How PropScout builds NBA & WNBA player projections

Every daily player projection ships with a confidence tier, a Mean Absolute Error (MAE) score, and a bias-adjusted range. Here's exactly how those numbers are built.

Data sources

Each projection blends six categories of raw input, refreshed before every slate:

• Recent-game form — rolling last-5, last-10, last-15 and last-30 game averages per player, per stat.
• Home / away splits — split performance for the specific home-or-away context of tonight's game.
• Matchup defensive ranks — opponent's defensive rating vs. the player's position, filtered per stat (points allowed to guards, rebounds allowed to bigs, etc.).
• Pace & rest — projected possessions for the game plus days-of-rest for the player.
• Injury & minutes projections — official injury reports and projected minutes-per-game with usage rate adjustments when a starter is out.
• Prior-season baseline — anchor for players with thin recent samples (rookies, returning-from-injury, role changes).

Weighting

Recent-form windows are weighted so the freshest signal dominates but the model isn't whipsawed by a single hot or cold game. The exact weights are re-fit weekly against the previous 30 days of results to minimize MAE per stat.

Matchup adjustments layer on top: if a team allows the 3rd-most points-per-game to opposing point guards, that scales up point-guard scoring projections in that matchup. Pace scales possessions; a fast-pace opponent lifts every counting-stat projection.

Range & bias adjustment

Every projection publishes as three numbers:

• Point estimate — the model's expected value.
• MAE — Mean Absolute Error over the last 30 days for this player-stat combination.
• Bias — signed average error (positive = model has been under-projecting, negative = over-projecting).

The visible Range column is the point estimate corrected for bias, plus/minus the MAE — the true band the actual result has historically landed inside.

Confidence tier scoring

Every projection is labeled with a stat-aware confidence tier. Tiers are derived from MAE / projection line — the relative error — with thresholds tuned per stat so ±0.5 blocks and ±0.5 points aren't scored the same way.

• Elite Very tight historical accuracy relative to the line size. Rare.
• Sharp Reliably accurate. The model's strongest reads land here.
• Fair Reasonable projection with average error. Useful, not authoritative.
• Loose Low sample or high recent error. Treat as directional only.

Combo props

Combo projections (PRA, PR, PA, RA, S+B) are built from the underlying single-stat projections, but graded on their own historical accuracy. Combos require at least 12 graded games before they'll be labeled Sharp or Elite, and are capped to Loose if the underlying single-stat bias is directional (|bias| / MAE ≥ 0.45) so small-sample combo noise never earns a Sharp label.

Isotonic calibration

Once per week, the model's raw output is passed through an isotonic calibration step against the last 30 days of actual results. This corrects for any systematic over- or under-projection that has crept in and keeps the published bias close to zero across the full slate.

What we don't do

• We don't pull sportsbook lines into the projection. Every number is generated independently, then compared to the market inside the app.
• We don't hide accuracy. Every projection ships with its own MAE and bias — you can see exactly how it's performed.
• We don't "tout" picks. PropScout publishes projections and confidence tiers, not betting advice.

See tonight's projections →