Did Polymarket Beat a Simple Elo Model at the 2026 World Cup?

Scoring $2.67 billion of prediction-market prices against a rating table you could compute by hand, across 101 matches.

1 August 2026 · 7 min read

Every 2026 World Cup match as one square, coloured by whether Polymarket and an Elo model both called it correctly — Each square is one match, in date order. The two forecasts named the same most likely outcome in 90 of 101 matches; the eleven that differ are picked out in colour.

90 / 101

Matches where both agreed

62.4%

Pick rate each — ceiling was 71.3%

0.501

Market Brier, vs 0.528 for Elo

28.7%

Draws — market priced 22.8%

Across 101 matches of the 2026 World Cup, Polymarket's three-way markets and a plain Elo rating — pre-tournament ratings, no updating, no injury news, no tactics — named the same most likely outcome 90 times. Where they disagreed, they split four apiece and both missed the other three. Both finished on a 62.4% pick rate.

About $2.67 billion changed hands across those markets, a median of $25.4 million per game. It did not buy a single extra correct call.

It is worth knowing what 62.4% means here. Draws happened in 28.7% of matches, and neither forecast names a draw as its single likeliest outcome more than twice — so the ceiling was 71.3%. Both landed at 87.5% of what was reachable.

They got there differently. On the 72 matches that produced a winner, the rating table was the better picker: 87.5%against the market's 84.7%. The market drew level only because it twice named a draw as the likeliest result and was right both times. The model never named one at all.

What the money did buy

The market was still better. Just not at picking winners.

Scored on Brier — which rewards saying 70% when you mean 70%, rather than merely getting the ranking right — the market came in at 0.5014 against the model's 0.5278. A no-skill forecast that puts a third on each outcome scores 0.6667.

Bar chart of Brier scores: Polymarket 0.501, Elo model 0.528, no-skill 0.667 — Brier scores across all 101 matches. The market's advantage is 0.026, about 5% relative.

The advantage is consistent. The market posted the better score in 67 of 101 matches, a sign test at p = 0.001, and log loss agrees on direction: 0.847 against 0.891.

It is also small, and honesty requires saying so plainly. A paired test on the average gap gives p = 0.11, with a bootstrap 95% interval of [−0.006, +0.058] that includes zero. The market wins often, by a little. It loses rarely, by a lot. Those rare losses eat the average.

We stress-tested the baseline rather than leaving it a soft target. Its draw parameter is fixed at the historical 25% rate for tournament football, not fitted to 2026 — fitting it would hand the model information nobody had in June. Refit it to the draw rate this tournament actually produced and it improves to 0.5245. The market still wins. The result is not an artefact of a handicapped baseline.

Does 70% mean 70%?

Group every match by how confident each forecast was in its own favourite, then check how often that favourite actually won.

Calibration curve for Polymarket and the Elo model plotted against the perfectly-calibrated diagonal — Calibration against the diagonal. Points above the line mean the forecast was underconfident; below means overconfident. Bucket sizes are labelled because several are thin.

Both track the diagonal reasonably. The market sits slightly above it through the middle of the range, meaning it was mildly underconfident: when it said 65%, favourites won about 83% of the time. The rating model wanders further in both directions. The one place the market falls below the line matters, and it has a name.

The blind spot both forecasts shared

Draws happened in 28.7% of matches. The market priced them at 22.8%. The Elo model said 21.4%.

Grouped bars showing predicted versus actual shares of home wins, draws and away wins — Predicted versus realised share of each outcome. The crowd improved the draw estimate by 1.4 points against an error of roughly 6.

The crowd inherited the naive model's blind spot almost intact. Whatever collective intelligence $2.67 billion buys, it did not buy an understanding of how often World Cup football ends level.

So where did the edge come from? Take the Elo model and swap in only the market's away-win probabilities, leaving the rest naive. The score improves from 0.5278 to 0.5112— roughly two-thirds of the market's entire advantage, from one dimension. The market's real skill was knowing when a visiting side would win. On draws it was barely better than arithmetic.

One failure mode, six times

The market's six worst individual calls all look the same.

