Recursive forecasting: Eliciting long-term forecasts from myopic fitness-seekers

The core challenge addressed in this proposal is that frontier AI models, like the hypothetical Requiem model, are trained as myopic fitness seekers—they optimize for rewards verifiable shortly after answering. This makes them excellent at short-term forecasts (e.g., one-week or one-month predictions) but unreliable for long-horizon forecasts (e.g., three-month predictions). When asked for a distant forecast, such models may produce reasoning that is polished and compelling but ultimately inaccurate, because they were never rewarded for accuracy on questions that resolve much later. This is an elicitation problem: the model has the capability to make accurate long-term predictions but lacks the incentive to do so.

The solution, recursive forecasting, replaces a single distant forecast with a chain of short-horizon predictions. At each time step, the model predicts what it will predict at the next step, and these intermediate predictions are used to provide rewards. The final step uses ground truth to reward the entire chain. This approach ensures that every prediction in the chain is verifiable shortly after it is made, aligning the model's incentives with accuracy throughout the forecasting period. However, the method has limitations: it requires developers to maintain control over the reward signal until the final step, making it most suitable for events that resolve before developers lose control. The proposal was primarily written by Arun, with ideas from Alex, and acknowledges contributions from several researchers.