Model Overview
The Footy Bets model is a regression-based system with ensemble methods that estimates match outcome probabilities, expected margins, and expected total scores for every AFL game. It is not a black box. This page explains, at a high level, how it works, what it uses, and where it falls short.
Input Variables
The model draws on over 50 variables across five categories:
- Possession metrics: Disposals, contested possessions, uncontested possessions, handball-to-kick ratio, and disposal efficiency. These measure how effectively each team moves the ball.
- Scoring metrics: Scoring efficiency (points per inside-50 entry), goal accuracy, set-shot conversion, and score from turnover rate. These measure how well teams convert territory into points.
- Defensive metrics: Intercept marks, spoils, defensive one-on-ones, rebound-50s, and opposition inside-50 rate. These measure how effectively teams prevent scoring opportunities.
- Venue and travel factors: Ground-specific home advantage (calibrated per venue), interstate travel adjustment, day/night game split, and roof/open-air classification.
- Contextual factors: Head-to-head history (last 6 meetings), days since last game (rest advantage), weather forecast (rain probability, wind speed, temperature), and team selection changes from the previous week.
Model Structure
The core engine is a gradient-boosted regression model trained on historical AFL data from 2018 to 2025. It outputs an expected margin for each game, which is then converted into win probabilities and implied odds. A secondary model estimates expected total score using a similar feature set with additional pace-of-play variables.
We use an ensemble approach: the final output is a weighted average of the gradient-boosted model, a regularised linear regression, and an Elo-based power rating. The ensemble reduces overfitting and provides more stable predictions than any single model component.
Refresh Cadence
The model is updated after every round of the AFL season. New data from the most recent round is incorporated within 24 hours of the final game. Pre-round predictions for the following week are published by Wednesday, with adjustments made for late team changes where possible.
Early in the season (Rounds 1–4), the model leans more heavily on pre-season ratings and historical data because the current-season sample is too small for reliable inference. By mid-season, current-season data dominates the weighting.
Limitations — What the Model Cannot Do
We believe transparency about limitations is as important as explaining strengths. The model has real blind spots:
- Early-season small samples: Rounds 1–3 produce the least reliable predictions. Teams are still settling into structures, new players are integrating, and the data is thin. We flag early-season predictions with lower confidence.
- Injury timing: The model adjusts for known team changes but cannot predict injuries that occur during warm-up or in-game. Late withdrawals of key players can invalidate the pre-game estimate.
- Coaching changes and tactical shifts: A mid-season coaching change or a major structural adjustment (new game style, different forward line setup) takes several weeks to register in the data. The model is reactive, not predictive, on tactical innovation.
- Motivation and intangibles: Finals pressure, rivalry intensity, player milestones, and team morale are real factors that the model does not attempt to quantify. We believe the market often overprices these factors, but we acknowledge they exist.
- Individual brilliance: A single player having an extraordinary game can override structural advantages. The model deals in probabilities, not certainties.
Our Philosophy
We do not claim to predict winners. We estimate probabilities and compare them to the market. When our estimated probability is meaningfully different from what the bookmaker's odds imply, we flag it as a potential value opportunity. Some of those bets will lose. Over a full season, the process should produce a positive return if the model is well-calibrated — and we track that rigorously on our track record page.