What the WNBA model cannot measure

How many WNBA shots does it take before shot-making means anything?

About 216.4 attempts before WNBA shot-making is half reliable, and 630.3 attempts before it reaches 0.7. Shot selection settles far earlier, at 10 attempts for 0.5 and 14.1 attempts for 0.7. Selection is a property of where a player shoots from, which stabilises quickly; making is whether the ball went in, which does not.

How many attempts before a number means anything

Split-half reliability of WNBA shot selection and shot-making, by attempt count
AttemptsSelection reliabilityShot-making reliability
100.6630.086
200.7540.090
300.8040.123
500.8650.212
1000.9260.308
2000.9580.485
4000.9800.662
8000.9870.728

The observable ceiling on this population is 800 attempts, where selection reads 0.987 and making 0.728. These thresholds are a property of each league's own talent dispersion, not constants, which is why they differ across the 9 leagues tested.

Are corner threes really better, or is it just distance?

In the WNBA, corner threes convert at 37.30% against 34.00% above the break, a gap of 3.30 percentage points. A distance-only model already predicts 2.04 of that, leaving a residual of 1.26 points (SE 0.57), which is distinguishable from zero. The corner effect is mostly the shorter shot, and the model prices it already.

What accuracy costs as the data gets cheaper

Full NBA feed to hand-logged scoresheet costs 0.020 log loss and 0.024 AUC, and IMPROVES calibration (ECE 0.0122 -> 0.0025). Measured across 4 grades of feed on one fixed subsample of 250,000 shots; the full table is on the page.

What the model cannot see

No per-shot defender distance, touch time or pose. Predictive validity is untested and untestable with the data available, so no claim is made about forecasting development.

Over Expected is a basketball shot-value platform covering NBA, EuroLeague, EuroCup, Liga ACB, Lega Basket Serie A, Basketball Bundesliga, Greek Basket League, ABA Liga, WNBA.