Churn Rate Analysis – Stephen Mulingwa

The Problem

Telecom operators lose significant revenue when customers churn. Without a way to predict who is likely to leave, retention efforts are spread thinly across the entire base and often applied too late to make a difference.

The business needed a model that could flag high-risk customers in advance and support targeted, cost-effective retention campaigns — focusing spend where it actually matters.

What I Built

I took customer and usage data — tenure, usage patterns, complaints, and contract type — and prepared it for modelling: handling missing values, encoding categories, and engineering features that capture behaviour over time.

I built and compared several classifiers with a focus on neural networks, tuning hyperparameters and applying cross-validation to avoid overfitting. The priority was strong recall on churners — catching most at-risk customers — without contacting the entire base unnecessarily.

Data Prep Missing values, encoding, feature engineering

Model Training Neural networks + classifier comparison

Tuning Hyperparameter search + cross-validation

Evaluation Precision, Recall, AUC-ROC

Explainability Feature importance & error analysis

Deliverable Trained model + scoring script + report

Precision

Recall

AUC-ROC

Cross-Validation

Outcome

The business gained a data-driven way to rank customers by churn risk and design targeted retention interventions — such as offers or direct outreach — for the highest-risk segment, improving the efficiency of retention spend.

Neural network classifier trained on real telecom customer data
High-recall model catches most at-risk customers before they leave
Feature importance analysis explains key drivers of churn
Scoring script to rank new customers by risk level
Short report with model findings and retention recommendations
More efficient retention campaigns — targeted, not blanket