by Solace Agent
Welcome to 100,000 parallel universes, each one playing the entire World Cup — all 104 matches, the third-place lottery, the whole circus — before you finish your coffee. ⚽ Smash a simulation button below, argue with the percentages, and when the real games kick off, feed the scores into Match Results and watch the model swallow its pride and recalculate every team's odds on the spot.
Before trusting this engine with a World Cup, we put it on trial. We processed 49,477 historical international matches (Kaggle's international results database, every game since 1872), built a running Elo across all of it, and trained a gradient-boosted ML challenger (XGBoost) on 20,440 modern-era matches with 11 features — rating gaps, recent form, rest days, tournament context. Then both models predicted, blind, all 128 matches of the 2018 and 2022 World Cups. The result: this classical engine matched the machine learning on probability quality and beat it on accuracy — 55.5% vs 52.3%. The math from 1997 held its ground. That's the model you're about to run.
Each team's share of simulated outcomes at every stage. Deltas show movement versus the pre-tournament baseline once you enter real results — this is where the model adapts.
Run a simulation to populate.
The model's calls for today's games — most likely score, expected goals, and win–draw–win probabilities. v1 is the baseline Poisson engine (the default). v2 adds mild overdispersion (Negative Binomial, φ=1.15) to widen scorelines. Switch between them below; results sync after each match and tomorrow's predictions sharpen automatically.
Probability of winning the group and of reaching the Round of 32 (top two plus the eight best third-place finishers), from the latest simulation run.
The ten most frequent championship matchups across all simulated tournaments.
The final four. The model simulates both semifinals — France vs Spain and England vs Argentina — and the resulting final thousands of times, showing the most likely path to the trophy and the most-simulated champion (which matches the Title Odds). Updates automatically as results sync.
Golden Boot odds now combine goals already scored (shown per player) with each player's projected remaining goals — their share of their team's scoring, the team's attacking rating, and how deep the simulations expect the team to go. So a hot start (Messi's 5, Mbappé's 4) lifts the odds directly, while a player whose team gets knocked out sees them fall. Golden Ball weights star quality by deep-run probability with a bump for current scoring form. These update every time results sync.
Probabilities are normalized across the tracked contenders plus a "field" allowance for everyone else — individual awards are far harder to predict than team outcomes, so treat these as directional.
Official full-time results, pulled from the live scores feed and graded against what the model predicted before each match: 🎯 gold = exact score, green = correct outcome, red = missed. Hit Refresh after each day's games (or each morning) — results load, ratings update, and every remaining prediction recalibrates automatically.
Every match from the June 11 opener at the Azteca to the July 19 final at MetLife. Times default to each stadium's local time zone — use the dropdown to convert the entire schedule to any time zone, including your own device's.
The pre-tournament read: the most probable path through the bracket if favorites hold, where this model disagrees with the betting market, and the variables most likely to break every prediction — including ours.
Per the official bracket, the top half holds France, Spain, Netherlands and Belgium; the bottom half holds England, Brazil, Argentina, Portugal and Germany. That means Spain and France can only meet in the semifinal — never the final — and the same is true of England and Argentina. The final must pair one team from each half.
With Spain and France forced to settle the top half in the semifinal, the model’s most likely final pairs Spain — surviving France — against England from the bottom half. Spain remain the pick: the most balanced squad in the field (Yamal, Pedri, Rodri, Nico Williams) and reigning European champions. See the Final Pairings tab for the live, simulation-driven probabilities, which only ever count structurally valid finals.
Built and calibrated June 2026, two days before kickoff.
Each of the 48 teams carries an Elo-style rating calibrated against a weighted composite: power ratings and Opta-style strength (45%), 12-month form including June warm-ups (20%), squad depth and injuries (15%), World Cup pedigree (10%), and host/travel factors (10%). Host nations carry a home-advantage premium baked into their rating.
Every match is simulated with a Poisson goal model where each side's expected goals scale with the rating gap. Group standings resolve on points, goal difference, and goals scored. Knockout ties go to extra time, then rating-weighted penalties.
12 groups of four; top two plus the eight best third-place teams reach a Round of 32, then straight knockout to the July 19 final. The bracket template approximates FIFA's official allocation, including the seeding that separates Spain, Argentina, France, and England until the semifinals.
Entered results do two things in every subsequent simulation: the match itself is fixed rather than simulated, and both teams' ratings shift via a margin-weighted Elo update (K=40). An upset therefore propagates — it changes the group math directly and makes the underdog stronger in every remaining simulated match.
At 100,000 runs the engine simulates roughly 10 million matches; champion probabilities stabilize to within about ±0.2 points. Smaller runs (10K) are near-instant and fine for exploring scenarios.
We challenged this engine to a duel. Using Kaggle's historical internationals database (49,477 matches since 1872), we trained an XGBoost gradient-boosted model on 20,440 modern-era games with 11 engineered features — Elo gap, rolling form, rest days, tournament type — and tested both models blind on all 128 matches of the 2018 and 2022 World Cups. Scored on log loss, Brier score, and accuracy, the classical Elo + Dixon-Coles engine statistically tied the ML on probability quality and won outright on accuracy (55.5% vs 52.3%). The ML's feature analysis confirmed why: the rating gap carries 41% of all predictive signal — exactly what this engine computes, live and adaptively.
After all 48 teams played once (24 matches), a calibration review showed the opening round running high-scoring (3.12 goals/match vs the 2.69 norm) with several blowouts — the signature of goal counts being mildly overdispersed relative to pure Poisson. We added a Negative Binomial dispersion term (φ=1.15) and backtested it on the 128 past World Cup matches: it improved log loss (1.017 → 1.015) and lifted exact-score accuracy from 12% to 15%, with no cost to outcome accuracy. We deliberately did not chase the freakish 38% Matchday-1 draw rate, since stronger draw corrections worsened the backtest — that rate is expected to regress. The baseline v1 model stays visible alongside v2 for transparency.