Tontin-BETe Called This Final Back in May. It Simulated It 50,000 Times

On May 28, before a ball was kicked, we published Tontin-BETe's predictions for the World Cup: 20,000 simulated tournaments, Spain first (~13%), Argentina second (~11%), when the market had them fifth or sixth, and a line that reads differently today: "the most likely final is Spain–Argentina, and it's a coin flip."

This Sunday, it's Spain–Argentina.

The model didn't guess a scoreline: it called the exact final matchup a month and a half out. So we did the only thing that made sense: simulate the final. This time, fifty thousand times.

The Numbers of the Final

• Champion: Spain 50.7% vs Argentina 49.3%. A point and a half apart, with nearly identical Elo (1888 vs 1889). There's no real favorite: a single detail decides it.
• In 90 minutes: Spain 35.8% · draw 29.6% · Argentina 34.6%. Almost three in ten finals end level in regulation: 29.6% go to extra time and 15.8% reach penalties.
• Goals: an average of 2.96 per simulated final. Both teams score in 59.7% of cases and Over 2.5 runs at 57%. A dull 0-0 is not projected.
• The most likely scoreline is 1-1 (11.2%), exactly the one that pushes the final to extra time. Next come 1-2 (8.6%), 2-1 (8.3%) and 1-0 (7.8%).

Personal Disclaimer, Part Two

Back in May I wrote that I hated the result and that I hadn't touched a single number. I stand by both. The model gives Spain a point-and-a-half edge: if I correct it so Argentina wins, it stops being a model and becomes a wish. 🇦🇷

The Thesis: the Model Backs Argentina More Than the Market Does

The market prices the final like this: Spain 41.4% · draw 31.3% · Argentina 27.3% (implied probabilities). Our model gives Argentina 34.6%, seven points above the market (ten, if you look at the raw Poisson, without the draw adjustment).

It's not an error to fix: it's the thesis of the whole tournament. Tontin-BETe deliberately doesn't copy the market, and its divergences are measured against results, not against odds. May's theses, so far: the final, called and nailed. France in the final rounds, delivered: it reached the semis. Morocco champion, no, though it did make the final eight. The only open thesis closes on Sunday.

And yes: I'm Argentine, and the model's most pro-Argentina number is exactly the one I like most. That's why it's the one we audit the most, not the one we celebrate the most. It was published before the match, dated: the pitch delivers the verdict.

Fifty Thousand Simulations, a Different AI, the Same Numbers

Before publishing, we asked Claude (Fable, by Anthropic) to reimplement the engine from scratch: its own code, its own seed, plus the exact analytical calculation of the distribution. It ran its 50,000 finals and nailed the same numbers, with deviations of ±0.3 to ±0.5 points, exactly the statistical noise you'd expect between two independent Monte Carlo runs.

That doesn't validate the thesis (the pitch does that on Sunday): it validates the implementation. There's no bug. The numbers are what they are.

The original prediction: it all started on May 28, with the tournament yet to begin. The model played the entire World Cup 20,000 times, put Spain first, Argentina second, and called this final a coin flip. The full story is here 👉 Tontin-BETe Simulated the 2026 World Cup 20,000 Times: Here's What the Math Says.

---

Method note: proprietary Elo on a neutral field, a Poisson goals model (λ Spain 1.5 · Argentina 1.47), 50,000 simulations with extra time and penalties. The 1X2 carries the structural draw adjustment ("empate_modal"): reviewing 39 matches already played in the tournament, 8 of 13 errors from the raw model were draws it didn't see coming. The champion probability is robust to the adjustment; the goal distribution is the raw Poisson. This is a statistical simulation for educational purposes, not financial advice: the coin can land either way. By Esteban Aleart, Founder & Lead Engineer at PairProgramming.