Polymarket’s Binary Markets Are Well-Calibrated. Here’s What “92% Accuracy” Misses.

Binary yes/no markets track reality. Multi-outcome markets complicate the headline stat. Here’s the on-chain data.

Data: Polygon via Allium · Jan 2022 – Feb 2026 · Inspired by McCullough’s Dune dashboard

108,861

Resolved tokens

~28,145 actual questions

+2.6 pp

Binary bias (10–90%)

Events happen more than prices predict

-5.4 pp

Worst multi-outcome bucket

40–50% range

75%

Multi tokens in tails

Padding the denominator

Key Findings

1.Binary markets are well-calibrated. Prices track reality across probability ranges with no longshot bias — slightly conservative, if anything.
2.Multi-outcome markets overprice by 3–5pp in the competitive range. A token priced at 45 cents wins just 39% of the time.
3.The “92% accuracy” headline tells an incomplete story. Like crediting a forecaster for predicting no snow in Phoenix in July — near-certain outcomes dominate the count.
4.Binary and multi-outcome markets have opposite biases. Binary prices underestimate event probability; multi-outcome prices overestimate it. Mixing them creates an artifact.

Above the diagonal = underpriced. Below = overpriced. Band = 95% CI for binary.

Positive = underpriced. Negative = overpriced. Midpoint shown on axis.

Indigo = binary Rose = multi-outcome Units: percentage points

Markets priced 10–90% only. Indigo = binary, rose = multi-outcome.

Accuracy for markets priced 10–90%, by quarter (2023+). Solid = binary, dashed = multi-outcome.

Multi-outcome tokens outnumber binary 7:1, but represent roughly equal question counts.

By token count

By question count