Study finds LLM trading results shift with market information disclosure

Forkast reported on a preprint dated November 5, 2026, that examines how large language models perform when the amount of market information they receive changes. Ten LLMs were put through their paces in three controlled market settings, one involving trade across several resources, one built on bilateral bargaining, and one based on double auctions. According to the preprint, disclosure levels were varied while the models and incentives were held constant.

Outcomes differed by model and setting. The preprint says GPT-6 Sol's realized gains fell from 97.7% to 90.8% under full disclosure, while Claude Sonnet 5's rose from 72.1% to 85.6% under the same condition. In a double auction, the authors wrote that Claude Opus 5.5's realized gains dropped from 94.6% to 64.1%. A ten-resource study found that complete information raised candidate quality across every model, and a communication-length test showed GPT-5.1 welfare rising 30.6 points under the least information.

The authors caution that the findings are bounded to the selected controlled markets, which limits how far the results reach beyond those settings. Because the tests covered only those environments, the results do not establish how the same models would behave in other markets.