In a first paper presented by Alejandro Lopez-Lira (University of Florida), the authors document the capability of LLMs, like ChatGPT, to predict stock price movements even without direct financial training. Using a data set of news headlines related to the U.S. stock market, they show that LLMs have significant predictive power in asset markets. Their analysis also suggests that LLMs can complement human decision-making through their enhanced information processing capabilities, thus reducing market inefficiencies and altering information diffusion dynamics in the economy. 

The second paper presented by Luciano Somoza (ESSEC Business School) documents limits to the rationality of LLMs. In particular, the authors show that frictions in the algorithm can degrade the performance of LLMs. The authors analyze this phenomenon in two different contexts: i) earnings forecasts from corporate calls and ii) predictions of market reactions to stock-specific news. They show that the predictive accuracy of LLMs follow an inverted U-shaped pattern, suggesting that beyond a certain threshold, additional context degrades predictive accuracy of the models rather than improving it. This characterizes information overload: more data is not always better.