Clustering 100,000 bank customers into six risk profiles, plus a short list of outliers that need a manual review.
A single credit score hides the different ways a customer becomes risky. Could 100,000 banking customers be grouped into clear risk profiles a bank can act on?
The synthetic dataset had real-world problems built in: duplicates, impossible values and missing income and credit scores. Those had to be fixed before any model could be trusted.
I chose nine variables covering repayment capacity, debt and repayment behaviour, and left out demographics because they aren’t direct risk signals and raise fairness concerns.
In R, I cleaned and scaled each variable, tuned K-means with elbow and silhouette analysis, and picked six clusters where the statistics and the business meaning agreed. A DBSCAN pass then flagged customers who fit no cluster at all.
Six clear segments, from debt-free high earners to a distressed group of 4.1% of customers with a debt-to-income ratio of 11.57. DBSCAN isolated another 3.3% (3,303 people) whose debt-to-income ratio was 32 times everyone else’s.
The bank gets segments to plan collections around, plus a short list for manual review today. Code and report on GitHub.