We examine the potential of ChatGPT and other large language models (LLMs) to predict stock market returns using news headlines. ChatGPT scores significantly predict daily stock returns, outperforming traditional methods. A model involving information capacity constraints and LLMs rationalizes this predictability, which strengthens among smaller stocks and following negative news and further predicts LLMs' wide availability will improve market efficiency. Only advanced LLMs accurately interpret hard-to-read news and deliver higher Sharpe ratios, while basic models cannot accurately forecast returns, suggesting return forecasting is an emerging capacity of bigger LLMs. We introduce an interpretability technique to evaluate LLMs' reasoning.