Customer Intelligence

What it is

Customer Intelligence estimates which customers are likely to leave, groups customers into segments, estimates their future value and produces a ranked list of who to call first.

The problem

A business with limited staff cannot call every customer. It needs to know whom to contact and which customers are worth the effort, with probabilities it can trust.

What I built

• A calibrated gradient-boosting churn model, with a logistic baseline to compare against.
• RFM segments (recency, frequency, monetary value) such as Champions and At Risk.
• A customer lifetime value estimate that is checked against later revenue.
• A ranked action list of the customers with the most revenue at risk.

How it’s checked

For churn I report PR-AUC and precision and recall at a fixed contact budget, with confidence intervals. For lifetime value, I predict a held-out later period and compare it with what happened.

What the results do and don’t tell you

The churn model is tested on a public telecom customer dataset (7,043 customers, 26.5% churned, a 1,409-customer test set):

• PR-AUC 0.841 (95% interval 0.807 to 0.871) and ROC-AUC 0.933.
• If you contact the top 20% of customers by risk, about 81% of them really are churners and you reach about 61% of all churners.

The lifetime value check uses simulated retail transactions (1,200 customers over 540 days), where predicted holdout revenue was within 1.2% of actual. That shows the method works on data built to behave this way. It is not evidence about a real shop. Telecom churn is also easier than churn in many other businesses, so expect lower numbers elsewhere.

My role

I built all of it: the models, the segmentation, the value check, the action list and the app.

Stack

Python, scikit-learn, gradient boosting, Streamlit.