From Long News to Accurate Forecast: Importance-Aware Fusion and PRM-Guided Reflection for Time Series Forecasting
Researchers propose a novel framework combining importance-aware news compression and process reward models to improve LLM-based time series forecasting across finance, energy, and cryptocurrency markets. The method addresses practical limitations of existing approaches by intelligently filtering news articles within context windows and guiding iterative retrieval, achieving better accuracy with fewer refinement iterations.
This research tackles a fundamental challenge in applying large language models to financial forecasting: effectively leveraging news data without exceeding computational constraints or conducting wasteful iterative searches. The framework introduces two key innovations that work in tandem. An importance reward model learns to score news articles by their forecasting utility, enabling selective compression that preserves signal while respecting token limits. A process reward model then supervises supplementary news retrieval by ranking candidates based on current prediction errors and previously selected articles, replacing blind iterative collection with intelligent quality control.
The approach addresses real deployment constraints that practitioners face when incorporating news into production forecasting systems. Financial markets generate vast quantities of textual data daily, and existing pipelines struggle to process relevant articles efficiently without sacrificing information quality. By training both components offline using historical ground truth, the framework achieves practical efficiency—inference requires no reflection loops or expensive real-time computations.
The experimental validation across multiple domains strengthens the research's credibility. Performance improvements on finance, energy, traffic, and bitcoin forecasting benchmarks suggest the method's generalizability beyond any single asset class. Notably, the framework maintains effectiveness even when relevant articles span thousands of tokens, indicating robustness to information-dense scenarios common in cryptocurrency markets during high-volatility periods.
For market participants and practitioners, this work demonstrates that intelligent news filtering outperforms naive corpus expansion. The reduction in refinement iterations directly translates to lower computational costs and faster prediction cycles, competitive advantages in high-frequency decision-making environments.
- →Importance reward models identify high-utility news articles for forecasting, enabling effective compression within fixed context windows.
- →Process reward models eliminate wasteful iterative news retrieval by intelligently ranking supplementary candidates based on error profiles.
- →The framework improves prediction accuracy across finance, energy, traffic, and cryptocurrency benchmarks with fewer refinement iterations.
- →Offline training of filtering logic eliminates expensive reflection loops during inference, improving computational efficiency.
- →The method remains effective when relevant news articles span thousands of tokens, addressing real-world information density challenges.