data-portfolio

Churn Prediction — MTN Nigeria Telecom

Dataset note. This Kaggle snapshot (974 MTN Nigeria customer rows) is approximately random-uniform in the covariates we tested — see the sister notebook 08-customer-survival-mtn-tenure/README.md for details. Logistic AUC and RF F1 collapse; XGBoost AUC = 0.71 reflects non-linear interactions in high-cardinality one-hots, not substantively interpretable signal. This notebook is a teaching demo of binary-churn classification end-to-end — logistic regression, random forest, XGBoost, ROC / PR / calibration diagnostics, retention-queue ranking. For a real benchmark see the IBM Watson Telco Customer Churn dataset (7,043 customers).

Business question

Which customers are most likely to churn in the next quarter, and which features drive their risk? A binary classifier ranks customers by churn probability so retention campaigns can target the highest-risk segment.

(Companion to project 08, which models the same data as a survival problem — same data, two analytical lenses, both useful.)

Data

Source: MTN Nigeria Customer Churn (Kaggle) — 974 customers, 17 features (age, state, plan, satisfaction, tenure, revenue, data usage, churn label).

Run python download_data.py.

EDA targets

Class balance (~29% churn)
Univariate hazard signals: low-satisfaction, low-tenure, plan type
State-level differences

Modeling

Family	Model
Linear	Logistic regression with L2 penalty (interpretable baseline)
Trees	Random Forest, Gradient Boosting (XGBoost)

Validation

80/20 stratified split (test-set base rate = 29%).
Metrics: ROC-AUC, PR-AUC, F1, balanced accuracy, Precision@k and Lift@k for the retention-queue framing.
Calibration plot with Wilson 95% CIs per bin so the ~19-obs-per-bin uncertainty is visible rather than smoothed over.

Headline numbers (test):

Model	ROC-AUC	PR-AUC	P@50	Lift@50	P@100	Lift@100
XGBoost	0.71	0.51	0.52	1.78×	0.40	1.37×
Random Forest	0.59	0.36	0.34	1.16×	0.38	1.30×
LogReg	0.48	0.27	0.20	0.68×	0.30	1.03×

Only XGBoost’s top-50 queue carries meaningful lift (1.78× over random). LogReg’s top-50 is actually worse than random, a telltale that linear scoring is picking up anti-signal on this (near-random) data. See the dataset note at the top for context.

Deployment sketch

Retention queue ranked by XGBoost score; audit Lift@k monthly against the holdout to catch score drift.
Not deployable on the current dataset — the dataset note at the top applies. On a real churn benchmark (IBM Watson Telco) the same code produces interpretable features and defensible top-k lift.

Business outcome

Retention spend allocated where predicted lift is highest; top-k is the metric that matches the campaign-targeting use case.
Companion to project 08 (survival analysis on the same data) — two analytical lenses, both constrained by the same data provenance issue.

This site is open source. Improve this page.