acceptodds
Withdrawn from ICLR 2027

HIP: Forecast-Calibrated Event Representations for Multi-Horizon Time Series Forecasting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Forecast-relevant events do not by themselves specify how their influence should be represented across a future trajectory. We introduce Hierarchical Impact Pathways (HIP), a structured intermediate representation that maps semantic event hypotheses and temporal attributes into horizon-wise conditioning signals for multi-horizon forecasting. HIP explicitly separates what an event is hypothesized to affect from how its influence is represented across future prediction steps. Because semantically plausible event descriptions are not necessarily predictively useful, we further introduce RL-HIP, which calibrates event retention and pathway attributes using downstream forecasting feedback while keeping the forecasting model and pathway policy separately trained. We evaluate the framework on electricity-demand and Bitcoin forecasting with multiple numerical backbones. RL-HIP provides strong and consistent improvements for electricity demand, while the gains on Bitcoin are smaller and more architecture-dependent. These results suggest that the utility of external events in multi-step forecasting depends not only on their relevance, but also on how their predicted influence is represented and calibrated across future horizons.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Multi-horizon time series forecasting, event-conditioned forecasting, multimodal representation learning, structured event representation, forecast calibration, reinforcement learning.

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

This paper proposes Hierarchical Impact Pathways (HIP), a structured representation for incorporating external events into multi-horizon time series forecasting. HIP decomposes event impact hypotheses into semantic and temporal attributes, including affected components, direction, intensity, and uncertainty, which are transformed into horizon-specific conditioning signals. The authors further introduce RL-HIP, which uses LLM-generated pseudo-labels for initialization and PPO with downstream forecasting feedback to calibrate the structured event representations.

Experiments are conducted using four numerical backbones and representative LLM forecasters, with additional controlled ablations and forecast-horizon analyses. RL-HIP achieves consistent MAE improvements of approximately 5.4–10.9% on the AULF numerical benchmarks, while improvements on Bitcoin are smaller and less consistent.

Critical strengths
  1. Well-motivated and interpretable representation design.

The paper identifies a meaningful distinction between determining event relevance and modeling how event influence varies across forecasting horizons. The explicit decomposition into semantic and temporal attributes provides a structured interface that can be integrated with different forecasting backbones. This is a useful perspective beyond directly injecting static event embeddings or textual descriptions.

  1. Encouraging empirical improvements.

RL-HIP consistently improves all four numerical backbones on the main AULF benchmark, with substantial reductions in forecasting error. The authors also provide controlled analyses of policy calibration, temporal conditioning, and event retention. Importantly, they acknowledge negative and inconclusive results rather than claiming universal improvements.

Critical weaknesses

1. The benefit of temporal conditioning is not clearly established.

The Static vs. Temporal comparison shows clear improvements for only two of the four backbones. Moreover, the two settings have substantially different conditioning magnitudes, making it difficult to attribute the gains specifically to horizon-dependent modeling.

2. The necessity of RL-based calibration remains unclear.

In Table 15, simple global shrinkage achieves performance indistinguishable from Full RL-HIP. Is the additional complexity of PPO and attribute-level calibration justified then?

3. Limited evidence of generalizability and the value of event information.

The strongest improvements are concentrated on the electricity dataset, while the results on Bitcoin and AUD/USD are less convincing. The Context-only control also changes the training distribution, making it difficult to isolate the contribution of actual news content. Direct comparisons with recent event-aware forecasting methods are also missing.

Questions for authors
  1. Can the authors compare Static and Temporal conditioning with matched signal magnitudes? This would help clarify whether the temporal structure itself contributes to the improvements.

  2. Does RL-HIP still outperform simple global shrinkage under the main experimental setup?

  3. Can the authors provide a controlled experiment to better isolate the contribution of actual news content, such as shuffling or masking the news while keeping the training setup unchanged?

  4. Could the authors clarify how the RL policy is trained for Bitcoin?

  5. Could the authors clarify the MSE and RMSE calculations for the Bitcoin LLM forecaster in Table 3? The reported values seem unusually far apart.

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

This paper proposes HIP, a structured event representation method for multi-step time series forecasting. It maps external events into conditional signals for various forecast horizons using semantic attributes and temporal decay functions. The authors further introduce RL-HIP, which employs downstream forecasting feedback and PPO (Proximal Policy Optimization) to refine event selection and path attributes. Experiments cover electricity load and financial time series forecasting, with consistent performance improvements observed in the electricity load tasks.

Critical strengths

The paper presents a comprehensive framework for structured representation and calibration in event-augmented time series forecasting—a research problem with clear practical relevance. Table 2 demonstrates that RL-HIP yields performance gains across four numerical forecasting models on the AULF dataset. Furthermore, the authors provide several ablation studies and discuss some of the method's limitations.

