Key facts
- Researchers asked three generations of ChatGPT for twenty next-day stock picks on every US trading day from May 2024 to December 2025.
- The picks underperformed the sell-side analysts they were benchmarked against over the 20-month period, per the report.
- About one in three recommendations rested entirely on the model's training data, the researchers said.
- Among picks that used retrieved pages, the newest source was more than 100 days old in a quarter to three fifths of cases.
A study by researchers at RMIT University, Monash University and Sun Yat-sen University found that stock picks from three generations of ChatGPT underperformed the sell-side analysts they were benchmarked against over a 20-month period, according to a report published by RMIT University.
The researchers asked the chatbot, through the same public website a household would use, for the twenty stocks it would buy the next day on every US trading day from May 2024 to December 2025. They supplied no documents, screening model or expert framing, and compared each recommendation with sell-side analyst recommendations issued on the same day.
The report said the chatbot returned a tradeable list of twenty stocks on the great majority of days it answered. It said about one in three recommendations rested entirely on the model's training data, and that the share did not fall from GPT-4 to GPT-4o to GPT-5.
Among recommendations that used retrieved pages, the newest source was more than 100 days old in a quarter to three fifths of cases, depending on the model, the report said. It said same-day articles existed for roughly seven in ten recommendations.