How well can large language models predict the future?