30-Day AI Trading Bot Challenge: Two Bots Beat the S&P 500 with Real Money

In a 30-day experiment from late February to late March 2026, two former finance professionals (ex-JP Morgan and ex-Goldman Sachs) each deployed an AI trading bot with $10,000 of real money via the Alpaca brokerage API. ## Setup Both bots ran autonomously with no human intervention after day 0. Each was given a system prompt defining its strategy and access to market data and trading execution. As a bonus, the bots were configured to email each other daily — resulting in trash talk and attempted prompt injection between them. ## Results | Bot | Strategy | Starting | Final | Return | vs S&P 500 | |-----|----------|----------|-------|--------|------------| | "Bull" (Nate's) | "Be a financial adviser, do research, trade as you see fit" | $10,000 | $9,981 | -0.19% | Beat by 8.27% | | Salmon's | Pareto-optimized aggressive strategy | $10,000 | $9,624 | -3.76% | Beat by 4.70% | | S&P 500 | — | — | — | -8.46% | Benchmark | The winning strategy was notably unsophisticated — it instructed the AI to spin up sub-agents for research and trade however it judged best. The more complex, explicitly optimized strategy performed worse. ## Key Observations Both bots outperformed the market during a significant downturn, suggesting that AI trading bots may be better at *not losing money* during drops than at generating alpha during bull runs. The simpler strategy outperforming the complex one mirrors the broader finding in quantitative finance that simpler models often beat overfit ones. However, 30 days during a specific market regime (a downturn) is not a meaningful sample size. Both bots may have simply been cautious by default (holding cash or defensive positions), which mechanically outperforms in a falling market. The test would need to run through multiple market regimes — bull, bear, and sideways — to demonstrate genuine edge. The bot-to-bot prompt injection attempts are an interesting side note on AI security in multi-agent systems.