When News Gets Old: Portfolio Allocation with Reinforcement Learning and Large Language Models

Paper: arXiv 2610.30564

Authors: Bolin Song, Surui Xiang

Abstract

Financial news can influence investment decisions long after publication. Timely news may reach an investor before prices fully adjust, whereas older news may already be reflected in prices. We examine two ways to handle aging signals: retain them and give a portfolio policy timing information, or filter them out before portfolio selection. A fixed large language model (LLM) extracts signals from public SEC disclosures. We compare these treatments in reinforcement learning (RL) policies and LLMs that choose portfolios directly. We evaluate returns after trading costs in historical tests over 2023-2025. RL choices respond to timing information in a controlled test, but historical returns show no consistent benefit from adding it. For the LLM, filtering old news produced the highest average net return, reaching 18.43% per year when news delivery was delayed by four weeks. This exceeded the returns of all RL strategies that used news.

Complexity vs Empirical Score

• Math Complexity: 5.0/10
• Empirical Rigor: 7.0/10
• Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: The paper presents a novel approach to integrating LLM-extracted news signals with RL for portfolio allocation, specifically addressing the ‘aging’ of news. It demonstrates solid empirical rigor through historical backtesting with various delay settings and cost considerations, though the dataset size is somewhat limited. The mathematical complexity is moderate, focusing on established RL frameworks rather than deep theoretical derivations.

Research Flowchart

  flowchart TD
    A[Research Goal: How to handle aging financial news for portfolio allocation?] --> B{Methodology: Compare 2 approaches for aging signals};
    B --> C1[Approach 1: Retain signals + timing info];
    B --> C2[Approach 2: Filter out old signals];
    C1 & C2 --> D{Computational Processes: Develop RL policies & LLM-driven policies};
    D --> E[Data/Inputs: Fixed LLM extracts signals from SEC disclosures];
    E --> F[Evaluation: Historical tests over 2023-2025, considering trading costs];
    F --> G[Key Findings: LLM with filtered old news achieved highest average net return (18.43% annually with 4-week delay)];
    G --> H[Outcome: Filtering old news outperformed retaining timing info for portfolio selection, especially for LLMs];