How Our Model Did in the Group Stage — An Honest Look at 72 Matches of ELO Predictions

Prediction models exist to be wrong occasionally. The point is not to be right every time — that would require genuine foreknowledge of the future, not a mathematical framework for quantifying historical performance. The point is to be calibrated: to have your 70% predictions win roughly 70% of the time, your 30% predictions win roughly 30% of the time, and your 50/50 calls split roughly half and half. With the World Cup 2026 group stage now complete, we have 72 matches of data against which to measure the FootUps ELO model. Here is what it got right, what it got wrong, and what the honest analysis looks like.

The Headline Number: 66.7% Accuracy

The model predicted the correct match outcome — win, draw, or loss — in 48 of 72 group stage matches. That is a 66.7% hit rate. The obvious question is: how does that compare to the baseline?

In any international football tournament, a naive model that simply picks the higher-ranked team to win every match gets roughly 55-60% correct, because draws are frequent and upsets are real. A model that adds draw probability and calibrates properly should expect to do better than that. A 66.7% hit rate against a naive baseline of around 57% represents a meaningful improvement — roughly one fewer wrong prediction per five matches.

For context: commercial bookmakers, whose implied probabilities are derived from markets involving millions of pounds of information, typically achieve 67-70% accuracy in group stage match prediction at major tournaments. The FootUps model, built on ELO ratings with squad quality adjustments and tournament form correction, sits at the lower end of that range. That is not a boast, but it is not an embarrassment either. It is a fair result for a model that is publicly visible, transparent about its methodology, and does not have access to injury news or private squad information.

Where the Model Was Confident and Right

When the model gave a team more than 60% chance of winning, that team won 27 times from 38 attempts — a 71.1% conversion rate. This is the model's strongest signal: when the ELO gap is large enough to produce a 60%+ win probability, the favourite wins nearly three times out of four. The six times out of ten that it didn't were the majority of our notable misses, and most of them were draws, not underdog wins.

Examples of the model's most accurate strong-favourite calls: Argentina beating Algeria (pre-match 73.9%, result 3-0), Mexico beating South Africa (72.4%, result 2-0), Switzerland beating Bosnia 4-1 (61.4%, the Bosnians offering some resistance but quality winning out). These are the matches that don't make headlines precisely because the model was right.

The Twenty-Four That Got Away — and What They Have in Common

Of the 24 matches where the model's top prediction was wrong, 14 were draws. Only 10 were genuine underdog wins. That distinction matters: a draw in a match the model predicted as a win is a calibration miss — the draw probability should perhaps have been higher — but it is not a failure to identify the better team. The better team usually did dominate those matches. They simply failed to convert that dominance into three points.

The ten genuine underdog wins tell a different story. Here are the most significant, ordered by how improbable the model considered them:

Match	Result	Winner's probability
Ecuador vs Germany	2–1 to Ecuador	6.8%
South Africa vs Korea Republic	1–0 to South Africa	12.3%
Turkey vs USA	3–2 to Turkey	17.8%
Paraguay vs Turkey	1–0 to Paraguay	25.8%
Mexico vs Korea Republic	1–0 to Mexico	39.0%
DR Congo vs Uzbekistan	3–1 to DR Congo	37.3%
Australia vs Turkey	2–0 to Australia	35.9%

Ecuador's 6.8% win over Germany is the tournament's most extreme probability miss. South Africa's 12.3% win over Korea Republic is the second. Together, they account for two of the three biggest upsets in the group stage and represent matches where something happened that the model simply couldn't anticipate: extraordinary individual performances, tactical decisions that caught opponents unprepared, or the kind of single-day over-delivery that any human performance is capable of but that no rating system can forecast in advance.

The Draw Problem

The model's biggest systematic miss was not outright upsets but draws — particularly draws in matches where a heavily-favoured European team failed to convert expected superiority. Spain 0-0 Cape Verde (Spain were 79.3% to win). England 0-0 Ghana (73.6%). Belgium 0-0 Iran (59.1%). Portugal 1-1 DR Congo (76.2%).

In each of these cases, the model correctly identified the better team. The match simply produced the outcome that the model had assigned roughly 18-20% probability. When you have 72 group stage matches with draw probabilities averaging around 22%, you should expect approximately 16 draws by chance alone. The tournament produced 16 group stage draws. The model's draw probability calibration was, in aggregate, close to correct — it just wasn't possible to know which of the 72 matches would produce them.

The improvement to make here is not in how draws are modelled on average, but in identifying which specific match types are most likely to produce them: high-ranked European side versus compact mid-tier African or Asian team, group stage (lower stakes than knockout), and conditions that favour the lower team's defensive organisation over the higher team's attacking system. Future refinements to the model will attempt to weight these contextual factors more heavily.

Comparing to Bookmaker-Implied Probabilities

Bookmakers don't publish explicit accuracy statistics, but their implied win probabilities (derived from odds) are available and have been compared to actual outcomes by academic researchers across multiple tournaments. The typical finding is that commercial bookmakers achieve 67-70% group stage accuracy at major international tournaments, with slightly higher accuracy for knockout rounds where draw probability is compressed.

The FootUps model at 66.7% sits at the boundary of that range. In the group stage, the model and the bookmakers were roughly equivalent in predictive power on most matches. The main differences appeared in matches involving unexpected upsets: the Ecuador-Germany and South Africa-Korea matches had market-implied probabilities for the underdog of roughly 8-10% and 14-16% respectively — slightly higher than our model, suggesting the market assigned more uncertainty than we did. In those specific matches, the bookmakers were better calibrated.

More meaningfully: for the matches where both the model and the market agreed on a heavy favourite (10 or more such matches), both the model and bookmakers were correct approximately 74-75% of the time. When a team is that heavily favoured, the signal is strong enough that being slightly more or less confident doesn't change the outcome often. The value in a model like FootUps is in the medium-probability range — the 40-55% win probabilities — where the calibration quality actually distinguishes models from each other.

The Knockout Stage: A Tighter Sample

The knockout stage has produced 13 results to date (16 last-thirty-two matches plus the France-Morocco quarter-final on July 9). The model's accuracy across those 13 is 10 correct from 13, or 76.9%. The three knockout misses include Norway beating Brazil (model: 17.3% for Norway), Morocco beating the Netherlands (22.8%), and Switzerland beating Colombia (46.8% — a near-coin-flip that went the non-favourite's way).

The 76.9% knockout accuracy is higher than the group stage figure, partly because knockout matches remove the draw option from the final outcome calculation (there must be a winner), and partly because the strongest teams tend to assert themselves more decisively in the knockout rounds. The sample is also smaller — 13 matches versus 72 — and should be interpreted with corresponding uncertainty.