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

Bolin Song; Surui Xiang

cs.LGOct 22, 2026

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

Authors: Bolin Song, Surui Xiang

Organizations: Cornell University · United Nations

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.

Figures & tables

Fig. 1: News Signals and Portfolio Decisions. An LLM extracts news signals, while report dates determine news age. The RL policy and direct LLM baseline choose portfolios independently, with separate holdings under common backtesting rules. Table I specifies each method’s inputs.

Policy	Input setting	Market data & holdings	News signals	News age & time since arrival	Decision maker
Equal Allocation		—	—	—	Fixed weights
Buy and Hold		—	—	—	Fixed positions
RL	Market Data	$\checkmark$	—	—	RL
	Market Data + News	$\checkmark$	$\checkmark$	—	RL
	Market Data + News + News Age	$\checkmark$	$\checkmark$	$\checkmark$	RL
	Market Data + Filtered News	$\checkmark$	$\checkmark^{*}$	—	RL

TABLE I: Decision inputs and comparison methods.

		Annualized return after costs			Results with a four-week news delay
Policy	Input setting	0 weeks	1 week	4 weeks	Sharpe ratio	Maximum drawdown	Annual trading volume ( $\times$ )	Total trading costs
Equal Allocation		17.63%	17.63%	17.63%	1.13	10.44%	1.09	0.12%
Buy and Hold		17.19%	17.19%	17.19%	1.13	10.45%	0.00	0.00%
RL	Market Data	19.07%	19.07%	19.07%	1.00	12.99%	25.12	2.69%
	Market Data + News	16.22%	16.95%	17.81%	0.96	12.49%	23.72	2.55%
	Market Data + News + News Age	15.41%	15.11%	18.02%	0.98	12.88%	24.75	2.64%

TABLE II: Historical returns and trading outcomes.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Comparison	Difference (%)	95% interval (%)
News + Age $-$ News	$+0.21$	$[-3.73,4.29]$
Filtered News $-$ News	$+0.22$	$[-2.93,3.44]$

Appendix

TABLE III: Paired return differences at four-week delay.

	Return change (%)
Model	2023	2024	2025
GPT-4.1 mini	$-3.42$	$+6.40$	$+0.48$
Llama 3.3	$+0.72$	$+0.54$	$-1.76$
GPT-6 Astra	$+5.35$	$+0.50$	$-4.36$

Appendix

TABLE IV: Timing effects in LLM portfolio policies.

	RL		Direct LLM
Year	News + Age $-$ News	Filtered News $-$ News	News + Age $-$ News	Filtered News $-$ News
2023	$+1.18\;[-4.39,6.71]$	$-1.41\;[-5.15,2.08]$	$-3.42\;[-17.35,10.18]$	$+4.34\;[-4.90,18.77]$
2024	$+0.51\;[-3.86,4.82]$	$+1.44\;[-4.84,8.35]$	$+6.40\;[-7.90,20.99]$	$-3.93\;[-24.10,13.41]$
2025	$-1.06\;[-10.63,8.98]$	$+0.63\;[-5.38,6.12]$	$+0.48\;[-8.78,8.43]$	$+6.83\;[-6.60,20.02]$

Appendix

TABLE V: Annual return differences at four-week delay, with paired 95% intervals (%).

Finance

Finance studies markets, institutions, and financial decisions using data and models to understand risk, value, and resource allocation.

See all 281 papers

Large Language Model-Driven Small-Capitalization Trading: Integrating Financial News Sentiment, Macroeconomic Indicators, and Technical Signals
Aug 12, 2026Alireza Kargarzadeh, Nariman Khaledian, Navid Parvini +1Portfolio Optimization Finance
Tailstate Intelligence Ltd. · Independent Researcher · Zanista AI Ltd.
Reading the News Without Hindsight: NoLBERT Topics and the Cross-Section of Stock Returns
Oct 15, 2026Kevin Foley, Jonathan Hartadi, Shivesh Prakash +2Financial Forecasting Stock Price Forecasting
University of California, Berkeley, Berkeley, California, USA
The Price of Thought: Does Test-Time Reasoning Pay in LLM Trading?
Sep 25, 2026Jiayi Chen, Guiling WangLLM Reasoning Efficient Language Model Reasoning
Department of Computer Science New Jersey Institute of Technology Newark, New Jersey, USA

Aug 12, 2026J/K move · Enter open · S save

Large Language Model-Driven Small-Capitalization Trading: Integrating Financial News Sentiment, Macroeconomic Indicators, and Technical Signals

Alireza Kargarzadeh, Nariman Khaledian, Navid Parvini, Arman Khaledian

Tailstate Intelligence Ltd. · Independent Researcher · Zanista AI Ltd.

Portfolio Optimization Finance Algorithmic Trading

Open paper AskPDF arXiv

Large language models can extract richer signals from financial news than fixed sentiment lexicons, and recent work has explored feeding such signals into portfolio construction. We study an uncertainty-aware construction that feeds model-predicted risk -- decomposed into aleatoric and epistemic components -- directly into the covariance matrix of portfolio allocators, rather than treating portfolio risk as fixed or adjusting only expected returns. We evaluate the pipeline on Russell 2000 equities under three stock-selection regimes: a pure-alpha trigger that isolates abnormal stock moves not explained by macro indicators, a pure-beta trigger that captures macro-indicator moves before the stock itself fires, and a beta trigger in which both channels agree. Across the full holding-period grid, the separated pure-alpha and pure-beta legs usually dominate the beta intersection on Sharpe and return. Two horizons are especially informative. At one day, pure beta can work under low and moderate transaction costs because it captures immediate lead-lag spillovers from liquid macro and sector indicators into exposed small-cap stocks, but this advantage disappears at 100 bps when turnover and microstructure noise dominate. At 40 days, pure beta works for a different reason: slower macro repricing overtakes the firm-specific pure-alpha channel. The strongest conservative row is pure beta with GPT-4o mini sentiment, a Student-t target, a 40-day holding period, and risk parity allocation, reaching Sharpe 2.33 at 100 bps. The results suggest that stock-selection regime and allocator choice matter at least as much as the sentiment model, and that separating firm-specific and macro-exposure triggers is more informative than requiring both to fire simultaneously.

Abstract

Finance

Explore similar work

Large Language Model-Driven Small-Capitalization Trading: Integrating Financial News Sentiment, Macroeconomic Indicators, and Technical Signals

Reading the News Without Hindsight: NoLBERT Topics and the Cross-Section of Stock Returns

The Price of Thought: Does Test-Time Reasoning Pay in LLM Trading?