From classical statistical theory to multi-signal fusion—every layer is academically validated and production-ready.
01
Academic-Grade Probability Models Dixon-Coles Poisson Engine
Built on the 1997 Dixon & Coles framework published in the Journal of the Royal Statistical Society, forming a complete probabilistic inference system across 15 method modules. Same methodology lineage as Opta, FiveThirtyEight, and peers.
Poisson DistributionELO RatingsLog-Linear DecompositionDixon-Coles ρ CorrectionBayesian Prior ConstraintsMonte Carlo Integration
A single model has blind spots. The system weight-fuses four signals—statistical probabilities, market-implied odds, ML enhancement, and head-to-head history—for accuracy above any single path alone.
Full Evaluation Loop Quantifiable, Verifiable, Evolving
Every prediction logs full intermediate variables (four-path probabilities, weights, sample counts) and supports backtesting across any season. Built-in grid search continuously optimizes fusion weights and calibration parameters—no manual accuracy tallying.
Probability Distribution · Bivariate Poisson Process
Home and away goals follow independent Poisson distributions G ~ Poisson(λ), providing the mathematical foundation for all market derivations.
M
Stochastic Simulation · Monte Carlo Integration
Computes complex conditional probability scenarios (e.g. handicap lines spanning multiple spreads) with guaranteed numerical precision.
E
Strength Assessment · Dynamic ELO Ratings
Team ratings update automatically after each match. World Cup K-factors: group stage 30 / knockouts 50 / final 60.
A
Attack Modeling · Log-Linear Decomposition
Expected goals λ decomposed into attack α, defense β, and home advantage γ—with independent parameter estimates per team.
T
Form Modeling · Exponential Time Decay
Recent matches carry higher weight w = 0.5^(t/45), with a 45-day half-life to capture shifting team form.
S
Venue Modeling · Home/Away Dual Weighting
Home teams blend 65% home + 35% away; away teams blend 65% away + 35% home—to neutralize venue bias.
ρ
Low-Score Correction · Dixon-Coles ρ Factor
Corrects correlation bias in four low-score outcomes (0-0/1-0/0-1/1-1), with significant empirical gains in draw prediction accuracy.
H
Head-to-Head · Independent H2H Signal Path
Win-draw-loss frequency from past meetings serves as a fourth fusion signal (minimum 2 valid matches), capturing tactical matchup dynamics.
X
Expected Goals · xG Over Actual Goals
Uses expected goals (xG) instead of actual goals to estimate λ, reducing single-match noise in parameter estimation.
I
Squad Injuries · Position-Differentiated Weights
Injuries to forwards, midfielders, defenders, and goalkeepers apply differentiated impact coefficients to λ, with squad value dynamically adjusted.
$
Market Probability · Public Market Implied Odds
Public market data normalized into probabilities as the second fusion signal—capturing team momentum, user expectations, and unstructured match intelligence.
B
Bayesian Update · Posterior Parameter Estimation
In small-sample scenarios, priors constrain posteriors (λ ∈ [0.2, 4.0]), preventing MLE overfitting on sparse data.
D
Form Decay · Exponential Time Weighting
wₜ = 0.5^(t/45)—a 45-day half-life keeps the model highly responsive to recent form changes.
F
Travel Fatigue · Geographic Distance Factor
Long-haul travel applies a negative correction to away-team λ—especially critical in cross-continental World Cup schedules.
C
Set Pieces · Dedicated Independent Modeling
Independent estimates of corner and free-kick conversion rates, completing scoring paths beyond open-play goal models.
Market ProbabilityPublic market data · probability normalization
ML Enhancementλ-delta + logit model · nonlinear capture
Head-to-Head StatsH2H frequency · independent validation
1X2 · Handicap · Correct Score · Total Goals · Half/Full Time
All five markets share the same λ parameters, ensuring internal logical consistency
Technical Trust
Why It'sWorth Trusting
Technical strength shows not only in model accuracy, but in explainable processes, verifiable results, and a system that evolves sustainably.
TRUST 01
Explainable Probability Output
Alongside 1X2 and score probabilities, every prediction retains four-signal sources and weight allocation. Any result can be traced back—how much came from Poisson derivation, market signals, or head-to-head adjustments. Full-chain transparency.
TRUST 02
Post-Match Review & Continuous Learning
Every completed match is automatically added to the review sample. Built-in grid search finds optimal fusion weights and calibration parameters across full-season data. The engine evolves with the tournament—not a fixed, unchanging model.
TRUST 03
Strict Data Quality Controls
Predictions follow strict temporal cutoffs: only pre-kickoff historical data, excluding the match itself and unfinished fixtures. Up to 20 samples per dimension to prevent data leakage inflating backtest metrics. Anomalies are identified and filtered before entering the model.
Engine Pipeline
From Data toPrediction Output—The Full Chain
A layered architecture where every stage can be independently validated and iterated—stable, transparent, traceable predictions.
01Data CollectionMulti-source real-time signal ingestion