Tenure Prediction
The exact prediction target is not defined in the founder brief (customer vs employee tenure; months-to-churn vs probability of staying beyond N months). This project offers two framings; the founder's chosen target will replace them.
The scenario and data are provided, but tasks that depend on the undefined business rule are marked and must not be assumed.
Business scenario
Orion Telecom wants to anticipate how long new subscribers will stay, to set acquisition spend per customer.
Problem statement
Framing A: predict the probability that a subscriber stays at least 12 months. Framing B: estimate expected tenure in months, treating active subscribers as censored.
Learning objectives
- Define the target without leakage
- Handle censored observations
- Evaluate and explain the model
Instructions
Features must be known at signup (plan, contract, payment method, region, age, acquisition channel). For Framing A, include only subscribers who signed up at least 12 months before the data end. Do not use usage or tickets after signup.
Tasks
- Build the modelling table with signup-time features.
- Framing A: logistic regression and gradient boosting; AUC and calibration on a time-based split.
- Framing B: Kaplan-Meier curves by contract type (lifelines or manual).
- Interpret drivers.
- Recommend acquisition spend logic.
Deliverables
- Notebook
- Model card
- Recommendation
Evaluation rubric
| Criterion | Weight | Excellent | Proficient | Developing | Insufficient |
|---|---|---|---|---|---|
| Target & leakage control | 25% | Target correctly defined, censoring handled, no leakage. | Correct and complete with minor gaps in rigour or presentation. | Partly correct; important steps missing, unvalidated or unclear. | Missing, incorrect, or not supported by the data. |
| Modelling & evaluation | 25% | Appropriate models with time-based evaluation. | Correct and complete with minor gaps in rigour or presentation. | Partly correct; important steps missing, unvalidated or unclear. | Missing, incorrect, or not supported by the data. |
| Interpretation | 20% | Drivers explained in business terms. | Correct and complete with minor gaps in rigour or presentation. | Partly correct; important steps missing, unvalidated or unclear. | Missing, incorrect, or not supported by the data. |
| Interpretation & recommendations | 15% | Actionable recommendations. | Correct and complete with minor gaps in rigour or presentation. | Partly correct; important steps missing, unvalidated or unclear. | Missing, incorrect, or not supported by the data. |
| Presentation | 15% | Concise, well-labelled outputs a manager can act on without explanation. | Correct and complete with minor gaps in rigour or presentation. | Partly correct; important steps missing, unvalidated or unclear. | Missing, incorrect, or not supported by the data. |