How Alphr's AFL Predictions Work

The 4-Model XGBoost Ensemble

Alphr's AFL prediction system is not a single model, it is a four-model ensemble where each component is optimised for a different task. The outputs are combined into the head-to-head, margin, and over/under tips published on the AFL hub each round.

1. AFL Win Predictor, a binary classifier returning the home-team win probability. This drives every head-to-head tip.
2. AFL Margin Predictor, a regression model estimating the winning margin in points. Feeds the margin & line picks.
3. AFL Dual-Score Predictor, regression models forecasting combined scoring, used for total-points (over/under) tips.
4. Edge Filter, only a tip with ≥7% positive edge over the bookmaker's implied probability is published. This single rule explains why Alphr's tips outperform a naive "tip every match" approach.

159 AFL Features per Match

Each match is described by 159 engineered features grouped into seven families:

• Team ELO ratings, custom AFL ELO with margin-adjusted K and season regression (top feature, ~14% importance).
• Positional player ELO, dual-track lineup ratings aggregated across the match-day 22 (midfielders, forwards, defenders, rucks).
• Rolling form, last 5/10 game wins, points for/against, average margin.
• Venue & travel, venue home-win rate, ground size, interstate travel, rest days.
• Weather, temperature, rain, wind speed and conditions at venue.
• Head-to-head history, recent results between the two clubs.
• Market signals, live Sportsbet head-to-head, line and totals odds at lock-in time.

Training, Validation & Anti-Leakage

Alphr uses a strict temporal train / validation / test split: training on the 2012–2023 seasons (12 years), validation on 2024 for hyperparameter tuning and early stopping, and a blind out-of-sample test on the full 2025 season. Every feature uses anti-leakage window functions, only data available before the first bounce is allowed. That means the historical accuracy you see is a real estimate of forward-looking skill, not a backtested overfit.

Why the Edge Filter Matters

Predicting an AFL winner is easier than beating the bookmaker. Sportsbet's opening line is already a sharp prediction. To make money from AFL tips you need an edge, your model probability must exceed the implied probability built into the odds. Alphr requires ≥7% positive edge before publishing a tip. See the AFL best bets page for the highest-edge selections each round.

2025 Performance Snapshot

On the blind 2025 AFL season Alphr's head-to-head model hit 81.9% (177 / 216, ROC-AUC 0.8808), the margin-direction model hit 76.9% (166 / 216), and the over/under model hit 60.2% (130 / 216), a combined 73.0% strike rate across 648 published tips. Read the full breakdown on the AFL accuracy page.