I let 7 AIs bet real World Cup markets for 27 days. Here's what 1,276 bets actually revealed.
Author: https://www.reddit.com/user/Obside_AI/
Abstract:
This post summarizes an experiment where seven AI models trade real Polymarket FIFA World Cup markets using simulated capital. Each model began with $10,000 in play money and places bets one hour before kickoff using real market prices and odds. After 96 settled matches and 1,276 bets, the models are collectively profitable, earning a combined simulated profit of $26,535 against human-priced markets. The results suggest that the models perform best not by simply picking winners, but by identifying mispriced probabilities, especially in mid-range odds, draws, and less obvious outcomes. The experiment also shows that high model confidence does not always translate to profitable bets, since confident predictions are often already priced into the market. Different models display distinct betting behaviors: some prioritize consistency, others take more diverse or aggressive positions, and model cost does not appear to correlate with performance. Overall, the project uses prediction markets as a live test of how LLMs reason under uncertainty, revealing both their ability to exploit human biases and their tendency to repeat common betting mistakes such as overconfidence and chasing favorites.