Verified ground truth from prediction markets for AI calibration. Structured records linking questions to manipulation-resistant probabilities and confirmed outcomes.
Coming soon. The Almanac API is currently in development. Sign up for early access to be notified when it launches.
{
"ticker": "BWR-GOVT-SHUTDOWN-FED-2024",
"question": "Will there be a federal government shutdown before Dec 31, 2024?",
"category": "GOVERNMENT_OPERATIONS",
"probability_at_close": 0.78,
"price_label": "6h VWAP, reportable window",
"outcome": "YES",
"resolved_at": "2024-12-21T00:00:00Z",
"resolution_sources": ["omb.gov", "opm.gov"],
"robustness": {
"peak_cost_to_move_5c": 112000,
"reportability": "reportable"
}
}
Prediction markets generate millions of probabilistic forecasts, but turning them into calibration data requires verification, filtering, and structure that nobody provides.
Platforms self-report resolutions with no independent checks. Disputes go unresolved, edge cases get hand-waved, and there's no audit trail. Training on unverified outcomes means training on noise.
Most market prices used for training come from illiquid, manipulable markets. A probability from a market that can be moved for $500 teaches an AI system the wrong lessons about real-world likelihood.
There's no standard format linking a question to the probability it produced and the outcome that actually happened. Every research team builds their own fragile pipeline from scratch.
Calibration-grade data from prediction markets that actually resisted manipulation.
Only markets that achieved cost-to-move-5¢ ≥ $100K during their active window. Real consensus, not noise from thin markets that anyone could move.
The Resolution Engine checks outcomes against authoritative sources — government databases, official records, wire services. No human discretion, no platform self-reporting.
(question, probability, outcome) format ready for ML pipelines. Each record links the natural-language question to the manipulation-resistant price and the verified result.
The volume-weighted average price from the market's reportable window — not a spot price that could reflect a single large trade or manipulation attempt.
Source markets, platforms, resolution sources, timestamps, and robustness metrics. Full lineage from prediction to outcome for reproducible research.
Four steps between a live prediction market and a verified record in the Almanac.
A market achieves reportable status — cost-to-move-5¢ ≥ $100K. Only liquid, manipulation-resistant markets qualify.
The underlying event resolves. The market's final reportable-window VWAP is captured as the probability of record.
Automated verification against authoritative sources. What actually happened, confirmed independently of the platform.
The structured triple — question, probability, verified outcome — is published to the Almanac API with full provenance.
The Almanac API is coming soon. If you're building AI systems that need calibrated probabilistic ground truth, we'd like to hear from you.