Model Documentation
How The Model Works
A full technical breakdown of how SportSphere HQ generates AFL disposal predictions — every factor, every weight, every filter. No black boxes.
Every prediction is a weighted combination of six data inputs. Each factor is given a sensitivity weight that determines how strongly it pulls the prediction away from the player's baseline average.
2026 Season Average
65% blend
The player's current-season average disposals. By Round 6, this dominates the season blend at 65% weighting — we trust recent data more than historical as sample size builds.
2025 Season Average
35% blend
Full prior-season average. Provides a stable baseline, especially early in the season when 2026 data is thin. The blend ratio shifts each round (see Section 5).
Opponent Adjustment
25% sensitivity
How many disposals this opponent typically concedes to this position vs league average. An opponent conceding +12% more than average to DEFs lifts the predicted line. Sensitivity is capped at 0.30 to prevent over-weighting noisy early-season data.
TOG-Adjusted Rate
40% sensitivity
Disposal rate per 1% time-on-ground, normalised to an 82% TOG baseline, then scaled to the player's expected TOG for this game. A player playing more minutes should be expected to accumulate more disposals.
CBA / Form Trajectory
10% sensitivity
Centre bounce attendance effect for midfielders. A player getting more CBAs than the positional average gets a small uplift. Dampened to 0.10 sensitivity to prevent over-inflation on players with volatile CBA counts.
Play Style Factor
60% sensitivity
TRANS (transition) players get a 1.05 multiplier in dry conditions — they generate disposals through run and carry. STOP (stoppage) players get 0.90 — their disposals come from contested ball, which is more volatile. HYBRID players sit at 1.00.
Raw edge (model prediction minus bookie line) alone is not sufficient to identify a bet worth taking. A 4-disposal edge means something very different on a player with a standard deviation of 3.5 vs one with a standard deviation of 8.0.
The Edge/Vol ratio divides edge by the player's estimated disposal standard deviation to produce a signal-to-noise measure. This is the core innovation of the model.
Worked Example
LOW volatility player
Bookie line: 27.5
Model prediction: 31.5
Edge: +4.0
Std Dev: 3.5
Edge/Vol: 1.14 → HIGH CONVICTION
HIGH volatility player
Bookie line: 18.5
Model prediction: 22.5
Edge: +4.0
Std Dev: 8.0
Edge/Vol: 0.50 → borderline — filter with caution
HIGH CONVICTION
≥ 0.90
Strong statistical signal. Edge is large relative to the player's typical variance. Highest-confidence picks.
BET
0.50 – 0.89
Meaningful edge that clears the statistical noise threshold. Worth including in analysis.
SKIP
< 0.50
Edge exists but may be within normal variance. The model shows a lean but not a high-confidence edge.
HIGH CONVICTION picks (E/V ≥ 0.90) achieved 60.7% accuracy across 84 picks. Filtered picks (E/V ≥ 0.50) achieved 60% across 100 picks. Rounds 3–6, 2026 season.
03
Position-specific thresholds
Bookmakers price different positions differently. MID markets are tighter (more efficient) than FWD markets, which are the widest. The model uses separate STRONG thresholds per position to reflect this.
MID
≥ 3.0 disposal edge
Most liquid market. Bookmakers price MIDs well. Standard threshold.
DEF
≥ 3.0 disposal edge
Similar liquidity to MID. DEF markets have been profitable historically.
FWD
≥ 4.5 disposal edge
EXCLUDED from bet filter. Model accuracy on FWDs is 38% — below break-even. FWDs are listed for analysis only.
RUCK
≥ 5.0 disposal edge
Very conservative. Disposals are a poor proxy for RUCK performance — contest stats drive their role, not disposals.
Premium (line ≥ 27)
+2.0 bonus edge required
Players with high bookie lines are priced more efficiently. An extra 2.0 disposal edge is required for STRONG on these players.
04
Conditions and multipliers
External game conditions affect disposal counts. The model applies multipliers on top of the base prediction.
Dry conditions
×1.00
Baseline. No adjustment.
Wet conditions
×0.95 base
Rain reduces total disposals across the game. Additionally: TRANS players (×0.95) are more penalised as run-and-carry is harder. STOP players (×1.04) benefit — congestion and contested ball increases.
Roof venue (Marvel Stadium)
×1.02
Indoor venues produce marginally higher disposal counts. The roof removes weather risk and the surface tends to produce faster, higher-disposal games.
2026 rule changes — DEF
×1.03
Rule changes boosted intercept marking (+11% league-wide) disproportionately benefitting defenders who read the play.
2026 rule changes — RUCK
×0.90
Ruck contest rule changes reduced centre bounce frequency by 16%, directly cutting RUCK disposal counts.
05
Dynamic season blending
Early in the season, the 2026 data is thin (2–3 games). The model blends 2025 full-season averages with 2026 data using round-dependent weights. As the season progresses and 2026 data becomes more reliable, it takes over.
Rds 1–3
60% 2025 / 40% 2026
Early season — 2026 data is thin. 2025 full-season averages dominate.
Rds 4–7
20% 2025 / 80% 2026
2026 data becomes reliable. Heavy weighting shifts to current season.
Rds 8–11
20% 2025 / 80% 2026
Stable mid-season blend. 2025 retained as a small anchor.
Rds 12–17
10% 2025 / 90% 2026
2026 data now authoritative. 2025 used only for stability.
Rd 18+
5% 2025 / 95% 2026
Finals and late rounds — almost entirely current season data.
06
What the model doesn't do
Transparency means being clear about limitations, not just strengths.
Real-time injury news
The model does not scrape team selection or last-minute injury updates. Always check official team lists before acting on any analysis. A key teammate out can dramatically change a player's disposal count.
FWD accuracy is 38%
FWDs are listed for analysis but are excluded from the bet filter. The model does not explain why FWD predictions underperform — it's a known limitation, not a solved problem.
Early-season opponent factors are noisy
Opponent concession data is unreliable before Round 6 (small sample size). The model halves opponent sensitivity (0.30 vs 0.60) to compensate. Treat early-season DvP data with caution.
This is a decision-support tool
The model identifies statistical edges. It does not account for every variable — weather changes, game-day motivations, selection surprises. It is a tool, not a guarantee.