How AI Football Prediction Models Work — Sezi's Methodology

Can AI predict football matches? The short answer is yes — but it's not magic. It's mathematics. Here's exactly how Sezi's model works, layer by layer.

The 4-Layer Architecture

Layer 1: Dixon-Coles Poisson Model

Football goals follow a Poisson distribution. The Dixon-Coles model estimates each team's attacking strength and defensive weakness, then calculates the probability of every possible scoreline. It also corrects for the "draw inflation" effect in low-scoring matches.

Layer 2: ELO Rating System

Adapted from chess, ELO ratings measure team strength on a continuous scale. After every match, ratings update based on: result, goal difference, opponent strength, and home advantage. For World Cup 2026, Sezi uses World Football Elo with Bayesian shrinkage for teams with limited data.

Layer 3: XGBoost Ensemble

This is where machine learning shines. XGBoost takes features from Poisson and ELO, plus: recent form (last 5-10 matches xG/xGA), injury impact, rest days, head-to-head history, and tournament context. It captures non-linear interactions that simple models miss.

Layer 4: Isotonic Calibration

Raw model outputs tend to be overconfident ("95% sure" but only right 70% of the time). Isotonic regression maps predicted probabilities to observed frequencies. When Sezi says "62% confidence," it actually means 62%.

The Trust-Contract

Every Sezi prediction passes 47 automated checks. Data source verification, model assumption documentation, accuracy monitoring, and a 535-file copy-lint that ensures zero forbidden betting language.

Continuous Learning

Sezi updates after every match. Our CLV benchmark (Closing Line Value vs. Pinnacle closing odds) shows +0.0222 CLV, 55% positive — meaning Sezi systematically generates predictions that beat the market.

_Sezi — Understand the game with numbers. sezi.io_