Everything you might want to know about Odds Raven, consensus probability, and our data.
Odds Raven is a prediction market aggregator. We track markets across Polymarket, Manifold, PredictIt, and Gemini, then compute a cross-platform consensus probability for every question that exists on 2+ platforms. Think of it as a meta-aggregator for probabilistic forecasts.
Some of the markets we track are real money (Polymarket uses USDC, PredictIt accepts up to $850 per position, Gemini Predictions uses USD), and some are play money (Manifold Markets uses Mana). Our consensus probability weights equally by default (Model B), so play-money predictions still count. Some models (A, F) weight by liquidity, giving real-money markets more influence.
Every 6 hours. Fetchers pull fresh data from each platform's public API at 00:00, 06:00, 12:00, and 18:00 UTC. The cross-platform matcher (which finds new matches) runs once daily at 03:00 UTC to save LLM cost.
It’s a single probability computed from all platforms that list the question. Our default is the geometric mean of odds (log-odds pooling), which the IARPA ACE forecasting tournament identified as the most accurate aggregation rule in large-scale controlled testing. Full methodology with citations is on the methodology page.
Six different aggregation methods: A (liquidity-weighted linear), B (geometric mean of odds — default), C (threshold imputation for missing candidates), D (equal-weight linear — MNAR-safe baseline), E (Bayesian Track Record — pending resolution data), and F (volume-weighted log-odds — experimental). Every question page shows all six side-by-side. See methodology for the research behind each one.
Markets are inefficient across platforms. Liquidity differences, audience differences, and lag in information diffusion mean a 15% probability on Polymarket can coexist with 23% on Manifold for the same question. Our cross-platform consensus is our best estimate of the true probability given all the evidence.
Five-phase matcher: (1) parent-event grouping (e.g. both platforms' 2028 Dem Nominee event), (2) candidate alignment within matched parents, (3) LLM-verified binary matching with Claude Haiku, (4) star-graph merge to prevent false transitive matches, (5) Claude Opus validation catches inversions (“Putin out” vs “Putin stays”), scope mismatches, and timeframe mismatches.
A question with 3+ possible answers rather than yes/no. “Who will win the 2028 election?” is multi-outcome. “Will Trump win?” is binary. We show multi-outcome events as single cards with the top 3 candidates and a “+N more” link to the full ranked list.
Filters the listing to only questions that have 2 or more platforms actively tracking them — i.e. where we can compute a real cross-platform consensus. Single-platform questions are hidden.
We track them while they're open, then stop when they resolve. Resolved markets feed into Model E (Brier score leaderboard), which is still collecting data.
All our data is exposed via a JSON API at /api/v1/questions. See each question's detail page for per-question and history endpoints. API docs page coming soon.
No. We're an independent aggregator. We have no commercial relationship with Polymarket, Kalshi, Manifold, PredictIt, or Gemini. We link out to the source platform for every market we display.
A small team. Contact: hello@oddsraven.com. See about.
The only prediction market aggregator with academically validated cross-platform consensus.