Autonomous LLM Agents Trading in Decentralized Financial Markets - Wan Chu Cheong

Event Date: 21 October 2026

Speaker: Wan Chu Cheong, London School of Economics

Venue: Strathclyde Business School, Cathedral Wing, CW404a

Time: 2pm

Abstract:

This study examines autonomous large language model (LLM) agents as a new class of market participant that independently acquires information and executes trades without contemporaneous human intervention. Using data from Polymarket, a real-money prediction market used by both human traders and autonomous agents, I find that agent wallets trade more accurately than human wallets within the same market, and that their accuracy improves following a major LLM model upgrade. The agent accuracy advantage is larger when annual filings contain more tokens and longer sentences and when year-over-year disclosure changes contain more litigious or forward-looking language, suggesting that agents have a comparative advantage in processing large volumes of qualitative textual information. In contrast, this relative advantage diminishes as disclosures become more numerically intensive, uncertain, or vague, and in markets involving M&A rumors, where relevant information is less verifiable and requires greater judgment. Greater agent trading volume is also associated with lower subsequent market price error, with this association more pronounced in markets that receive no initial human participation. Polymarket price changes are associated with movements in the underlying equity price, and this association is stronger in markets with high agent trading volume.

Published: 14 October 2026