How the model works
Every prediction on this site comes from the same transparent method. Here is exactly how it is built, factor by factor.
How Our Predictions Work
Every prediction on this site is generated by a multi-factor statistical model combined with Monte Carlo simulation. No human picks, no gut feelings - just data and mathematics applied consistently across all 104 matches.
The model combines 5 independent factors, each capturing a different dimension of team strength. When live data isn't available for a factor, its weight is automatically redistributed to the factors that do have data - so predictions are never artificially compressed toward 50/50.
The 5-Factor Model
Each match prediction combines these five factors with base weights. The final probability for Team A is a weighted average of all factor probabilities:
The foundation of our model. FIFA's official ELO system tracks cumulative team strength across all competitive matches. Higher ELO = historically stronger team. The probability that Team A beats Team B is:
The 600-point scale is FIFA's official parameter. A 200-point ELO gap gives the stronger team roughly a 76% win probability. A 50-point gap (common among top teams) gives about 58%.
Analyzes each team's last 10 competitive matches with two adjustments:
- ›Recency decay (0.9^i): The most recent match has full weight, each older match is worth 10% less. Match 1 = 1.0x, match 5 = 0.66x, match 10 = 0.39x.
- ›Opponent strength: Results against stronger teams count more. Multiplied by opponent's ELO / 1600 (baseline). Beating a 2000-ELO team counts 1.25x; losing to a 1200-ELO team counts 0.75x.
Results are scored as: Win = 1.0, Draw = 0.5, Loss = 0.0. The final score ranges from 0 (all losses) to 100 (all wins against strong opponents).
A composite score built from three tactical metrics, each normalized to 0-100:
- ›xG Differential (50% of metrics): Expected goals for minus expected goals against per 90 minutes. Measures overall attacking vs defensive quality. Range: -1.5 to +2.0.
- ›Defensive Record (30% of metrics): Goals conceded per 90 minutes (inverted - fewer = better). Range: 0.3 (elite) to 2.5 (poor).
- ›Pressing Intensity (20% of metrics): PPDA (passes allowed per defensive action) - lower means more aggressive pressing. Range: 5 (very intense) to 15 (passive).
USA, Mexico, and Canada are co-hosts. This factor is the sole home advantage mechanism in the model. When a host nation plays in their own country, they receive a score boost (base 50 = neutral):
- ›+20 home score for playing on home soil (crowd support, familiarity, logistics)
- ›+8 additional for Mexico City matches (2,200m altitude significantly stresses opponents)
- ›+5 crowd bonus for venues with 70,000+ capacity (intensity of large home crowd)
These boosts are deliberately modest. World Cup crowds are heavily mixed with tourists and neutral fans, so home advantage is weaker than in domestic competitions. The ELO factor uses raw ratings only - home advantage is not double-counted.
Historical matchup record between the two specific teams. Requires at least 2 prior meetings to activate. The win rate differential is converted to a factor capped at ±5% to prevent ancient history from overriding current team quality.
Example: If Team A has won 70% of meetings (diff = +0.40), the shift is +0.04 - a modest 4% boost. This ensures H2H is a tiebreaker, not a dominant factor.
Dynamic Weight Redistribution
Not all data is available for every team. When form, metrics, or H2H data is missing, the model doesn't just use 50/50 as a placeholder - it redistributesthat factor's weight proportionally to the factors that do have data.
Example: If H2H data is unavailable (no prior meetings):
Formula: adjusted_weight = base_weight / sum_of_available_weights
If only ELO and home advantage are available (no API data at all), ELO gets 80% weight and home gets 20%. The model always uses whatever data it has optimally.
Monte Carlo Simulation
Group stage standings are determined by running 10,000 full tournament simulations. This produces stable probability distributions that account for the randomness of football.
How Each Simulation Works
- 1.Compute win probabilities for all 6 group matches using the 5-factor model
- 2.For each match, roll a random number to determine outcome (win/draw/loss) based on probabilities
- 3.Generate realistic scorelines using Poisson-distributed goal counts (avg 1.8 for winner, 0.7 for loser)
- 4.Rank teams by FIFA tiebreaker rules: points → goal difference → goals for → ELO rating
- 5.Top 2 advance automatically. 3rd place has a 67% chance (8 of 12 third-place teams advance)
Draw Rate
Group stage matches use a 26% base draw rate (historical World Cup average). The draw probability adjusts dynamically based on match closeness - evenly-matched teams draw more often than lopsided matchups. Intra-confederation matches draw at ~28%, inter-confederation at ~22%.
Knockout Matches
In the knockout rounds, draws aren't possible (extra time and penalties decide). The draw probability is redistributed proportionally between the two teams:
Goal Simulation
For group tiebreakers, we need goal difference - not just win/loss outcomes. Goal counts follow a Poisson distribution, the standard statistical model for rare events (goals in football):
Draw: Both teams score from Poisson(λ=1.0) - avg ~1 goal each
Winner: Scores from Poisson(λ=1.8) - avg ~2 goals
Loser: Scores from Poisson(λ=0.7) - avg ~1 goal
Constraint: Winner always scores at least 1 more than loser
Data Sources
FIFA/Coca-Cola World Ranking
Official FIFA rankings and ELO ratings. Updated after every international window. Source: inside.fifa.com
API-Football (api-sports.io)
Live form data (last 10 matches per team), head-to-head records, and live match scores. Data refreshed daily via automated cron jobs.
Team Metrics (xG, PPDA)
Expected goals, defensive records, and pressing intensity metrics. Sourced from public football analytics databases.
Venue & Schedule Data
Official FIFA World Cup 2026 venue assignments, match schedule, and capacity data for all 16 stadiums across USA, Mexico, and Canada.
Free CSV Fallback
When the API is unavailable, form data falls back to a free CSV source. The model automatically adjusts weights when data is missing.
Data Freshness
› Form data: refreshed every 24 hours via Upstash Redis cache
› H2H records: cached for 7 days (historical data changes slowly)
› Live scores: polled every 30 seconds during active matches
› Page predictions: revalidated every 1 hour via ISR (Incremental Static Regeneration)
Model Limitations
Every prediction model has blind spots. Ours is transparent about what it cannot capture:
Key Player Injuries & Suspensions
A squad missing its star striker is statistically the same team. Real impact: significant.
Manager Tactical Matchups
Some managers historically dominate others through tactical preparation. Not modeled.
Squad Chemistry & Morale
Locker room dynamics, tournament fatigue, and psychological momentum are invisible to statistics.
Set Piece Specialization
Teams with elite set piece routines overperform their underlying metrics. Not captured in our xG data.
Tournament Experience
Teams with deep World Cup pedigree may handle knockout pressure better. ELO partially captures this but not fully.
Football's Inherent Randomness
A deflected shot, a referee decision, a moment of individual brilliance. A 55% probability means the other team wins 45% of the time.
Bottom line: These are the most likely outcomes, not certainties. The model tells you which team wins more often if this match were played 10,000 times. In any single match, anything can happen - that's what makes the World Cup magical.
Transparency
Every number on this site is computed from the formulas described above. No manual overrides, no editorial picks, no sponsor influence. The same model runs identically for every team - from the defending champions to the smallest qualifying nation. If the data says an upset is likely, that's what you'll see.