Churn Risk Analyzer

A customer churn prediction and retention ROI model built on IBM Telco data. Predicts which customers are likely to leave, ranks them by risk, and calculates the net revenue value of a targeted retention campaign (i.e recommends the interventions worth paying for).

Headline result: Targeting 601 flagged customers is estimated to generate $63,024 in net retained revenue at 699% ROI.

Table of Contents

• Problem Framing
• What This Project Does
• Stakeholders
• Data
• Modelling approach
• Results
• Cost-benefit analysis
• Limitations
• Getting Started
• Tech Stack
• Author/Contributors
• License

Problem Framing

Most businesses only know a customer churned after it happened, that is, when it's too late. This project predicts churn 30–60 days in advance using behavioural signals, so retention teams can act before revenue walks out the door.

Churn prediction is a binary classification problem. The target variable is whether a customer churned (1) or stayed (0). The dataset is imbalanced — 26.5% churned, 73.5% stayed — which shapes all modelling decisions.

Recall was chosen as the primary evaluation metric over accuracy. A missed churner costs ~$65/month in lost revenue; a false alarm costs ~$15 in an unnecessary retention offer. The asymmetry justifies optimising for recall.

What This Project Does

Layer	Output
Prediction	Churn probability score per customer
Prioritisation	Ranked list of at-risk customers
Cost-benefit	Which interventions are actually profitable to run
Business brief	Revenue at risk, churn segments, recommended actions

Stakeholders

Role	How They Benefit
Marketing team	Concentrate spend on highest-value interventions.
Customer success	Shift from reactive to proactive outreach.
Business analysts	Tie model performance to measurable revenue saved.

Data

• IBM Telco Customer Churn dataset via Kaggle
• Size: 7,043 rows × 21 columns
• Target: Churn (Yes/No became 1/0)
• Cleaning: TotalCharges col. converted from string to float (11 nulls filled with 0), customerID dropped, binary columns mapped to 1/0, 10 multi-category columns one-hot encoded.
  • Final shape: 7,043 rows × 30 columns.

Feature	Description
tenure	Months as a customer
contract_type	Month-to-month, one year, two year
monthly_charges	Monthly bill amount
support_calls	Number of support interactions
services_used	Add-ons and products subscribed to
churn	Target label (Yes / No)

Modelling approach

Three models were trained and evaluated:

Model	Rationale
Logistic Regression	Interpretable baseline; performs well on linear relationships
Random Forest	Ensemble method; tests non-linear pattern detection
XGBoost	Sequential boosting; strong on tabular data with imbalance

Evaluation included classification report, confusion matrix, ROC-AUC curve, and precision-recall curve. Threshold tuning was applied across 0.10–0.55 to find the optimal business tradeoff between precision and recall.

Results

Model	Threshold	Recall	Precision	F1	AUC
Logistic Regression	0.30	0.757	0.523	0.619	0.84
Random Forest	0.25	0.807	0.513	0.627	0.82
XGBoost	0.25	0.826	0.514	0.630	0.84

Selected model: XGBoost at threshold 0.25.

XGBoost achieved the highest recall (0.826), best F1 (0.630), and highest average precision (0.66) on the precision-recall curve. Threshold tuning reduced missed churners from 199 to 65, which is a 67% reduction.

XGBoost tuned confusion matrix:

Top churn drivers (XGBoost feature importance):

1. Internet service: fiber optic
2. Contract type: two year (protective)
3. Contract type: one year (protective)
4. Tenure
5. Payment method: electronic check

Feature importance chart:

Cost-benefit analysis

Assumptions: $64.76 avg monthly revenue, $15 offer cost, 30% retention success rate, 12-month customer lifetime.

Tier	Customers	Net Value	ROI
High risk (≥ 0.70)	109	$19,320	1,060%
Medium risk (0.50–0.70)	213	$21,936	806%
Low risk (0.25–0.50)	279	$21,768	491%
Total	601	$63,024	699%

ROI by Tier Chart:

Limitations

• Dataset represents a single US telecom provider so generalisation to other markets or business models requires validation.
• 7,043 rows is a small dataset; performance may shift with more data
• No call log, NPS, or usage frequency data available, thus, richer behavioural signals would likely improve recall further
• Retention success rate of 30% is an industry benchmark, not a measured conversion rate for this business.
• Cost-benefit figures are estimates; actual campaign ROI depends on real offer costs and customer lifetime data.

Getting Started

pip install pandas numpy matplotlib seaborn scikit-learn xgboost imbalanced-learn plotly joblib

Run notebooks in order:

1. notebooks/data_exploration.ipynb: data overview and visualisation.
2. notebooks/cleaning.ipynb: preprocessing and feature encoding.
3. notebooks/modelling.ipynb: model training, evaluation, threshold tuning.
4. notebooks/cost_benefit.ipynb: ROI analysis and recommendations export.

Outputs generated:

• outputs/confusion_matrix_lr.png
• outputs/roc_curve.png
• outputs/precision_recall_curve.png
• outputs/feature_importance.png
• outputs/threshold_tuning.png
• outputs/roi_by_tier.png
• outputs/retention_recommendations.csv
• outputs/xgb_churn_model.pkl

Tech Stack

Models used: Logistic Regression · Random Forest · XGBoost

Evaluation: Confusion Matrix · Precision / Recall · ROC-AUC

Author/Contributors

_marzafiee

License

• Distributed under the MIT License.
• See LICENSE for more information.

---

For full details, see the report in the repo titled Data Consultancy Report_TelcoX.pdf