When News Gets Old: Portfolio Allocation with Reinforcement Learning and Large Language Models
This paper studies how to handle aging textual signals in weekly portfolio allocation: keep stale news with explicit timing information, or filter out reports older than a fixed threshold. A frozen LLM extracts ternary $z_{i,t} \in \{-1,0,1\}$ signals from SEC disclosures, then either a Double DQN RL policy or the same LLM selects portfolios for a ten-stock US universe. The core finding is that filtering old signals outperforms retaining them with age metadata for the LLM agent, yet nearly none of the return differences are statistically significant. The work matters because trading costs make signal timeliness practically relevant, but its claims of superiority rest on noisy, short-sample evidence.
The paper proposes an elegant controlled comparison of news-aging treatments—retaining signals with timing annotations versus filtering at an eight-week threshold—applied to both RL and direct-LLM portfolio policies. The experimental design is conceptually clean: a fixed LLM extracts $z_{i,t} \in \{-1,0,1\}$ signals from SEC disclosures, and all policies trade the same ten-stock universe under identical execution rules. However, the headline result that filtering yields superior LLM returns rests on noisy annual data with fragile statistical support, and the conclusion that RL "responds to" timing information without translating that response into reliable returns ultimately leaves the reader uncertain about whether either treatment is robustly preferable.
The strongest element is the paired comparison architecture: because every method shares the same extracted signals, market data, and trading cost model, differences in net returns can be attributed to the information treatment rather than data or execution variation. The controlled perturbation test—shifting all report ages backward while holding market features and holdings fixed—cleanly isolates whether a policy's choices actually depend on timing inputs. The LLM-filtering result is also practically interesting because it suggests a simple heuristic—discarding stale signals—can outperform more complex conditioning.
The principal weakness is that almost no result survives statistical scrutiny at conventional levels: the headline 18.43% versus 17.81% contrast is indistinguishable from sampling noise. This null pattern holds despite action-change rates of 52% in the controlled test, revealing a sharp disconnect between behavioral response and economic value. Compounding this, the directional impact of aging is inconsistent: "The mean does not decline steadily as reports age: some trained policies increase signal-aligned allocation, while others decrease it" (Appendix A-C), so even the qualitative behavior of RL is unstable across training seeds. The test period is extremely short (2023–2025), stock-specific, and partially pre-inspected; the authors note that "The evaluation is retrospective: 2023 had served as validation in a prior analysis, and 2024–2025 had already been inspected" (Appendix B-B), undermining claims of out-of-sample validity. The eight-week threshold $h=8$ is announced as pre-fixed but no sensitivity analysis supports it across alternatives like $h=4$ or $h=12$. Trading-cost fragility is severe: raising fees from 10 to 30 basis points collapses RL News + Age returns from 18.02% to 12.32% (Section IV-D), a loss larger than any claimed benefit from news.
The evidence supports the existence of an RL behavioural response to timing features but not that this response improves risk-adjusted returns. The 95% time-block intervals for News + Age versus News at four-week delay are $[-3.73, 4.29]\%$ and for Filtered News versus News are $[-2.93, 3.44]\%$ (Appendix A-A and Table III), both comfortably including zero and in some folds showing negative point estimates. The LLM policy exhibits even more volatile year-to-year variation, with timing effects swinging from $-3.42\%$ to $+6.40\%$ depending on the year and model (Appendix A-B, Table IV); the paper confirms that "All nine corresponding paired intervals contain zero" (Appendix A-B). Comparisons to related work are brief and mainly cite architectural hybrids rather than benchmarks against simpler age-decay baselines like exponential discounting of signals, which would have been a more informative rival to the hard eight-week filter. The synthetic-market validation also yields near-zero mean effects ($+0.35\%$ and $-0.32\%$) with confidence intervals containing zero (Appendix B-D), offering little external validity.
Reproducibility is mixed. Hyperparameters and network architectures are reported in detail—the Double DQN target is $y_t = r_{t+1} + \gamma(1-d_t)Q_{\bar{\theta}}(o_{t+1}, \arg\max_a Q_\theta(o_{t+1},a))$ with Adam at $3 \times 10^{-4}$ (Appendix B-B)—and the code is referenced as a "reproduction package," though no public repository URI appears in the text. Critical barriers remain: the GPT-6 Astra model is specified as "an API alias, not a dated snapshot" (Appendix C-A), and with temperature not even sent for Astra, the exact generative trajectory cannot be recovered. Only ten US stocks are used, selected by an oddly specific procedure ("choose the first alphabetical eligible ticker in each of eight declared business groups, then fill alphabetically"), which is unlikely to generalise. The 82.3% format-compliance rate for real-text extraction means 17.7% of signals are coerced to zero, a non-trivial source of unmodelled noise. Finally, because holdings and signals are shared across delay views but the evaluation is retrospective with prior inspection of 2023–2025, independent reproduction would need access to the exact pre-commitment protocol to rule out specification searching.
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.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.