H&M Customer Retention Analysis

The question

Which H&M customers are quietly drifting away, how much revenue is at risk, and which ones are actually worth saving first? The context is a fashion retailer under pressure from faster, cheaper rivals. The data carries no competitor information, so I never claim to know why customers leave.

What I did

• Worked a 10% customer sample of the public H&M dataset, 137,571 customers and 3.2 million purchases, with DuckDB SQL on a laptop, and checked the sample against the full 31.8 million rows.
• Built a retention funnel and sorted every customer into on track, slipping, drifting or gone quiet by comparing their silence against their own usual gap between visits.
• Trained a logistic regression churn model, then read its predictions against spend rather than accuracy alone.
• Shipped a Tableau Public dashboard and a script where a free local AI model writes a monthly summary, with a built-in number checker added after an early version mixed up customers and revenue.

What I found

• About a fifth of yearly revenue sits with customers who have gone quiet or are drifting, and roughly 12.1% is likely gone for good.
• The model scored 0.77 AUC on a later period it never saw, but its riskiest customers were small spenders, so chasing accuracy alone would have protected almost none of the money.
• Ranking the same top 20% of customers by risk times spend catches 49.4% of the revenue that later goes quiet, against 11.0% for a plain churn model, about 4.5 times the money from the same sized list.