1 · What the model does
Every World Cup 2026 match probability on Onside is the output of an ensemble that blends multiple independent indicators of team strength, calibrated against the long-run historical baseline for international football. The output is a win / draw / loss probability vector for every fixture — and, by running the full bracket forward many thousand times, a per-team probability of reaching each knockout round and lifting the trophy.
2 · The principles we held to
Calibrated, not optimistic. Where the football itself doesn't give you a clear favourite, the model shouldn't pretend otherwise — draw rates and tight-fixture splits hew to the historical baseline rather than the consensus narrative. Transparent on receipts, private on internals. We publish every prediction the model has ever made and grade each one as the tournament plays out; what we don't publish is the exact weighting recipe, because that's the work. Stable under live results. As real outcomes land, completed matches lock in as certainties and propagate through the simulator — the bracket-wide probabilities sharpen naturally without us moving the goalposts.
3 · The simulator
Tournament-wide probabilities — who reaches the round of 32, the quarter-finals, the final, the trophy — come from a Monte Carlo simulator that runs the full bracket forward repeatedly, sampling each remaining fixture from the model and respecting completed results. Aggregated across many thousand runs, the per-team counts give the round-by-round probabilities you see on the simulator and winner pages. The number of runs is well past convergence; the meaningful uncertainty sits in the per-match model, not the sim.
4 · Live result integration
Live results come through a 4-second-timeout client with Redis caching, so a slow upstream never hangs a page. Every completed group fixture is treated as a certain prior in the simulator; remaining group fixtures and every knockout are sampled from the model. By the semi-final round only two matches remain to sample, so the champion probabilities you see become very precise. The pages refresh on an hourly cadence to balance freshness against compute.
5 · What the model deliberately does not do
It is not a tipster. We do not model injury news after the squad-submission deadline, individual key-player absence, in-tournament tactical change, weather, late substitutions, or stadium effects beyond the basic host bonus the historical record supports. The model is calibrated for the long run; on any single match, variance is high. Treat the percentages as probabilities, not predictions.
6 · How to verify our numbers
The Model Record page is the receipt — every match the model called, the live grading, biggest correct call, biggest miss, and a calibration grid that shows whether 60%-favourite calls actually win 60% of the time. A full post-tournament reconciliation, including Brier score by round, will be published within 14 days of the final.