Hangi Müşteriler Ertesi Yıl Geri Geliyor?Which Customers Come Back the Next Year?

Can you look at a customer list and tell who will come back next year? In this study I derived four customer segments from an online retailer's first year of data, then tested them against a second year the model never saw.

Data and cleaning

The data is UCI's open Online Retail II dataset: two years of invoices from a UK-based online gift retailer, many of whose customers are wholesalers. The raw data was not ready for analysis; I cleaned it in four steps.

• Returns were kept. Instead of deleting 17,586 return rows, I deducted them from each customer's revenue (£709,953 in total). Had they been deleted, a customer who placed a huge order and then cancelled it would look like one of the most valuable.
• Rows without a customer were left out. 22.8% of rows have no customer ID; they account for 13.7% of revenue. They cannot be segmented because nobody knows whose they are.
• Overlapping days. The file's two sheets both contain December 1–9, 2010. Most of the 34,335 exact duplicate rows come from there.

Method

For each customer I computed three variables: days since the last purchase (recency), number of invoices (frequency) and net revenue after returns (monetary). All three are right-skewed: 33.5% of customers placed a single order, while a few spent hundreds of thousands of pounds. So I took logarithms and standardized them; otherwise K-means locks onto a handful of giant customers.

How many clusters?

I chose the number of clusters with two measures: the silhouette score and stability, meaning how consistently customers land in the same cluster when the data is resampled and the model rebuilt.

The highest silhouette is at two clusters, but that solution only separates active from inactive customers, which is hard to act on. The best silhouette after two, together with high stability, is at four, so I chose four. A silhouette of 0.34 also shows the clusters are not sharply separated: customers sit on a continuum, and the segments cut that continuum into usable pieces. The K-means post covers the method in detail.

Four segments

17% of customers bring 64% of revenue

Champions are one in six customers and nearly two thirds of revenue. Dormant customers are the largest group (35.5%), yet they account for only 5.5% of revenue.

The segments separate the next year

The segments saw nothing but first-year data, yet behavior in the second year separates clearly. 96% of Champions came back and spent £6,348 per customer on average. Among Dormant customers the return rate is 40% and the average spend £247.

What to do with these segments

• Champions. They come back anyway. Discounts erode margin; the priority is not losing them: priority service, early access, guaranteed stock.
• Regulars. A quarter did not return the next year. The best group for testing reminders when the gap between orders grows.
• New and small. A third bought nothing in the second year. The goal is to speed up the second and third order.
• Dormant. Low-cost win-back campaigns. Because 40% come back anyway, any campaign's effect must be measured against a control group.

Limitations

Conclusion

Three simple variables and four clusters were enough to separate behavior a year later: 96 of every hundred customers in the best segment came back, against 40 in the weakest. That is where segmentation earns its keep: it turns a customer list into groups that each call for a different decision.

The most instructive part for me was the validation. Clustering produces a result on any data; only a period the model has not seen shows whether that result is useful.

Data: Chen, D. (2019). Online Retail II. UCI Machine Learning Repository. Amounts are in pounds sterling.