Every knockout round produces the same ritual. A favourite goes out, and within minutes the timelines fill with the same verdict: "so much for the odds." The model was wrong. The market was clueless. The experts got humbled again.
It's the most confident sentence in football, and it's almost always innumerate.
We build forecasting AI and run it in public against real markets, so we watch this happen every single matchday. Here's the thing nobody selling you a "90% accurate" prediction wants you to understand: a good forecast and a wrong outcome are completely compatible. In fact, if your favourites never lost, your forecasts would be broken.
What a probability actually promises
When a model — or a market — says a team has a 70% chance to win, it is not saying they will win. It is making a much humbler and much more testable claim: *across all the matches I price at 70%, the favourite should win about seven times out of ten.*
Which means it should also **lose about three times out of ten.** Not as a failure. As the whole point.
If you lined up a hundred matches where the favourite was priced at 70%, and the favourite won all hundred, the model wouldn't be a genius — it would be badly miscalibrated. It was calling those teams 70% when they were really 95%. The losses aren't the bug. They're the receipt that the number meant what it said.
So when a 70% favourite crashes out of the quarter-finals, that is not evidence against the forecast. It is one of the three-in-ten we told you to expect. The only question worth asking is whether, over hundreds of these, the sevens and the threes land in the right proportion.
Football is built to humiliate forecasters
Some sports are kind to prediction. Basketball has hundreds of scoring events per game, so luck averages out and the better team usually wins. Football is the opposite: low-scoring, high-variance, decided by a single deflection, a marginal offside, a goalkeeper's one mistake in ninety minutes.
That structure means even genuinely dominant teams lose constantly. A side that is truly 70% to win — a strong, correct, honest 70% — will still be knocked out three times in ten by nothing more than the sport doing what it does. Add a penalty shootout, which is close to a coin flip between exhausted teams, and the knockout rounds become an upset factory by design.
None of that makes the favourites weaker than we thought. It makes football exactly as chaotic as the probabilities already said it was. The forecast priced in the chaos. The viewer just forgot it was there until it arrived.
So how do you actually judge a forecast?
Not by whether one call landed. A single result tells you almost nothing — a broken clock calls one match right too. You judge a forecaster the way you'd judge a casino: over a large number of bets, using a proper score.
The tool for this is the **Brier score**: take the probability you stated, subtract what actually happened (1 or 0), square it, and average over every call you ever made. Lower is better. A coin-flipper lands around 0.25. The genius of squaring the error is that it punishes confident wrongness far harder than humble wrongness — say 95% and miss, and it hurts roughly ten times more than saying 70% and missing. The metric structurally rewards knowing what you don't know.
Crucially, you have to score **everything** — every call, winners and losers, timestamped, with nothing quietly deleted. A track record that only remembers its best nights isn't a track record. It's an advertisement.
The honest version, in public
This is exactly the standard we hold our own AI to, out in the open. Before every World Cup match, our model's probabilities are locked and hashed into Bitcoin — so nobody, including us, can backdate or edit them. At the same moment we freeze the prediction-market price as a benchmark. After the final whistle, both get Brier-scored, and the result goes on a public board. Wins and losses. No delete button.
And the honest scoreboard is humbling: through the group stage, the market has been closer than our model on most nights. Our model only pulls ahead on average error because it refused to write off a couple of "impossible" outcomes that the crowd had rounded down to zero — and those landed. That's not a money printer. That's what an honest fight with an efficient market actually looks like.
Which brings us back to the favourites going out this week. When the next one falls and your feed declares the odds were garbage, you'll know the more interesting truth: the odds probably weren't wrong at all. You just watched one of the threes.
The forecast never promised you certainty. It promised you honest uncertainty — and honest uncertainty, unlike the guy selling 90% accuracy, occasionally has the decency to be wrong out loud.