Can AI be Used for Stock Selection? S&P 500 Backtest Shows 0.64% Weekly Spread between High and Low Groups

According to Insightful Beating monitoring, three economists conducted a backtest linking actual AI usage to the U.S. stock market. The research covered the period from January 2024 to April 2026, utilizing around 380 trillion Tokens on OpenRouter. OpenRouter is an AI model aggregation platform where users can access over 400 models through a single interface.

The researchers synthesized the weekly Token usage, expenditure amount, and active user growth rate into an "AI Usage Index," then observed the stock price reactions of each U.S. stock over the past 13 weeks. Stocks that tended to outperform the market during accelerated AI usage and underperform during deceleration were categorized into the high-exposure group.

They weekly went long on the high-exposure group while shorting the low-exposure group. The backtest revealed an average weekly difference of 0.641 percentage points between the two groups; even after controlling for factors such as company size, valuation, profitability, and stock price momentum, the difference remained around 0.56 percentage points.

However, the sample only covered 28 months, OpenRouter only captures a small fraction of global AI usage, and its users are more developer-centric. Whether this approach can remain effective in the long run is still to be validated.