Critical weaknesses

1. Insufficient evidence for the necessity of the complex RL-based calibration mechanism.

Table 15 shows that a simplified version, "Shrinkage-HIP," achieves performance comparable to the full RL-HIP model (MAE 177.100 vs. 175.719), with the difference being statistically insignificant. This suggests that the performance gains may stem primarily from the suppression of event-based conditional signals, rather than from the complex optimization of semantic and temporal attributes. The current ablation studies are inadequate to disentangle the contributions of these different factors. Moreover, the structured attributes refined via RL lack independent verification of their semantic consistency, leaving their interpretability in question.

2. Insufficient experimental comparisons and evidence for generalization. **

The paper primarily compares its approach against "No events," "Raw-event," and "LLM-HIP." It lacks direct comparisons—under a unified experimental setup—with other event-enhanced forecasting methods, making it difficult to fully demonstrate the advantages over existing work. Furthermore, the performance gains for Bitcoin are marginal; results for the AUD/USD pair do not show consistent improvements; and the "LLM Forecaster" for AULF fails to benefit from RL-HIP. Additionally, the reported results across multiple random seeds primarily reflect variations in the training of the final forecasting model, failing to account for the stochasticity inherent in the training of the PPO policy itself. These issues limit the robustness and generalizability of the experimental conclusions.

3. The effectiveness of the temporal structure modeling has not been sufficiently demonstrated.

The primary contribution of HIP lies in constructing "horizon-wise event representations" via explicit temporal attributes. However, the "Static vs. Temporal" comparison in Table 13 does not control for the magnitude of the conditioning tensors. For instance, in the PatchTST model, the average tensor norms for the two settings are 7.068 and 20.686, respectively; consequently, the observed performance differences cannot be attributed solely to the temporal structure. Moreover, the "Temporal" approach shows a clear advantage only with certain backbones. Therefore, the independent contribution of the proposed temporal representation mechanism remains unclear.

Questions for authors
  1. Given that the performance difference between "Shrinkage-HIP" and "Full RL-HIP" is not significant, could you verify this finding across a broader range of backbones and clarify the necessity of the complex RL-based calibration mechanism?
  2. Could you provide comparisons with a wider array of event-enhanced forecasting methods to better highlight the advantages of your approach?
  3. Could you re-evaluate the "Static" versus "Temporal" comparison while controlling for the norms of the conditioning tensors—perhaps by incorporating simple learnable temporal weights or a gating mechanism—to substantiate the independent value of the HIP temporal structure?
Reviewer_AZonOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper attempts to identify how the influence of a relevant event in the past is distributed along the future. This leads to a model called Hierarchical Impact Pathways. This is essentially a structured way of classifying an event into what type of effect it may have and when. RL-HIP is a related method that then tunes this classification based on downstream results.

Critical strengths

The paper has a clear motivation; it is somewhat well understandable. The presented empirical results seem to be promising.

Critical weaknesses

The most important weakness, which in my eyes disqualifies the paper, is that I think there is no comparison to other event-aware or horizon-conditioned methods. Given that the introduction of a new method appears to be the only contribution here, it should be very comprehensively tested against the state-of-the-art. What the authors present instead are in a sense, comparisons of the same TS forecasting methods based on HIP and not on HIP. But is there a single HIP-based forecaster that can achieve sota performance under fair conditions?

Second, I think the HIP principle isn't actually strongly validated by the numerical examples. The LLM-HIP approach is almost consistently the worst method (Table 2). It requires calibration to work (the authors acknowledge this). But this could also just mean that the calibration step is doing the heavy lifting. It does not isolate HIP as a meaningful format. This is especially so since, for the LLM forecaster, no HIP-based method performs well. This could be because the LLM forecaster actually learns a better representation, which would be an argument against HIP as a good format.

The third issue is that I am not sure if I get any particular insight from the paper other than that the demonstrated method works sometimes. Is an approach like this sometimes necessary? Are there toy examples where one can show that HIP is the right format and that, with other approaches, prediction must fail? Is it clear for which problems HIP is not a good method?

Questions for authors

Q1: Can the authors compare performance with a selection of event-aware approaches or just other state-of-the-art time series prediction methods? Q2: Is there a theoretical justification for HIP? What are the conditions under which it is expected to work? Is it possible to demonstrate a theoretical result here? Q3: Is there a way to guarantee that the calibration step does not add something to the model that boosts its performance but is independent of the HIP principle?

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

Related papers

Discussion (0)

Sign in to comment.