2026 World Cup · prediction-market microstructure
When a goal hits, which market knows first?
xResidual is a study of how two large real-money prediction markets, Kalshi and Polymarket, priced the 2026 World Cup, benchmarked against the sharp bookmaker line. It answers a question the literature names but hadn't measured on live in-play sports: which venue discovers price first. The answer, across 86 captured matches: Polymarket leads, but the lead is un-harvestable, because the book vanishes at the goal. Everything ran on a millisecond capture pipeline, and every prediction was pre-committed before kickoff and graded in the open.
What's inside
The research note is the headline; the rest is the live apparatus behind it.
Research note
Cross-venue price discovery, the law-of-one-price decomposition, and the methods. The headline artifact.
read →
Forecast board
A model's probabilities for every market, pre-committed and scored against the price on CLV and calibration.
open →
Group stage
All 72 group games, called up front, then conditioned and graded as each one resolves.
open →
Bracket
The model's projected knockout path, filling in with real teams as groups settle.
open →
Scorecard
How the model is actually doing: running accuracy and log-loss vs the baseline, the v1→v2→v3 ladder, and the biggest surprises.
open →
Market lab
Cross-venue microstructure: lead-lag price discovery, order-flow imbalance, and the goal-overreaction tests.
open →
How it works
The full methodology: de-vigging, the goal model, calibration, and the price-discovery estimation.
read →
Repository
All code, the pre-registration committed before kickoff, and the findings log. Public and reproducible.
github ↗
The results
Six weeks of tick capture across 86 matches. Eleven predictions committed before kickoff, graded in public after the final.
6 / 2 / 3pre-reg graded
Eleven calls, committed before kickoff, graded in the open. Six passed, two failed, three inconclusive for documented data reasons. The headline pass is the one I'd least like: the market is better calibrated than my own pre-committed model (Brier 0.487 vs 0.503, slope 1.07). Both failures independently reproduce what 2026 microstructure papers found.
6 pass · 2 fail · 3 inconclusive11 pre-registered
81.0%Polymarket info share
Polymarket discovers price first. It leads 72% of decisive events (median +600ms) and carries an 81.0% Gonzalo-Granger information share across 63 cointegrated matches, leading 61 of 63. Computed on de-vigged mids, so it's robust to the ~59% trade-direction-classification problem that limits other work.
Polymarket 81%Kalshi 19%
0%harvestable at size
And the lead is un-harvestable, which is the point. At each goal the book collapses to ~0.5% depth, so the median 12.0¢ stale-quote gap can't be traded at size in the typical match. A genuine price-discovery lead, not alpha; the collapse-and-refill is a market-maker's defence against toxic flow, adverse selection, observed in real time.
Book depth at the goal~0.5% of normal → median match yields nothing
0.15ppvenues agree
The visible cross-venue gap is mostly margin, not disagreement. De-vigged, Kalshi and Polymarket title prices agree to ~0.15pp; the "5 to 8 cent" gap the press quotes is the house overround (Kalshi ~5.4% vs Polymarket ~3.0%). A relative-value convergence trade on it returned a clean null (Sharpe −1.95), reported as the negative result it is.
House margin (the vig)true belief gap 0.15pp
How it's built
An Elo-plus-squad-value goal model (Skellam goal-difference, Dixon-Coles correction, a format-aware Monte-Carlo sim) serves as an independent reference, not a competitor to the market: it exists to ask where model and market disagree and who is right, checked against a third source before any "edge" is believed. Price discovery is estimated with Hasbrouck information share and Gonzalo-Granger component share on de-vigged mid-prices, separately for the pre-match and in-play windows. Everything is reproducible from public data and covered by 115 unit tests; collection ran 24/7 on an always-on pipeline, so every figure regenerates from a clean clone. The stance is consistent throughout: the market is the subject, not the opponent. Analysis travels because it's analysis; predictions only travel when they're right.