DEMO / 07 · Demo ready
Customer retention lab
Leakage-controlled churn-risk demonstration on synthetic B2B accounts: point-in-time features, cohort retention, calibration and capacity-limited ranking.
What decision is supported
Which accounts a retention team with limited capacity should contact this period, and whether the model's risk scores can be read as probabilities.
Who uses it
A customer success lead at a fictional B2B software vendor with a fixed number of outreach slots, and an analyst who must show that no feature was computed after the outcome it predicts.
What data enters
Authored synthetic account histories with usage, billing and support events and a churn outcome per period. No real customer data is used.
What is computed
Point-in-time feature snapshots that only use events before the cut-off, cohort retention curves, a scored model with a calibration check, and a ranking cut at the team's capacity with the expected number of churners in the contacted set.
What action is suggested
A contact list of the given size with each account's calibrated risk, and a note on how many churners the list is expected to contain versus a random list of the same size.
What evidence supports it
Lift over random and calibration error are computed on the synthetic fixture and labelled demo.
No metric is claimed for this project. Anything computed by the repository on its synthetic fixture carries the label demo.
What fails or is uncertain
- Churn is defined by the generator; real churn definitions vary by contract and are often ambiguous.
- The model is a baseline classifier; the emphasis is the leakage test and the capacity-limited evaluation.
- No causal claim is made about whether contacting an account changes its outcome.