Live Accuracy · WC 2026 · 54%

Model90 · Prediction Engine

The World's Most
Advanced Football
Prediction Engine

Bivariate Poisson statistics, Elo strength ratings, expected goals (xG), six contextual intelligence signals, and a meta-model stacking layer — unified into a single prediction engine refreshing every 30 minutes.

Live Performance Data

The Numbers Don't Lie

54%
WC 2026 Accuracy
Group Stage
0.193
Brier Score
38% better than random
0
Matches Analyzed
4 leagues · Training corpus
0+
Teams Modeled
Individual attack & defense params

System Architecture

8-Stage Prediction Pipeline

Raw match data flows through a chain of interconnected model layers. Click any stage to explore its implementation details.

Technical Deep Dive

Inside the Model

Core Engine

Dixon-Coles

Bivariate Poisson Distribution

Team-specific attack (α) and defense (β) parameters are optimised via weighted maximum likelihood. Poisson probability computed for every scoreline from 0-0 to 8-8. Low-score ρ correction improves calibration.

λ_H = exp(α_H − β_A + γ)
λ_A = exp(α_A − β_H)
P(k) = e^(−λ) × λ^k / k!
αTeam attack strength
βDefense weakness
γHome advantage coefficient
ρLow-score correction factor
Temporally weighted: decay λ = 0.005 — recent matches dominate parameter estimation.

Elo Edge

Dynamic Team Strength

FIFA World Rankings Convention

Every team carries a dynamic strength score updated after each result. K=40 competitive, K=20 friendly. Goal-difference multiplier follows FIFA convention. +100 home Elo offset.

E_A = 1 / (1 + 10^((R_B − R_A) / 400))
R′_A = R_A + K × G × (S_A − E_A)
P(draw) = 0.26 × exp(−|ΔElo| × 0.0015)
K40 (tournament) / 20 (friendly)
GGoal-difference multiplier (FIFA)
E_AExpected score (0–1)
+100Home advantage Elo offset
Elo → 3-way: empirical draw model converts differential to home/draw/away probabilities.

xG Model

Expected Goals Layer

Shot-quality parallel Poisson

Shots-on-target converted to xG at 0.35 goals-per-shot-on-target. A parallel Poisson model trained on xG is blended at 30% weight with the goals model.

xG = shots_on_target × 0.35
λ_final = 0.70 × λ_goals + 0.30 × λ_xG
0.35Shot-on-target → goal conversion
0.70Goals model blend weight
0.30xG model blend weight
xG model activates only when artifact exists — early-stage fallback to goals-only Poisson.

Live Tournament Validation

54% Accuracy on FIFA World Cup 2026

The engine called 54% of the first 24 Group Stage matches correctly, achieving a Brier score of 0.193 — 38% better than a random-guess baseline of 0.222. Validated on live tournament data in real time.

54%
WC 2026 Accuracy
0.193
Brier Score
0.222
Random Baseline
−38%
Outperformance
See Today's Predictions →Explore Results

Model90 predictions are statistical probability estimates — not guarantees of outcome. Where passion meets probability.