Can Language Models Forecast Inductively?

Judgmental forecasting requires integrating sparse evidence with knowledge about the actors, institutions, and processes that shape how events unfold. The same observation can imply different futures depending on these underlying relationships. Here, we investigate whether language models (LMs) recover those relationships or rely on surface regularities. Across a battery of experiments spanning real-world prediction markets and controlled synthetic environments, we test whether models update their beliefs coherently, recover latent predictive relationships, and transfer what they learn to new settings. Our results provide evidence for each of these capabilities. Models selectively update their forecasts in response to causally relevant evidence, use contextual information to infer predictive relationships when observations are sparse, and revise these inferences as direct evidence accumulates. Training on families of predictive relationships further improves forecasting on unseen structures, although transfer is incomplete in more challenging settings. Together, these experiments provide behavioral evidence that LMs can forecast inductively: their predictions respond to evidence and context in ways that reflect underlying predictive structure, and some of this behavior transfers beyond the settings in which it was learned.