LLMs Can Teach Themselves to Better Predict the Future

Large language models (LLMs) have demonstrated remarkable capabilities in a wide range of areas, often approaching or exceeding human performance. One area where human performance has not yet been surpassed is judgemental forecasting [1], where probabilistic forecasts are assigned to future events. Successful forecasts by top-performing human forecasters include substantial reasoning about facts of the world, various trends, and competing pieces of evidence [2], making it a great place to study model reasoning capabilities in a messy real-world environment. Moreover, forecasting is a central task in decision-making across sectors as diverse as finance, policy, and law. It is central to inform resource allocation, manage risks, and plan organizational decisions. Modern LLMs have already been shown to conduct financial analysis [3], evaluate the impact of events on time series [4], and improve climate policy decision-making [5]. This makes improving LLMs' forecasting abilities potentially impactful and wide-ranging.