Model Insights
What the model considers important when predicting AFL match outcomes.
Active Model
The current model used to generate all predictions. Tip rate is the percentage of matches where the model correctly picks the winner, measured via walk-forward testing against historical seasons. Margin MAE (Mean Absolute Error) is how far off the predicted margin is on average — lower is better. Features are the individual data points the model considers; only the most useful ones are kept after automated selection.
What Matters Most
Each feature belongs to a category like "Team Form" or "Venue". This shows how much each category contributes to the model's decisions overall. A higher percentage means the model relies more heavily on that type of information when picking winners and predicting margins.
ELO-style offensive and defensive ratings for 10 key stats (scoring, disposals, clearances, contested possessions, etc.), adjusted for opponent quality. A team that scores well against strong defences gets a higher rating than one padding stats against weak opponents.
Recent performance trends — rolling averages of scoring, disposals, clearances, and other stats over the last 3, 5, and 8 games, plus win streaks and consistency measures.
How settled the team is — average shared games between teammates, squad continuity from last week, and how long positional groups (defence, midfield, forward) have played together. More familiarity generally means better team connection.
Underlying team capabilities — offensive efficiency (points per inside 50), defensive solidity (points conceded, intercepts), contested ball, disposal efficiency, and pressure.
The effect of player availability — missing players by position (ruck, inside mid, key forward, key defender), star player absences, role criticality scores, and win impact analysis.
Ground-specific factors — each team's win rate at the venue, how many games they've played there, and travel distance to the ground.
Situational factors — ladder position, days of rest between matches, whether it's a rivalry derby, and if either team is coming off a bye round.
How each team historically performs in different conditions — win rates in rain, wind, and heat compared to their overall record, and how their playing style adapts.
Coach-related factors — tenure at the club, overall and recent win rates, venue-specific record, new coach bounce effect, and head-to-head coaching matchup history.
How much the team's selected squad has changed from the previous week — high turnover can disrupt team cohesion.
Log-scaled versions of skewed stats like form shock and momentum — compresses extreme values so a team on a massive streak doesn't dominate the model disproportionately.
Historical matchup data between the two teams — overall H2H win record and how each team's playing style interacts with the opponent's.
Individual player rolling stats — recent averages for disposals, goals, clearances, and tackles for each player, plus form trends.
Top 20 Most Important Features
The individual data points the model finds most predictive. A feature with high importance has a strong influence on whether the model picks one team over another. For example, if "Home weighted recent margin" ranks highly, it means how well the home team has been winning recently is a strong signal for the model.
| # | Feature | Category | Importance |
|---|---|---|---|
| 1 | Home points per game averaged over recent matches team_form__home_avg_score_8 | Team Form | 0.0336 |
| 2 | Average time current players have spent under this coach (home-away difference) cohesion__diff_coach_squad_avg_tenure | Team Cohesion | 0.0332 |
| 3 | diff team shared win pct cohesion__diff_team_shared_win_pct | Team Cohesion | 0.0319 |
| 4 | Home team quality avg disposals player_impact__home_team_quality_avg_disposals | Player Impact | 0.0319 |
| 5 | diff score launches avg5 advanced_stats__diff_score_launches_avg5 | Advanced Stats | 0.0314 |
| 6 | diff missing disposal share player_impact__diff_missing_disposal_share | Player Impact | 0.0306 |
| 7 | Tackle pressure rating adjusted for opponent quality (home-away difference) adjusted_form__diff_off_tackles_rating | Opponent-Adjusted Ratings | 0.0279 |
| 8 | diff team avg shared wins cohesion__diff_team_avg_shared_wins | Team Cohesion | 0.0271 |
| 9 | Home contest off wins avg5 advanced_stats__home_contest_off_wins_avg5 | Advanced Stats | 0.0251 |
| 10 | Home points per game averaged over recent matches team_form__home_avg_score_5 | Team Form | 0.0240 |
| 11 | Current consecutive wins (positive) or losses (negative) (home-away difference) team_form__diff_win_streak | Team Form | 0.0232 |
| 12 | Home average margin adjusted for the quality of opponents faced team_form__home_sos_weighted_margin_3 | Team Form | 0.0222 |
| 13 | Home points given up per game — lower is better team_strengths__home_defensive_points_conceded_avg | Team Strengths | 0.0209 |
| 14 | Forward entries per game (home-away difference) team_form__diff_avg_inside_50s_8 | Team Form | 0.0208 |
| 15 | Team's current position on the ladder as a percentage (home-away difference) contextual__diff_ladder_pct | Context | 0.0206 |
| 16 | Coach's overall win rate at this club (home-away difference) coaching__diff_coach_win_rate | Coaching | 0.0205 |
| 17 | Away how many players changed from the previous week's team — high turnover disrupts cohesion roster_change__away_roster_retention | Roster Change | 0.0204 |
| 18 | Away tackles inside fifty avg5 advanced_stats__away_tackles_inside_fifty_avg5 | Advanced Stats | 0.0203 |
| 19 | diff defensive scoring shots avg team_strengths__diff_defensive_scoring_shots_avg | Team Strengths | 0.0203 |
| 20 | Home team shared win pct cohesion__home_team_shared_win_pct | Team Cohesion | 0.0200 |
Least Important Features
These features made it through automated selection but contribute the least to predictions. They might be redundant with stronger features, or the pattern they capture just isn't very predictive. For example, a weather feature ranking low means conditions don't meaningfully help the model pick winners in most matches.
| Feature | Category | Importance |
|---|---|---|
diff tackles inside fifty avg5 advanced_stats__diff_tackles_inside_fifty_avg5 | Advanced Stats | 0.0000 |
diff f50 ground ball gets avg5 advanced_stats__diff_f50_ground_ball_gets_avg5 | Advanced Stats | 0.0000 |
Log-scaled version of a skewed stat — compresses extreme values while keeping direction (home-away difference) transform__log_abs_diff_sos_weighted_margin_3 | Transformed Features | 0.0000 |
Defensive exits per game (home-away difference) team_strengths__diff_territory_rebound50_avg | Team Strengths | 0.0000 |
Home defensive rebound rating adjusted for opponent quality adjusted_form__home_off_rebound_50s_rating | Opponent-Adjusted Ratings | 0.0000 |
Away defensive rebound rating adjusted for opponent quality adjusted_form__away_off_rebound_50s_rating | Opponent-Adjusted Ratings | 0.0000 |
Clearance rating adjusted for opponent quality (home-away difference) adjusted_form__diff_off_clearances_rating | Opponent-Adjusted Ratings | 0.0000 |
Away def clearances rating adjusted_form__away_def_clearances_rating | Opponent-Adjusted Ratings | 0.0000 |
Home metres gained avg5 advanced_stats__home_metres_gained_avg5 | Advanced Stats | 0.0000 |
diff shots at goal avg5 advanced_stats__diff_shots_at_goal_avg5 | Advanced Stats | 0.0000 |