The quarter-finals are done, the field is down to four, and this week I did something I had been putting off: I let the model learn from the knockouts. Until now it froze every team’s strength at the end of the group stage and just tracked the bracket. Sensible, but half-blind. It could see who survived. It could not see how they were playing.
So I fed the knockout results back into the ratings. Now it knows the difference between winning well and just winning. And the instant it could tell them apart, it moved the favourite off Argentina.
Here is where it lands, what the upgrade changed, and the one call where I now disagree with my own code more sharply than at any point in this tournament.
The state of the race
Conditioned on every result through the quarter-finals, and now with knockout form baked into the ratings, the title picture:
Spain 30.6% · Argentina 27.9% · France 24.8% · gbeng England 16.7%
Four teams left, so those four hold the entire championship probability between them. The single most-likely final the model produces is still Spain vs Argentina, and its predicted champion is now Spain. Read the top two lines carefully, because the order just changed, and it changed for a reason.
The setup, for anyone new here
The engine is an ensemble of two models: a Dixon-Coles goals model and a World Football Elo rating, blended at the scoreline level and run through 10,001 Monte Carlo simulations from here to the final. It does not get nervous. It does not have a favourite nephew. It does maths, and this week it did some new maths.
The upgrade: I let the model watch the football
The old version had a blind spot I had been honest about since the group stage. It locked each team’s strength when the groups ended and then simply cascaded the bracket forward. That meant a team could win three knockout ties by a single goal, riding its luck, and the model would treat it exactly as it treated a team dismantling opponents 3-0. Survival and dominance looked identical on the page.
The fix was to feed the knockout results back into the ratings, so the model now updates on how teams are actually playing, not just on who is still alive. It is a small change in code and a large change in worldview. The moment the model could separate winning well from merely winning, it re-ranked the favourites. That is the whole story of this update.
What the model now sees
Argentina demoted, from 33.6% down to 27.9%. Three knockout wins, yes, but it conceded in every one of them. It survived Egypt and it survived Switzerland rather than dismissing either. The old model rewarded the clean bracket path. The new model looks at the performances underneath and marks them down. The bracket had been flattering it.
Spain surged to number one, from 25.7% up to 30.6%. It took Portugal apart and then took Belgium apart, beating strong teams convincingly. That is what actually moves a rating once the model is allowed to watch. Spain did not back into this. It earned it on the pitch.
A quiet flex I have earned the right to: my pre-tournament top two, Spain and Argentina, are both still standing, and the entire final four is exactly the calibre the frozen forecast flagged before a ball was kicked.
The semis
France v Spain · gbeng England v Argentina
A France v England final is very much alive, and if you have read these updates since the group stage you know exactly which of those two names I have been shouting about for weeks.
Where I overrule the machine, and it is not close
Here is the honest part, and it has sharpened since the last edition, so let me say it plainly.
Two rounds ago I said France and England, both, were operating on a different frequency. The bracket has since answered part of that for me: England drew Argentina, France drew Spain, and my two-team thesis has to survive contact with two very hard semi-finals. So I will narrow it and commit. The model has France third. I have France first, by a distance, and I will die on this hill.
Look at what France has actually done in the knockouts: three games, three wins, zero goals conceded. Sweden, Paraguay, Morocco, all dispatched without a scare. Sharper, deeper, meaner than anyone left. To my eye they are operating a full level above the field, and the eye test has not been this loud all tournament.
The model docks them on two technicalities. They drew the now-strongest side, Spain, in the semi, which is a draw problem, not a quality problem. And their goals scored were only modestly above expectation, a 1-0 over Paraguay where the maths wanted more. Fair, on the model’s own terms. But the model counts goals and opponents. I am watching football, and I am watching the best team in the tournament stack clean sheet on clean sheet on its way to the final.
Model says Spain. I say France, comfortably. One of us is about to be very wrong, and I have never been more relaxed about it. A forecast you only ever nod along with is not worth much. The value is in the disagreement, and in being honest about which side of it you are standing on.
Under the hood
For the curious: the two models are blended at the scoreline level and the remaining tournament is simulated 10,001 times, now conditioned not only on every actual result but on updated strength ratings that absorb knockout form. The rigour from earlier editions still holds. FIFA’s 2026 tiebreaker puts head-to-head above overall goal difference, a break from every previous World Cup, and it is coded properly rather than approximated. The knockout third-place allocation follows FIFA’s official deterministic table. A forecast is only as trustworthy as the plumbing beneath it, and this week the plumbing got an upgrade.
What is next
The semi-finals are loading. France v Spain, England v Argentina, no redraws, no second chances. The model gets re-run after every round, and I will keep publishing where it lands, and where, like now, I loudly disagree with it.
Model says Spain. My eyes say France. We are about to find out who was right.
The Champion Predictor is a quantitative model built for research and curiosity, not betting advice. For round-by-round updates through to the final, subscribe. The next one lands after the semi-finals.
from Risk Premium https://ift.tt/MkmpbSu
via IFTTT
No comments:
Post a Comment