Title: Enhancing Time Series Forecasting Through Importance-Aware Fusion and PRM-Guided Reflection

Abstract:

Integrating news data into time series forecasting offers a distinct advantage, as it captures sudden exogenous shocks that historical data patterns alone cannot anticipate. However, current pipelines relying on Large Language Models (LLMs) for this task encounter two significant practical hurdles: relevant news articles frequently surpass the model’s context window capacity, and the iterative retrieval of additional news is often unguided. This lack of direction results in redundant updates and sluggish convergence.

To overcome these challenges, we propose a novel framework that integrates importance-aware news compression with process-level retrieval supervision. Our approach begins by training an importance reward model designed to estimate the forecasting utility of individual articles. This estimation guides the allocation of compression budgets during sequential pairwise fusion, ensuring that critical information is retained within strict context limits.

Furthermore, we introduce a Process Reward Model (PRM) to manage the selection of supplementary news. By ranking multiple candidates based on the current error profile and the history of previously chosen articles, the PRM replaces inefficient, one-shot blind retrieval with a quality-controlled selection process. Both components are trained offline using historical data with ground truth labels. During inference, the system employs frozen filtering logic and compression modules, operating without any reflection loop.

Evaluations across benchmarks in finance, energy, traffic, and Bitcoin forecasting demonstrate that our method achieves superior prediction accuracy compared to strong baselines. Additionally, it significantly cuts down the number of refinement iterations required by iterative baselines and maintains robust performance even when relevant articles span thousands of tokens.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC