How we actually know.

The archive covers 413,212 settled Kalshi markets. For each one, the YES price is sampled at three fixed moments before close — 1 hour, 24 hours, and 7 days — read from hourly candlesticks, producing 512,797 price–outcome pairs.

Price selection follows a strict order: the last traded price is preferred; a bid/ask midpoint is used only when the spread is 20¢ or narrower; wide or empty books are skipped rather than filled in, which tilts the sample toward liquid markets. No price is ever fabricated.

Not every market yields all three pairs — a horizon that predates a market’s open is structurally absent, which is why the pair count is well below three per market.

A single game can spawn dozens of correlated markets — moneyline, spread, props, in-play ladders — all resolving on the same underlying outcome. Treating each as an independent observation overstates confidence: the effective sample size is the number of events, not the number of tickers.

We cluster observations by underlying event and compute confidence intervals with a cluster bootstrap over events (2,000 resamples), rather than a naive per-observation interval.

We count games, not tickers.

• Scalar and void settlements — no binary YES/NO outcome to score.
• Zero-volume markets — no real trading occurred.
• Buckets under 30 distinct events — too little data to say anything honest.
• Floor-pinned 1¢ prices in the cheapest bucket — Kalshi’s minimum tick would otherwise guarantee apparent overpricing regardless of the true rate, since a market can’t price below 1¢ even if its true probability is lower.

Last-trade prices embed the cost of crossing the book, not just the market’s belief. On structurally two-outcome events, quoted midpoints sum to about 100¢ — validating the method — while last-trade prices on the same events sum to about 101–102¢.

That gap means trade-price deviations reflect a buying premium, not necessarily market error. Findings are labeled by price source accordingly.

The results span 103 price buckets across categories and horizons. At 95% confidence, a handful of isolated interval breaches are expected by chance even if every bucket were perfectly calibrated. Headline findings are drawn only from consistent runs across adjacent buckets — never a single outlier cell.

Coverage begins May 2026 and is a rolling window, not a fixed historical archive. Prices are sampled at hourly candle granularity, so intra-hour moves aren’t captured. Economics buckets are quote-dominated and marked exploratory. Current coverage spans Sports, Climate & Weather, Economics and Financials; remaining categories are accumulating nightly.

Throughout: these results describe historical aggregates, not predictions. A bucket’s past calibration is not a forecast for any individual market.