Every SportSphere prediction runs on the same six-factor framework. The structure is identical across sports; the weights, thresholds, and conditions are tuned independently against each league's historical data.
Every prediction is a weighted combination of six data inputs: season averages, recent-form windows, opponent strength, role and minutes, conditions, and matchup. Each factor has a sensitivity weight that pulls the prediction away from the player's baseline.
Weights are not chosen by feel - they're grid-search optimised on the prior season for each sport before anything goes live. AFL weights differ from WNBA weights differ from NFL weights, but the structure is the same: six factors, transparent sensitivities, no black boxes.
Raw edge - model prediction minus bookmaker line - is not enough. A 4-disposal edge on a player with a standard deviation of 3.5 means something very different from the same edge on a player with a standard deviation of 8.0.
The Edge/Vol (E/V) ratio divides edge by the player's estimated standard deviation. It's a signal-to-noise score, and it's how tiers are assigned:
Players with thin recent data - rookies, returns from injury, limited minutes - have their edges automatically damped to reflect reduced confidence. Below the per-sport minimum sample threshold, nothing is published.
Where uncertainty is meaningful but not disqualifying, picks carry a visible "limited sample" flag. We surface the uncertainty rather than hiding it.
Every pick is logged before games start. Results are graded against official data. The track record is built publicly, round by round - no cherry-picking, no quiet edits after the fact.
We don't publish every pick the model generates. Each sport gets an archive analysis that identifies the segments where the model has statistically meaningful edge - and the segments where it does not. Only the segments with measurable edge appear on the public feed.
For AFL, the May 2026 analysis of 1,074 historical decided picks produced the current filter rule:
Sample sizes per segment vary. The filter is monitored against rolling 4-round performance - if drift appears, the rule is re-cut against the updated archive rather than left static.
Our AFL model averages roughly 16 picks per matchup once you include every prediction across both teams. Those 16 picks are not 16 independent edges - they're correlated expressions of the same underlying matchup signal. If the model is right that one team will dominate inside, it's right about several of their midfielders at once.
Archive analysis on 1,074 decided picks measured this directly:
The model's genuine edge is identifying which player in each matchup is the most mispriced - not pricing fifteen players in the same matchup better than the bookmaker. We publish accordingly: up to three picks per game, ranked by edge size.
Honest CI note: sample sizes per segment vary; the lower bound on top-1 (55.8%) sits comfortably above the ~52.4% break-even line. We monitor and adjust if rolling 4-round performance drifts.
We surface our picks as a unified portfolio across every sport we cover. Each pick is a position - line is cost basis, edge is alpha, volatility is risk, P&L is realised return. We measure performance the way an institutional analyst would, with attribution, exposure, drawdown, and risk-adjusted returns at the surface level - not buried in reports. /portfolio is the live institutional view across AFL and WNBA (with NFL, NBA, and NBL added as each sport goes live).
The structure above applies everywhere. The specific weights, position thresholds, and conditions multipliers are sport-specific. Read the deep dive for each sport:
Transparency means being clear about limitations, not just strengths.