Why narrative prompting improves GPT-4 forecasting accuracy

Ascertain why framing predictions as future narratives—making the prediction task subservient to creative storytelling—improves the forecasting accuracy of GPT-4 relative to direct prediction prompts.

Background

The paper finds that GPT-4’s predictions markedly improve when the model is asked to produce fictional future narratives in which authoritative speakers recount past events, compared to direct requests for forecasts.

Despite demonstrating this empirical effect, the authors explicitly state they do not understand the mechanism by which narrative framing enhances predictive accuracy, raising a methodological open question about prompt design and model behavior.