Additional reasoning in large language models does not improve trading returns
The Price of Thought: Does Test-Time Reasoning Pay in LLM Trading?
Artificial Intelligence
Summary
The paper explores if making large language models think harder when predicting stock prices actually leads to better investment profits. The authors tested various models and reasoning efforts on one year’s worth of US stock data with consistent setups. They found that spending more computing power on reasoning did not reliably increase profits after accounting for costs. Sometimes extra reasoning even led to unstable or unpredictable trading choices. This suggests that more model thinking doesn’t always pay off financially and should be carefully checked before use in trading.
What this means in practice
- •For quantitative traders: Decide whether to include additional LLM reasoning steps when building stock prediction models based on economic return impact rather than model scores alone.
- •For financial technology teams: Assess the trade-off between computational cost and economic benefit of complex reasoning in language models for financial decision systems.
Authors
Jiayi Chen, Guiling Wang
Abstract
While inference-time reasoning in large language models (LLMs) promises better decision making, its higher computational cost may not yield better economic outcomes. Yet reasoning controls are rarely evaluated as economic interventions, where changes in model outputs must translate into better portfolios after trading costs. We conduct a controlled study of representative LLMs from the DeepSeek, GPT, and Gemini families. We vary reasoning effort while holding information available at each formation date, prompts, output formats, and portfolio construction fixed. Our evaluation covers a full year of U.S. equities under three input conditions: numerical, identifiable news, and masked news. It includes more than 800,000 asset predictions and repeated model generations. Across all three model families, additional reasoning does not produce a reliable improvement in net portfolio returns. For DeepSeek, where we examine the full progression from no reasoning to maximum reasoning, performance is nonmonotonic. Repeated generations also produce unstable treatment effects and portfolio selections, even when overall scores remain similar. These findings show that additional reasoning can change financial decisions without reliably improving their economic value, motivating validation for each task before deployment.