- 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
- 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.
- 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
-
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.
-
Does RL-HIP still outperform simple global shrinkage under the main experimental setup?
-
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?
-
Could the authors clarify how the RL policy is trained for Bitcoin?
-
Could the authors clarify the MSE and RMSE calculations for the Bitcoin LLM forecaster in Table 3? The reported values seem unusually far apart.