← Blog
kickwizaiaflmachine-learning

KickWiz: Building an AFL Prediction Engine with Five AI Models

How we built an Australian Rules football intelligence system using an ensemble of five AI models — and what we learned about prediction, uncertainty, and the beautiful chaos of AFL.

Jose Martins 17 April 2025

The Problem with AFL Predictions

Australian Rules Football is one of the hardest sports to predict. Eighteen players per side, a circular ball, an oval ground, and a scoring system that rewards both goals and behinds — the variables are nearly infinite. Most prediction models fail because they try to reduce the game to a single number.

We took a different approach. Instead of one model trying to predict everything, we built five specialised models that each “see” the game differently — then we let them collaborate.

Five Models, One Prediction

The KickWiz prediction framework is an ensemble — a term from machine learning that means combining the outputs of multiple models to produce a better result than any single model alone.

Each of our five models focuses on a specific dimension of the game:

  1. Form Model — tracks recent performance trends at the club and individual player level, weighted by recency and opponent quality.

  2. Matchup Model — analyses historical head-to-head data between clubs, including home/away records, margin distributions, and quarter-by-quarter patterns.

  3. Player Impact Model — assesses the influence of key players on team performance. Injuries, suspensions, and form dips are factored in dynamically.

  4. Conditions Model — integrates weather data, ground size, and surface type. A wet MCG plays very differently from a dry Optus Stadium.

  5. Momentum Model — captures psychological and contextual factors: travel schedules, consecutive wins/losses, finals pressure, and inter-state travel fatigue.

How the Ensemble Works

Each model outputs a probability distribution — not just “Team A wins” but “Team A wins 64% of the time, with a likely margin of 12–22 points.” These distributions are weighted and combined using a meta-learner that has been trained on historical prediction accuracy per model per scenario.

The result is a confidence-aware prediction: you see not just the pick, but how confident the system is in it, and which models agree or disagree. Disagreement between models is itself a signal — it usually means the game is genuinely hard to predict.

Player and Season Predictions

Beyond match outcomes, KickWiz supports:

  • Player performance predictions: disposals, goals, marks, tackles — with confidence intervals
  • Season simulations: run Monte Carlo simulations across the remaining draw to forecast ladder positions and finals odds
  • Brownlow Medal tracking: accumulating vote predictions as the season progresses

What We Learned

Building KickWiz taught us something important: good AI predictions aren’t just about accuracy — they’re about calibrated accuracy. A model that says “70% confidence” should be right about 70% of the time. Getting calibration right across five different models, on a sport as chaotic as AFL, is the real engineering challenge.

We also learned that users trust predictions more when they understand them. Showing the model breakdown — “our Form Model strongly favours Richmond, but our Matchup Model is uncertain” — builds more trust than a black-box percentage.

KickWiz is available now. Download it, pick your teams, and let the models do the work.