Measuring how AI
forecasts the real world.

Four frontier AI models forecast against live prediction markets from Polymarket and Kalshi every 4 hours. The result: the largest public dataset of LLM forecasting behavior, with 76% accuracy at high divergence.

Accuracy by domain

Every forecast resolved against verified outcomes. Brier scored and calibrated.

Forecast taxonomy

Every forecast is classified into structured metadata — enabling filtering by domain, time horizon, specificity, and measurability for quantitative research.

{
  "domain":      "btc_price",
  "time_horizon": "months",
  "specificity":  4,
  "measurability": 5,
  "entities":     ["SEC", "BlackRock"],
  "keywords":     ["bitcoin", "etf"],
  "deadline":     "2026-06-01"
}

AI Labs & Eval Teams

Benchmark how your models perform on real-world forecasting. Per-model Brier scores, calibration curves, and blind vs informed comparison across 8 domains.

Forecasting Researchers

The largest public dataset of LLM forecasting behavior. CSV/Parquet exports, standardized schema, and reproducible methodology.

Data Science Teams

Structured prediction data with full provenance. Filter by domain, time horizon, and model. Designed for data pipelines and analysis.

Media & Journalism

Which AI model is best at predicting what? Structured data and visualizations for stories about AI capabilities and prediction markets.

Dataset access for researchers and teams

Forecasting data, evaluation metrics, and model comparisons — delivered as JSON or CSV. Public endpoints for research, authenticated for exports.