scorer model v1: P(player scores)

← lab · AUC 0.737 (real signal) · ran 7/4/2026

What this is: A one-off study — its numbers below are exactly what the run reported.

note	walk-forward chronological folds; players need 180+ prior minutes
brier	0.0708
auc mean	0.7368
base rate	0.0886
n player matches	25,745

calibration

n	bin	p pred	p actual
7,223	0.00-0.07	0.0358	0.0359
3,240	0.07-0.13	0.0951	0.0985
1,306	0.13-0.20	0.1631	0.1631
686	0.20-0.27	0.2313	0.2172
304	0.27-0.34	0.295	0.2566
86	0.34-0.40	0.362	0.2791
20	0.40-0.47	0.4258	0.55
8	0.47-0.54	0.5009	0.375

Reading the columnswhat each number actually means

AUC	predictability: 0.50 = coin flip, ~0.70 = ceiling for sports
Importance	how much the model leans on this factor (permutation importance)
Direction	sign of the raw correlation with the outcome
Survives all eras	effect points the same way in every historical era

Spec · the reproducible recipe

{
  "kind": "scorer_model",
  "features": [
    "goals_p90_l10",
    "shots_p90_l10",
    "shots_on_p90_l10",
    "start_rate_l5",
    "minutes_avg_l5",
    "is_forward",
    "is_mid",
    "is_home",
    "team_attack",
    "opp_defence"
  ]
}