PC Lab 7 — Bank Customers Segmentation

The setting

A traditional bank is bleeding customers to FinTech apps, and it has hired you as a consultant. What it does have is data: six months of detailed behavior for thousands of credit-card customers.

“We’re launching a marketing campaign, but we can’t talk to everyone the same way. Split our customers into at least three groups that actually mean something — and tell us which group to spend the budget on.”

Your goals: exploratory analysis on real customer data, clustering (with a defensible choice of k), and translating statistical clusters into customer segments a marketing department can use. This is unsupervised learning: no labels, no right answer — your interpretation is the deliverable.

Background

This week’s lecture, Banking in the Age of AI, frames the business problem. Toolkit: StandardScaler, K-means, the Elbow method, and optionally PCA and autoencoders for dimensionality reduction.

The data

BankCustomersData_PCLab7.csv — one row per customer, 18 variables of credit-card behavior: balances, purchases (one-off vs. installments), cash advances, frequencies (scored 0–1), credit limit, payments, tenure. The assignment PDF has a full glossary of every column.

Your assignment

Part 1 — Know your sample (Task 1)

• Import the data and check for missing values — fill the gaps in MINIMUM_PAYMENTS and CREDIT_LIMIT with the mean; check for duplicates.
• What is the average credit limit?
• Correlations and distributions of key variables — show only what’s interesting.

Part 2 — How many groups? (Task 2)

• Scale the data first (StandardScaler — clustering on raw dollars vs. 0–1 frequencies is meaningless).
• Draw the Elbow curve to choose the number of clusters. Real elbows are rarely textbook-clean: defend your choice.

Part 3 — Segment and interpret (Task 2, cont.)

• Run K-means with your chosen k and describe each cluster. Does it make sense as a customer type?
• The bank wants to focus on high-spending customers who use the card as a loan: which cluster gets the marketing budget?
• Find a smart way to summarize and plot the segmentation — one chart the board would understand.

Part 4 — Reduce the dimensions (Task 3 — optional)

• Apply PCA (2 components) or an autoencoder and look at your customers in the reduced space.
• Does it change your view on the right number of clusters? Can you improve the segmentation?

Deliverables

• One Jupyter notebook per group, by email to the instructor (clement.mazetsonilhac@unibocconi.it) and the TA (andrea.andolfatto@phd.unibocconi.it) in cc.
• Email title — and notebook name: PCLab#7 - Group X - Name1 Name2 Name3. Tell us how long it took.
• Deadline: Friday, midnight (end of Week 8).

Grading

1. Submitting before the deadline;
2. Code quality — comments, readability, use of functions;
3. Notebook structure — well organized, explain what and why;
4. Completing the tasks and innovating — produce less, but more useful output.