What is Bayesian win rate? — CrowdIntel

In detail

Raw win rate (wins / resolved_bets) is noisy for small samples. A wallet with 2 wins in 2 bets shows 100% raw win rate but has almost no statistical weight. Bayesian win rate applies a beta-binomial update: start from a prior centered on the category base rate (usually ~50% for binary markets), then update with observed wins and losses. The result is a shrunken estimate that converges to raw win rate as sample size grows but stays near the prior for tiny samples. This is the single biggest defense against small-sample noise in prediction-market analytics.

How CrowdIntel measures it

CrowdIntel stores bayesian_win_rate, confidence (standard error inverse), and z_score columns on every wallet_stats row, refreshed when the OutcomeTracker resolves new bets. The ConfidenceMeter UI component shows both raw and Bayesian win rate side-by-side. Sort and filter on leaderboards use Bayesian win rate by default.

Frequently asked

Why use Bayesian instead of raw win rate?

Raw win rate punishes long-track-record sharps and rewards lucky newcomers. A wallet with 80% over 200 bets is far more credible than a wallet with 90% over 10 bets. Bayesian methods mathematically formalize that intuition.

What prior does CrowdIntel use for the beta distribution?

A symmetric Beta(α=10, β=10) centered on 50% — equivalent to 20 hypothetical coin flips. This pulls small-sample estimates toward 50% and only lets data override the prior when the sample is large.

How much data until Bayesian win rate approaches raw win rate?

At ~100 resolved bets, the posterior is dominated by data (the prior contributes <20%). At 20 bets the prior still contributes ~50%. This is why investigations require ≥20 bets and the UI shows confidence bands.