🤖 AI

From Long News to Accurate Forecast: Importance-Aware Fusion and PRM-Guided Reflection for Time Series Forecasting

This paper proposes a novel time series forecasting framework that enhances accuracy and efficiency by combining an importance-aware news compression mechanism to handle long contexts with a process reward model-guided retrieval strategy to eliminate redundant updates in LLM-based pipelines.

Original authors: Mingyang Liu, Qingcan Kang, Yuke Wang, Shixiong Kai, Kaichao Liang, Hui-Ling Zhen, Tao Zhong, Mingxuan Yuan, Linqi Song

Published 2026-06-03

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mingyang Liu, Qingcan Kang, Yuke Wang, Shixiong Kai, Kaichao Liang, Hui-Ling Zhen, Tao Zhong, Mingxuan Yuan, Linqi Song

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict the future price of Bitcoin, the flow of traffic, or the demand for electricity. You have a very smart computer (an AI) that is great at looking at past numbers to guess what happens next. But sometimes, the future changes suddenly because of something outside the numbers—like a new law, a natural disaster, or a big company announcement. These events are like "ghosts" that don't show up in the historical data but completely change the outcome.

To fix this, people started feeding the AI news articles. But this created two big problems:

The "Too Much to Read" Problem: News articles can be huge (like a 50-page transcript of a CEO's speech). The AI has a "short-term memory" limit (context window) and can't read everything. If you just chop off the end, you might lose the most important part. If you summarize it poorly, you might lose the details that actually matter for the prediction.
The "Guessing Game" Problem: Some systems try to fix bad predictions by asking the AI, "What did I miss?" and then blindly grabbing more news. This is like a detective asking for more clues without checking if they are good clues. They might grab ten useless articles before finding the one that actually helps, wasting time and money.

This paper proposes a new system to solve both problems. Think of it as hiring a super-efficient editor and a smart detective.

1. The Super-Efficient Editor (Importance-Aware Fusion)

Instead of just cutting long news articles in half or summarizing them randomly, the system trains a special "Editor AI."

How it works: Before the news is fed to the prediction model, the Editor reads each article and asks, "How useful is this specific sentence for predicting the future?"
The Analogy: Imagine you are packing a suitcase for a trip, but you only have a small bag. A normal person might just throw in random clothes. This Editor, however, knows exactly which items are essential for the weather you are going to. It keeps the "warm coats" (important details) and throws away the "extra socks" (fluff), even if the coat is huge.
The Result: It compresses thousands of words of news into a tiny, perfect summary that fits in the AI's memory, keeping only the parts that actually help with the forecast.

2. The Smart Detective (PRM-Guided Reflection)

When the AI makes a mistake, it needs to find the missing piece of the puzzle. Old systems would just grab the first few articles they found that seemed related. This new system uses a Process Reward Model (PRM), which acts like a seasoned detective.

How it works: When the AI says, "I missed something," the PRM doesn't just grab a random clue. It looks at a list of potential new news articles and scores them: "If I add this specific article, will it actually lower my error rate?"
The Analogy: Imagine you are trying to solve a mystery. A naive detective grabs every file in the office. This Smart Detective looks at the files and says, "File A is just noise. File B is a red herring. But File C? That's the smoking gun." It picks the best file to add, not just any file.
The Result: The system finds the right news much faster. The paper claims it reduces the number of "guessing rounds" needed to get a good answer by about 25% on average.

The "Offline Training" Trick

Here is the clever part: The "Editor" and the "Detective" do all their hard work offline (before the system goes live). They practice on historical data where they already know the answers. They learn how to compress news and how to pick the best clues.

Once they are trained, they are "frozen." When the system is used in the real world (online), it doesn't stop to think or re-evaluate. It just uses the rules it learned during practice to instantly compress news and pick the right articles. This makes it fast and reliable.

Does it work?

The authors tested this on four real-world scenarios:

Electricity demand
Bitcoin prices
Traffic flow
Currency exchange rates

The Results:

Better Accuracy: Their method predicted these values more accurately than other strong AI models, especially in tricky areas like Bitcoin (where news moves markets fast).
Faster Learning: It needed fewer attempts to find the right news articles compared to systems that just grab news blindly.
Handles Long Text: It worked even when the news articles were thousands of words long, proving that the "Editor" can handle massive amounts of text without losing the important bits.

In short, this paper teaches an AI how to read the news efficiently (keeping only what matters) and search for clues intelligently (picking the best ones), making it much better at predicting the future when unexpected events happen.

