Read this before trusting any model you build on it. The churn label in this dataset is generated from a rule, not observed from behaviour. A gradient-boosted model scores 0.997 AUC on held-out rows — that is the rule leaking, not skill. Two hard cliffs do almost all the work: churn steps from 6% to 53% between day 30 and day 31 of inactivity, and collapses from 61% to 0% at exactly 5 watch hours. Demographics carry no signal at all (age, gender, region, device and genre together score AUC 0.48 — worse than a coin flip). Use this dashboard to recover the rule, not to forecast real subscribers.
Churn rate by inactivity and viewing. Drag the lines, or type thresholds.
Five flags reproduce the label. Tune them and watch the fit move.
Churn rate by attribute. Click any bar to filter the whole dashboard.
Churned against retained, on the variables that matter.
Monthly fees attached to churned subscribers in the current view.
Sort any column. Search matches id, region, tier, device, genre and payment.