Photo via Fast Company
Prediction markets have generated considerable buzz among investors and tech enthusiasts, with some social media users claiming significant returns by deploying artificial intelligence models to trade on platforms like Kalshi and Polymarket. However, recent research challenges these bold assertions, suggesting that AI-driven trading success may be far more elusive than popular narratives suggest.
Researchers at Arcada Labs conducted an ambitious real-world test, allocating $10,000 to each of six leading AI models and tracking their performance over 57 days on active prediction markets. According to the study published on arXiv, the results were sobering: every tested model lost money, with losses ranging from 16% to 30.8% on Kalshi. The findings underscore the complexity of translating AI capabilities into profitable trading decisions in unpredictable market environments.
The research revealed an interesting disparity between platforms, with models performing somewhat better on Polymarket than Kalshi—a difference the researchers attribute to market selection constraints. Grace Li, co-founder of Arcada Labs, notes that models given broader access to trade across multiple markets showed better outcomes than those limited to a standardized set. This suggests that autonomy in decision-making, rather than raw algorithmic intelligence, may be a critical factor in AI trading performance.
While the study tempers expectations around AI trading profitability, researchers believe the technology's trajectory remains significant. As models improve incrementally, AI-powered hedge funds may eventually become commonplace in financial markets. For Atlanta-based investment firms and financial professionals, the study serves as a reminder that emerging AI capabilities warrant careful evaluation rather than speculative enthusiasm, particularly as regulators intensify scrutiny of prediction markets and autonomous trading systems.
