Prediction markets now trade the same games on two serious venues. An obvious question nobody had publishable data to answer: when Kalshi and Polymarket disagree on the same MLB team-win market, who's right — and who moves first?
We could answer it because we capture both venues' order books continuously and align the sampled observations by time on one timeline — the same capture that powers our Polymarket–Kalshi MLB matched book. Sixteen days of June–July 2026 MLB, ~23,000 aligned rows per day, every row carrying both venues' bid/ask and the eventual settled outcome.
Everything below is reproducible: the study script ships with the data, and a free single-game sample of the exact dataset is on Hugging Face (Zenodo DOI: 10.5281/zenodo.20816908).
Result 1: Polymarket leads. Decisively.
Cross-correlating mid-price changes on a ~30-second grid, pooled across 309,693 aligned pairs:
- Polymarket's move now → Kalshi's move next step: corr 0.387
- Kalshi's move now → Polymarket's move next step: corr 0.057
- Lead-lag asymmetry: +0.394 (cluster-bootstrap 95% CI [0.323, 0.466])
- Direction positive on 16 out of 16 days, stable across both halves of the window (0.388 vs 0.421)
There is no ambiguity in this sample: on MLB game markets, information shows up on Polymarket first and Kalshi follows. Our earlier three-day pilot reached the same conclusion by three independent methods.
Result 2: divergences close fast — in about one snapshot
We flagged an "episode" whenever the two venues' mids diverged past a threshold, then watched who closed the gap:
| Threshold | Episodes | Median time to half-close | Closed within 10 min | Polymarket led | Kalshi led |
|---|---|---|---|---|---|
| 2c | 1,680 | ~34s | 98.4% | 51% | 37% |
| 3c | 1,133 | ~34s | 98.7% | 54% | 36% |
| 5c | 500 | ~33s | 99.0% | 56% | 32% |
The median divergence half-closes in roughly one 30-second snapshot. Cross-venue disagreement on liquid MLB markets is measured in seconds, not minutes.
Result 3: the "obvious" trade loses — we checked
If Polymarket leads, the tempting strategy writes itself: when the venues diverge, trade the laggard toward the leader. We simulated both directions with real touch prices (entering at the laggard's actual ask, not the mid):
- Trading Polymarket toward Kalshi loses money — about −0.5c per episode net of spread, and the 95% CI excludes zero. Obvious in hindsight: Polymarket is the leader, so this trade converges toward the venue that's behind. Do not do this.
- Trading Kalshi toward Polymarket shows +2.7c to +6.6c on paper — if you act within the same 30-second snapshot as the trigger. Enter one snapshot later and the edge is zero to negative. The convergence is faster than the measurement grid: this is a sub-30-second latency race, not a resting inefficiency. Add Kalshi's real taker fees (~1.75c per contract near 50c) and 1c ticks, and the bar rises further.
The honest headline isn't "free money between venues." It's the opposite, and it's more useful: these markets police each other within seconds, and the residual is a latency game with real infrastructure costs. Anyone selling you a cross-venue arbitrage signal on these markets is selling you a race you have probably already lost.
Honest limits
Quotes-only data (no queue or fill simulation — apply the halve-every-backtest rule); ~30s grid, so faster structure is invisible; one sport, 16 days of one summer; exit-at-mid is optimistic on the simulated trades; no score-event column in this cut, so event-driven vs. noise divergences aren't separated.
Reproduce it
The dataset that makes this measurable — both venues' books, one timeline, outcomes labeled — is what we sell, because neither exchange retains or publishes it, and it cannot be backfilled by anyone starting today (our comparison of historical order-book data sources covers the alternatives).
- Free sample (one complete game): Hugging Face · Zenodo DOI
- $9 tryout tape → full MLB matched-book archive
- Grade your own bets against our captured closes, free: /grader · /grader/kalshi
This is research and backtest data, not a trading signal. We publish our negative results — including this one — because that's the only kind of research worth paying for.