Technical Summary: From Long News to Accurate Forecast

Problem Statement

Incorporating external textual evidence, particularly news, into time series forecasting is a promising approach to capture abrupt exogenous events (e.g., policy changes, market announcements) that historical numerical data alone cannot reflect. However, existing Large Language Model (LLM)-based pipelines face two critical limitations in realistic settings:

Context Window Constraints: Relevant news articles (e.g., earnings transcripts, policy briefings) often exceed the context limits of forecasting models. Naive truncation discards causal details, while generic summarization methods are not optimized for preserving information specifically useful for forecasting.
Unsupervised Iterative Refinement: Current iterative pipelines often rely on "blind" supplementary retrieval. When a reflection signal indicates missing information, the subsequent retrieval is typically accepted without quality control. This leads to redundant updates, sensitivity to noise, and slow convergence.

Methodology

The authors propose a novel framework that separates offline refinement (where ground truth is available) from online deployment (inference). The system consists of four main components: an LLM forecasting backbone, a reasoning agent, an importance-aware fusion module, and a Process Reward Model (PRM).

1. Importance-Aware Fusion (Offline Training & Online Inference)

To address long documents, the framework introduces a module that compresses news based on estimated forecasting utility rather than generic importance.

Reward Model ( $R_\theta$ ): A reward model is trained to estimate an importance score $\rho \in [0, 1]$ for each article. The supervision signal is derived by comparing the forecasting RMSE when an article is aggressively compressed versus lightly compressed.
Sequential Pairwise Fusion: Articles are processed sequentially. The retention ratio for each article is determined by a monotone function of its importance score. This ensures that high-utility articles retain more information within a fixed context budget, while low-utility content is compressed more aggressively.
Deployment: During inference, the trained reward model and fusion logic are frozen and applied deterministically without iterative search.

2. PRM-Guided Supplementary News Selection (Offline Refinement Only)

To address uncontrolled retrieval, the framework replaces blind updates with a quality-controlled ranking mechanism.

Process Reward Model (PRM): A PRM is trained to rank multiple candidate news articles based on their expected contribution to error reduction. The model takes as input the current task context, the history of previously selected articles, and the candidate article itself.
Training Trajectory: The PRM is trained on offline trajectories where the "gain" ( $\Delta_t$ ) of adding a candidate is measured as the reduction in validation RMSE. The model learns to predict the probability that a candidate is beneficial ( $P(\text{beneficial} | \dots)$ ).
Refinement Loop: During offline training, a reasoning agent generates a reflection on missing information, retrieves a pool of candidates, and the PRM selects the highest-scoring article to update the news set and refine the agent's persistent filtering logic.
Deployment: Crucially, no reflection loop or PRM scoring occurs during online inference. The system uses the refined filtering logic and the frozen compression modules to retrieve and process news in a single forward pass.

Key Contributions

The paper summarizes its contributions as follows:

Formulation: It formulates news-augmented forecasting as an offline-refined pipeline that explicitly tackles long-document compression and uncontrolled supplementary retrieval.
Importance-Aware Fusion: It proposes a module that learns forecasting-oriented article utilities to adapt compression rates during sequential pairwise fusion, enabling the incorporation of long news under limited context budgets.
PRM-Guided Refinement: It introduces a strategy for supplementary news selection that replaces one-shot retrieval with candidate ranking based on process-level supervision from error-driven trajectories.

Experimental Results

The method was evaluated on four domains: Electricity, Bitcoin, Traffic, and Foreign Exchange (Exchange).

Prediction Accuracy: The proposed method achieved the best RMSE across all four domains.
- Bitcoin: Achieved an RMSE of 886.58, a 23.3% improvement over the second-best baseline (LightGBM at 1156.65).
- Exchange: Achieved an RMSE of 5.86, outperforming the strongest baseline (LightGBM at 7.09) by 17.3%.
- Electricity & Traffic: Showed marginal improvements, which the authors attribute to the strong periodicity and regularity of these domains, making them less sensitive to external news events compared to financial markets.
Convergence Efficiency: The PRM-guided selection significantly reduced the number of refinement iterations required to reach target accuracy compared to naive selection.
- Average Reduction: 24.8% fewer steps across all domains.
- Bitcoin: The most significant improvement, with a 37.6% reduction in convergence steps.
Ablation Studies:
- Removing external news ("w/o News") caused significant performance degradation, confirming the value of exogenous signals.
- Removing the reward model ("w/o Reward Model") also degraded performance, demonstrating that unguided compression introduces noise and discards key predictive signals.

Significance and Claims

The paper claims that its framework effectively bridges the gap between the need for rich external context and the constraints of LLM context windows. By decoupling the expensive search and supervision (offline) from the deployment (online), the system achieves high accuracy without the computational overhead of iterative reflection during inference.

The authors emphasize that their approach is particularly effective for domains where exogenous events drive dynamics (e.g., finance), while acknowledging that for highly periodic domains (e.g., traffic), the marginal gains are smaller. The work positions itself as a solution to the "bottleneck" of news-augmented forecasting, specifically addressing the trade-off between context efficiency and task relevance, and the risk of noisy iterative updates. Future work is proposed to extend the framework to multimodal data and larger news pools.