How the AFL model works
We predict each player's disposals, compare that to the bookie line, and publish only the bets the data actually rewards. No black boxes.
See the multi-sport framework—
Strong win rate
recorded · graded after each round
—
ROI
$1k flat stake
6
Inputs per prediction
weighted & blended
What goes into a prediction
Every line is a weighted blend of six inputs around the player's baseline.
2026 formblendedCurrent-season disposal average - weighted higher as games build up.
2025 baselineblendedPrior full season, for a stable anchor early in the year.
Opponent±How many disposals this opponent concedes to the position vs league average.
Time on ground×More minutes means more disposals - scaled to expected TOG.
Role / CBA×Centre-bounce attendance lift for midfielders in the engine room.
Style + conditions×Run-and-carry vs contested, plus wet-weather and roof-venue adjustments.
Confidence tiers
Edge alone isn't enough. We divide edge by the player's volatility (Edge/Vol) to score signal vs noise.
High Conviction
E/V ≥ 1.4× threshold
Edge is large relative to the player's usual variance - at least 1.4 times the position threshold below.
BET
E/V ≥ threshold
Real edge that clears the position threshold: 0.50 for MID/DEF, 0.55 for RUCK - raised by 0.10 for high-volatility players (σ > 6) and 0.25 for extreme volatility (σ > 8).
SKIP
below threshold
A lean, but inside normal variance - not published. At most two picks per team-position combination make the board.
What we publish (and thin weeks)
The model flags 100+ edges a round. These rules cut it to the picks you see.
OVERs: High Conviction only
BET and LEAN OVERs lose money at standard vig in our data, so only HC OVERs are published.
UNDERs: stricter cut
UNDERs work at every tier but pass a tougher line, edge and tough-defence filter - the UnderTaker.
Some weeks are quiet on purpose
A handful of picks means the model is declining bad bets - not that nothing happened.
By position
MID / DEFpublishedMost efficient markets and our most profitable. Standard edge thresholds.
FWDanalysis onlyModel accuracy is ~38% - shown for context, never bet.
RUCKconservativeDisposals are a weak proxy for ruck impact, so the bar is high.
What it doesn't do
Where the edge comes from matters - so does where it stops.
Late injury & team newsWe don't scrape selections - always check the official team list before betting.
ForwardsFWD disposal prediction is a known weak spot, so FWDs are excluded from picks.
Early-season noiseOpponent data is thin before Round 6, so its influence is dialled back.
It's decision supportThe model finds statistical edges - it isn't a guarantee on any single bet.