We Had Spain #1 Before the World Cup Started
Spain are World Cup champions. Ferran Torres in the 106th minute, 1-0 over Argentina in the final, Spain's second title.
On June 10, before a single ball was kicked, we published our model's title odds. Spain was #1.
That post is still up. Here's the table we published, next to the best forecasts in the sport:
Team | Impulse | Goldman Sachs | Opta | Markets (Kalshi/Polymarket) |
|---|---|---|---|---|
Spain | 18.8% | 25.7% | 16.1% | ~17% |
France | 15.5% | 18.9% | 13.0% | ~16% |
England | 10.0% | 5.0% | 11.2% | ~11% |
Brazil | 8.9% | 7.6% | 6.6% | ~8% |
Argentina | 8.0% | 14.3% | 10.4% | ~9% |
Our top five: Spain, France, England, Brazil, Argentina. Four of the five made the semifinals. The final was our #1 against our #5.
We stayed on Spain when the field left
Everyone had Spain high in June. The difference is what happened next.
After the group stage, the betting markets stampeded to France. By July 8, Kalshi and Polymarket both had France around 33% to win it all, the runaway favorite, with billions in volume behind them. We posted our board the same day: Spain on top at 20.2%, France at 18.9%, and our take in plain words: France was priced on momentum, not math.
Spain beat France 2-0 in the semifinal. Then they beat Argentina to win the whole thing. From the first published table to the trophy, our model's answer to "who wins the 2026 World Cup" was Spain.
The part that actually matters
Nobody at Impulse built this model.
We gave our AI agent three public datasets and a prompt describing what we wanted: a tournament forecast for all 48 teams, updating live as results came in. The agent did everything else. It built a strength rating over 49,000 international matches back to 1872, blended in current squad values and FIFA points, and simulated the full 48-team tournament 40,000 times. After every result, it re-ran the simulations and updated every team's odds, all tournament long.
No ML team. No modeling code from us. A prompt and a dataset, and the output went toe-to-toe with Goldman Sachs, Opta, and billions of dollars of prediction-market money for six weeks, in public, with every number on the record before the games were played.
That's the story. Not that a model picked Spain. That an agent you can prompt built a model that picked Spain, and held its own against the best forecasting operations in the world.
Same workflow, any prediction
NBA playoffs. NHL playoffs. The Fed. Now the World Cup. Same thing every time: public data, a prompt, an agent that ships the model.
The World Cup was the hardest one. Low-scoring sport, draws, single-elimination knockouts, a 48-team format nobody had ever run. We said in June we had no idea whether we'd look smart or foolish in a month. Now we know.
If you have data and something worth predicting, churn, demand, fraud, conversion, you can do exactly what we did. Upload the data, tell the agent what you want, done.
See the final World Cup forecast → worldcup2026.impulselabs.ai
Try the agent on your own data → impulselabs.ai