Streaming subscriptions and churn

What it is good for

• Build a cohort retention table and compare churn by plan and billing cycle.
• Fit a logistic regression for churn and interpret odds ratios.
• Explain why tenure for active customers is censored and why early cohorts have had more time to churn.
• Normalise support tickets by tenure before comparing customers.

Methods: proportions, cohort analysis, chi-square test, logistic regression, Kaplan–Meier (introduction), rates per exposure.

Data quality notes

• Censoring: customers with churned = 0 are still active, and their tenure is only a lower bound on their lifetime.
• Earlier cohorts were followed longer, so raw churn proportions are not comparable across cohorts without accounting for time.
• support_tickets grows with tenure; use tickets per month for comparisons.
• No missing values.

Provenance and licence

Signup month uniform over 2023 with a January and September bump. Plan: Basic 45%, Standard 38%, Premium 17%; annual billing 25% (15% fee discount). Each customer has a latent support-need rate (gamma, mean 0.3 tickets/month). Each month the customer cancels with probability = plan base hazard (Basic 5%, Standard 3.5%, Premium 2.5%) × 0.4 if annual × exp(1.5·(support rate − 0.3)), and is otherwise retained until 31 Dec 2024. Tickets ~ Poisson(rate × months active).

Synthetic data · CC0-1.0 (synthetic data generated by Statistics Quest) · generated deterministically (seed 20241023). It does not describe real people or organisations; any resemblance is coincidental.