Business Problem

For subscription-based telco operators, churn is a silent profit killer. Acquiring a new customer is typically several times more expensive than keeping an existing one, yet many organizations still react to churn only after customers have already left. The goal of this project was to build a data-driven churn-prediction engine that identifies at-risk customers early enough to intervene with targeted retention actions, and to do so in a way that is transparent and intuitive for business users.

Concretely, we focused on three questions:

• Who is likely to churn in the coming months?

• Which factors drive this risk the most?

• How can we translate model outputs into clear, actionable customer segments and visual stories for decision makers?Staying Focused on the Mission

Data & Methods

We worked with the well-known Telco Customer Churn dataset, which contains several thousand residential customers of a fictitious telecom provider.

The end-to-end pipeline has four main stages:

1. Data cleaning and feature engineering

2. Preprocessing and encoding

3. Model training and evaluation
   We train a Random Forest classifier on a stratified train/test split. The model is configured to handle class imbalance (class_weight=”balanced”) and to capture non-linear interactions between services, pricing and contract conditions.
   Performance is evaluated using:

   • ROC curve and AUC – showing how well the model separates churners from non-churners across all probability thresholds.

   • Confusion matrix and classification report – giving a more operational view of correctly captured churners vs false alarms.

4. Post-modeling analytics for business users
   Raw probabilities are transformed into:

   • Feature importance rankings (global drivers of churn)

   • Risk segments (low / medium / high churn risk buckets)

   • Tenure-based “risk tunnel” visualization to tell the story of churn over the customer lifecycle.

Results

Taken together, these results provide a full story for commercial and retention teams:

• The model is accurate enough to distinguish low- from high-risk customers, as shown by the ROC curve and the sharp gradient in churn rates across risk buckets.

• We can clearly explain why customers are at risk: short tenure, flexible contracts and higher charges emerge as key drivers, supported by other service and payment factors.

• The risk segments translate abstract probabilities into operational groups that can be targeted with differentiated offers, call-center scripts or digital journeys.

• The tenure-based risk tunnel delivers a simple but powerful message: invest heavily in the first months of the relationship, because that is where churn risk – and therefore the upside of good onboarding and service design – is highest.

The same pipeline can be re-trained on fresh data or adapted to other subscription businesses (media, utilities, SaaS), giving organizations a repeatable way to monitor churn risk and design smarter, data-driven retention strategies.

A new visualization, the so called tenure-based risk-tunnel was also introduced. The resulting chart shows a clear pattern

• Churn peaks in the very early months of the relationship, when customers are still evaluating the service and comparing alternatives.

• As tenure increases, the average churn probability declines steadily, and the band gradually narrows – long-standing customers are both more loyal and more homogeneous in their behavior.