My Model Picked Spain. I Wanted Argentina.

Before the knockout rounds started, I froze a bracket simulation and published its champion: Argentina. I was happy with that answer for reasons that had nothing to do with math; they were the defending champions, and they were the team I actually wanted. The model and my heart agreed, and agreement is a comfortable place to build from.

The model itself is simple on purpose. A bivariate Poisson setup with a Dixon-Coles correction, fit on roughly 200 matches of real 2026 data, refit after every round so it learns as the tournament goes. It priced six markets per knockout game against live Bet365 odds, sized paper bets with fee-aware half-Kelly, and published every number before kickoff. The rule I set at the start was the one that ended up costing me emotionally: the refit model always overrides the frozen one.

Round by round, it drifted. Argentina kept winning, but they kept winning narrowly, and the goals data cared about how you win, not that you did. Spain, meanwhile, kept posting the kind of shot volume and clean sheets that a goals model treats as a signature. By the final wave, my own dashboard put Spain at 72% to beat the team I had picked, published, and privately wanted. I went through the code that night looking for a bug I was half-hoping to find. There wasn't one.

The final finished 0-0 after ninety minutes; Spain won 1-0 in extra time. Two things graded that morning. The advancement call (Spain to lift the trophy, 72%) graded correct. The value bet attached to it (Spain to win in regulation at 44 cents) graded wrong, because the model liked a Spain that scores in ninety minutes and the real Spain needed one hundred and ten. Both outcomes are true at once, and holding both is the whole discipline: the forecast was right, the bet was not.

Here is the full ledger, because a results page that only shows wins is marketing. Across all four knockout rounds the model went 24 of 32 on advancement calls, 75%. The paper bankroll tells a harsher story: down $2,450 in the quarterfinals, up $3,651 in the semifinals, down $2,000 on the final, for a net of -$799. A model can out-predict my gut and still lose to the market after fees, because the market is not my gut; it is thousands of sharper guts, averaged. Beating me was never the real test. Beating Bet365 is, and one tournament of paper bets says: not yet.

What I keep from the month is not the 75%. It is the specific feeling of watching a final while rooting against my own published number, and understanding that the number did not care. I built the model to keep the fan in me honest, and it did its job so well that it hurt a little. I still wanted Argentina at full time. I do not think that is a flaw. The point was never to stop being a fan; it was to know, in writing, exactly what the fan would have gotten wrong.