▶ Simulate fills every unplayed game using the chosen Model and plays out the bracket — it keeps the 28 already-played real results and any scores you set yourself. The Upset dial (0 → 1) controls chaos: 0 = the favourite always advances (deterministic); higher lets underdogs through, so each Simulate gives a fresh bracket. ↺ Reset returns to today's real results.
You can also drive it by hand: for any unplayed game click a team to set the winner or type the goals; use ▲▼ to hand-order a group; click a team's name in the standings to track its path (up to 5); and in the bracket, click the side you think wins.
① Groups — results & standings
② Track teams' paths
③ Knockout bracket — click winners to advance
④ Data & provenance — every number the models use · complete match record from the martj42 dataset
Forecast — Monte-Carlo title odds: the tournament played 10,000 times
0.00
Each run samples every unplayed match from the selected Model's win probabilities at their natural upset rate — keeping the 28 played results and any scores you set — and aggregates how often each team reaches each round. Switch models in the Tournament tab to compare forecasts.
Simulation accuracy — each played WC2026 game vs what every model predicted (pre-tournament data only)
A fair back-test: H2H & form are cut off the day before the World Cup, so each model predicts the real results using only what it knew beforehand. A model predicts a win when it favours a side >60%; at 40–60% it's predicting a draw. ✓ the predicted outcome (that win, or a draw) was right · ✗ wrong. So a 72%-for-Brazil call that finishes 1–1 is a ✗ — only a 40–60% call is credited for a draw.