Fixture	Market gave the favourite	Result
Spain – Cabo Verde	90.0%	0–0
Argentina – Cabo Verde	86.3%	2–2, extra time
Ecuador – Curaçao	84.0%	0–0
England – Ghana	83.9%	0–0
Qatar – Switzerland	82.7%	1–1
Portugal – DR Congo	76.1%	1–1

Every one is a heavy favourite failing to win inside 90 minutes. Across the 14 matches where the market priced a favourite between 80% and 90%, it expected 11.9 wins and got 9 (binomial p = 0.043 — one bucket among several, so treat it as a lead rather than a finding).

Horizontal bars showing six matches where the market gave the favourite 76 to 90 percent and the game ended level — The six matches contributing most to the market's total error.

The interesting part is that the market cansee a stalemate coming. Before Algeria–Austria it priced the draw at 46.7% against the model's 24.6%. The match finished 3–3. The crowd knows how to price a grind. It just stops doing it when a giant plays a minnow.

The knockout reversal

Split the tournament and the picture inverts.

Pick rates by stage: the market leads in the group stage and trails in the knockout rounds — Pick rate by stage. In the knockout rounds the market called fewer matches correctly than the naive model.

In the 70 group matches the market clearly outperformed on Brier too, 0.4820 against 0.5143. In the 31 knockout matches its Brier edge nearly vanishes (0.5452 against 0.5584) and its pick rate falls below the baseline.

With 31 matches this is thin, so we are not claiming the crowd wilts under pressure. But there is a mechanical reason it might be real: knockout ties that finish level go to extra time, and a market resolving on the 90-minute result has to price a draw it may not want to price. We did not exclude the knockout rounds, because excluding them after seeing the split would be choosing the answer.

The margin was essentially zero

Sum the three legs of each match at the last pre-kickoff price and the mean is 1.0013 — thirteen basis points of overround. The median is 1.0050. A sportsbook three-way market typically runs 5 to 7 percent over.

Histogram of the summed three-leg price per match, clustered tightly around 1.00 — Distribution of the summed three-leg price. In 37 of 101 fixtures the legs summed below fair value.

That is not proof of free money: these are historical price points rather than simultaneous executable quotes, and legs can be stale relative to each other. But it does say something about how tightly this market was priced.

Methods

Prices

Polymarket CLOB three-way match markets (“Will X win”, “Will it end in a draw”, “Will Y win”), pulled at 15-minute fidelity and normalised across the three legs. The evaluated price is the last complete bucket before kickoff, which lands 14.95 minutes before the whistle for all 303 legs — so this is the market's view a quarter of an hour out, not at the whistle. Nothing from open play enters the score. Each leg had a median of 48 price points in the preceding twelve hours, so no quote is stale.

Results

A 2026 World Cup match dataset. Any match decided in extra time or on penalties was level after 90 by definition, so it counts as a draw. These markets resolve on the 90-minute result: the Spain–Argentina final settled as a draw despite Spain lifting the trophy on a 106th-minute goal. We validated the rule against Polymarket's own settlements for all 101 matches and found zero disagreements.

Baseline

Davidson three-outcome model on pre-tournament Elo ratings, with the draw parameter fixed by the historical 25% rate at even strength. No home-advantage term, since World Cup venues are neutral apart from the three hosts.

Exclusions

Three matches had no locatable Polymarket market: South Korea–Czechia, Mexico–South Korea, and USA–Bosnia and Herzegovina.

A warning for anyone using the same dataset

Its kickoff_time_utc column is unreliable. Checked against Polymarket's own gameStartTime, it was off by anywhere from −9.5 to +6 hours and exactly right in only 9 of 46 cases. Our first run trusted it, which pushed the price window past full time and produced “forecasts” like [0.999, 0.0005, 0.0005] — post-match settlements dressed up as predictions, scoring a triumphant 82.6% accuracy. If a prediction-market backtest looks brilliant, check the timestamps before checking the thesis.

What this is worth

One tournament, 101 matches, a small and statistically shaky average edge. Nobody should conclude from this that prediction markets are or are not efficient.

What survives is narrower and more useful. Money moved the probabilities in the right direction, mostly along one dimension, without improving the pick rate at all. And a systematic error that a rating table makes — refusing to price the stalemate when a big team meets a small one — survived contact with $2.67 billion.

That is a more interesting thing to know than whether the crowd won.