Two University Scholars Studied AI Stock Trading for 3 Months, and They Told Me AI Is Very Much Like Buffett
Summary
Two university research teams (from UIUC and Hong Kong University) ran three-month live-trading experiments and benchmarks (LiveTradeBench and AI-Trader on GitHub) to test large AI models on US equity markets. They found large models can capture short-term alpha and help trading when fed up-to-date, structured market and news data, but they struggle to sustain long-term outperformance across full market cycles. The teams warn static benchmarks can be gamed, model inference cost and latency limit high-frequency use, and models currently serve best as decision-support tools for medium-to-low-frequency strategies. Overall, AI trading shows promise but has clear limits and needs longer, more complete market-cycle validation before it can reliably generate sustained profits autonomously.