AI challenges Wall Street stock pickers.
A University of Florida professor is testing ChatGPT, DeepSeek, and Grok to see if AI can outperform human stockbrokers. Early results from backtesting simulations are promising, suggesting AI has significant predictive power in asset markets. While real-world application still requires human guidance, the potential for AI to automate many tasks currently performed by analysts is becoming increasingly clear.
Points clés
- Alejandro Lopez-Lira, an assistant professor of finance at the University of Florida, is testing AI models for stock picking.
- He is experimenting with ChatGPT, DeepSeek, and Grok.
- Lopez-Lira was impressed with AI’s ability to interpret news headlines as positive or negative for a stock.
- A backtest simulating historical returns using over 134,000 headlines for over 4,000 companies showed ChatGPT had “significant predictive power.”
- The GPT-4 version showed an average daily return of 0.38% and a compounded cumulative return of over 650% from October 2021 to December 2023 in the simulation.
- Approximately 76% of the simulated gains came from short sales, a strategy with inherent complexities in the real world.
- Lopez-Lira partnered with Autopilot, an investment app, to test ChatGPT’s picks in a real-world setting starting in September 2023.
- Real-world application requires a human to provide the AI with comprehensive information, as AI models are not trained on real-time data.
- In February, investing accounts using Grok and DeepSeek were added to the Autopilot experiment.
- The latest AI models being tested are OpenAI’s o3, xAI’s Grok 3, and DeepSeek R1.
À retenir
So, it seems the robots are coming for Wall Street’s jobs, or at least a good chunk of them. If a professor and some chatbots can potentially pick stocks better than seasoned professionals (in a simulation, of course, with no pesky real-world costs!), maybe it’s time for stockbrokers to start practicing their “Would you like fries with that?” or perhaps brush up on their AI prompting skills. After all, who needs a fancy suit and a corner office when you have a Python script and access to a large language model?
Sources
