Predict who's about to churn, quantify what it costs, and act before they leave.
An end-to-end churn-analytics project on 7,043 telecom customers: a machine-learning model scores every customer's likelihood of leaving, and an interactive Power BI dashboard turns those scores into a prioritized, money-aware retention plan.
For any subscription business, keeping a customer is far cheaper than winning a new one — but retention budgets are limited. This project answers three questions a commercial team actually asks:
- Who is likely to churn? — a probability score for every customer.
- How much is it worth? — the revenue at risk behind each customer and segment.
- Where should we act first? — the high-value, high-risk segment to target.
| Metric | Value |
|---|---|
| Customers analysed | 7,043 |
| Overall churn rate | 26.5% |
| Annual revenue at risk | ~$1.68M |
| High-risk customers flagged | 1,567 |
| Churn model — ROC AUC | 0.84 |
- Contract type is the #1 driver. Month-to-month customers churn ~15× more than two-year contract holders (≈43% vs ≈3%) — and they carry the bulk of the revenue at risk.
- The first year is the danger zone. Churn peaks in the 0–12 month tenure band and falls steadily as customers mature.
- Fiber-optic and electronic-check customers churn most — clear, targetable operational signals.
- Revenue at risk concentrates in month-to-month, short-tenure customers — the exact segment a retention budget should hit first.
data/WA_Fn-UseC_-Telco-Customer-Churn.csv raw data (7,043 customers)
|
scripts/build_dashboard.py clean -> train churn model ->
| score every customer ->
| derive risk tier + revenue-at-risk
v
data/churn_scored.csv Power BI-ready table (one row / customer)
dashboard/churn_dashboard.png rendered dashboard preview
- Churn model — Logistic Regression (standardized features), ROC AUC ≈ 0.84, producing a churn probability for every customer.
- Risk tier — High (p ≥ 0.50) · Medium (0.30–0.50) · Low (< 0.30).
- Revenue at risk —
MonthlyCharges × 12 × churn_probability: the expected annual revenue lost from a customer, weighted by how likely they are to leave.
The interactive dashboard is built in Power BI Desktop (free) from
data/churn_scored.csv: KPI cards, churn-rate breakdowns by contract / tenure /
payment method, a risk-tier split, and revenue-at-risk views — all cross-filtered by
slicers.
- Build guide:
powerbi/BUILD_GUIDE.md— step by step, ~45 min. - DAX measures:
powerbi/DAX_measures.md— copy-paste ready.
pip install -r requirements.txt
python scripts/build_dashboard.pyRegenerates data/churn_scored.csv and the dashboard preview, and prints the KPIs
and model AUC.
Customer-Retention-Intelligence-System/
├── data/
│ ├── WA_Fn-UseC_-Telco-Customer-Churn.csv raw dataset
│ └── churn_scored.csv scored, Power BI-ready output
├── scripts/
│ └── build_dashboard.py data pipeline + model + preview
├── notebooks/
│ └── Customer_Retention_Intelligence_System.ipynb exploratory analysis
├── models/
│ └── churn_model.pkl trained model
├── powerbi/
│ ├── BUILD_GUIDE.md how to build the dashboard
│ └── DAX_measures.md DAX for every measure
├── dashboard/
│ └── churn_dashboard.png rendered preview
└── requirements.txt
Python (pandas, NumPy, scikit-learn, matplotlib) for the data pipeline and model · Power BI + DAX for the interactive dashboard.
IBM Telco Customer Churn — 7,043 customers with contract, tenure, charges, services, demographics, and a churn label.
Built for subscription businesses — telecom, SaaS, streaming, e-commerce — to identify at-risk customers, quantify revenue exposure, and target retention (discounts, service fixes, contract upgrades) where it pays off most.