Forecasting Future with AI

🧪 Active Investigation What kinds of real-world events can language models meaningfully forecast?

CURRENT POSITION LLMs can meaningfully forecast structured, institutional, and process-driven events with clear timelines and leading indicators (e.g., elections, political appointments, regulatory approvals, corporate mergers), but perform poorly on chaotic shocks (wars, disasters), reflexive domains (financial markets), or purely individual, private human decisions.

KEY ASSUMPTIONS Forecastability depends more on signal structure than reasoning difficulty News coverage provides sufficient leading indicators for institutional events Calibration is more important than raw accuracy in forecasting

SUPPORTING EVIDENCE The OpenForesight paper shows strongest performance on institutional and process-driven events like elections and appointments OpenForesight evaluation shows highest accuracy and calibration on institutional political events such as elections, confirmations, and government appointments. Prior forecasting research shows calibrated probabilistic models outperform raw accuracy metrics in political and geopolitical forecasting.

OPEN QUESTIONS How well does this generalize beyond news-driven domains? Can models forecast longer time horizons without severe degradation?

WANT TO EXPLORE DEEPER? Read Full Thought → by parag