods, contextual bandits, or sparsity-matched selection rules that could optimize the same forecasting objective without sequential RL.

The electricity results raise a fundamental concern because CQL selects zero news on average but substantially outperforms the No News baseline in Table 2, suggesting that changes in forecaster training, rather than the proposed selection mechanism, may explain much of the improvement.

Questions for authors

Can the authors compare CQL with a supervised reward predictor, a contextual bandit, and a simple news-selection heuristic under the same forecasting backbone and context budget to establish whether sequential offline RL is necessary?

How much of the improvement in Table 2 remains when all methods use exactly the same frozen forecasting checkpoint, and why does the electricity CQL model outperform No News despite selecting no articles?

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

This paper aims to optimize news-driven time series forecasting by using offline reinforcement learning, enabling an agent to learn a news-filtering strategy based on actual prediction errors.

Critical strengths
  1. The paper distinguishes between the semantic relevance and forecasting utility of news. Rather than relying only on zero-shot inference, it uses prediction error to guide news selection. This is a valuable research direction.

  2. The experiments are evaluated across three domains and examine the impact of different news contexts on the prediction.

Critical weaknesses
  1. The independent contribution of Offline RL is not clear. According to Tables 2 and 3, CQL outperforms "No News", but for the "Fixed-Forecaster", the two perform nearly identically (59.66 v.s. 59.39). It is unsure whether the performance gains come from the different forecasting checkpoints or the RL-based news selection. Appendix C2 also reveals that 98% of the actions in the offline data are DROP. The optimal CQL policy for the Electricity dataset selects almost no news at all.

  2. The effectiveness of the reflection is unclear. Table 2 shows that reflection increases the RMSE for the Electricity dataset from 54.7 to 56.4. Appendix C4 indicates that for the electricity dataset in round 2, the RMSE increases from 54.7 to 88.8.

  3. The paper relies only on Gemma-2-9B as the forecasting backbone. It is unclear whether the learned news selection strategy can be transferred to other forecasting models.

Questions for authors
  • In the electricity and exchange rate datasets, over 98% of the actions selected in the offline data are DROP. How to prove that CQL has learned an effective strategy, rather than mimicking a distribution dominated by the "DROP"?

  • This paper regards news selection as a sequential decision-making process, but the current experiments only compare RL algorithms without validating the sequential dependencies. Can the authors remove the context of previously selected news items (Si−1S_{i-1}) from the state representation, or compare it with treating each news item independently?

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

Related papers

Discussion (0)

Sign in to comment.