We gave 5 LLMs $100K to trade stocks for 8 months (www.aitradearena.com)

In a groundbreaking experiment, researchers assessed the trading capabilities of five large language models (LLMs)—GPT-5, Claude Sonnet 4.5, Gemini 2.5 Pro, Grok 4, and DeepSeek—by providing each with $100,000 in paper money to trade stocks over an eight-month period. The study, conducted in a controlled environment called Trade Arena, utilized a backtesting method where LLMs were fed historical market data, news, and company financials filtered to ensure they only accessed information available at the time of each simulated trade. This approach allowed the researchers to scrutinize the decision-making processes and performance of the models without the risk of “future leakage.” The significance of this experiment lies in its potential to expand our understanding of model behaviors and predictive capabilities in financial markets. By evaluating the trading strategies of these LLMs, which predominantly favored tech-heavy portfolios, the researchers can differentiate between skill and luck in trading outcomes. Notably, Grok emerged as the top performer, while Gemini lagged behind, primarily due to its investment in non-tech stocks. As the researchers plan additional experiments, including real-time trading scenarios, they aim to refine LLM trading strategies and explore the nuanced interplay of various factors influencing performance, ultimately leveraging the market as a benchmark for model evaluation and enhancement.