Customer behaviour prediction for retailers
Customer Purchase Predictor
Predicts which customers will shop again soon, what they’re likely to buy, and how interests shift through the year — built and tested on 36,594 real orders.
- Demo on real, public retail data
- Retail
- Automation & prediction
- Working demo
Most retailers already hold years of sales history, but marketing still goes to everyone at once. The questions that matter are simpler: who is likely to come back soon, what are they interested in, and when should we talk to them?
To show how we answer them, we built the service on a public dataset of real transactions from a UK online gift retailer: 5,852 customers and 36,594 orders between December 2009 and December 2011. Every number on this page comes from that data.
The goal
Talk to the right customers, about the right products, at the right time.
Who will be back, and when
For each customer, it estimates when their next order is due from their own shopping rhythm, and which products they are most likely to want. For regular shoppers, the predicted window caught the next order 49% of the time, within 3 days.
What customers want, month by month
It tracks interest in each product category through the year. Christmas items went from almost nothing in spring to one pound in ten by November — and the rise was visible from July.
What sells together
Across 36,594 baskets, it finds products customers buy together, so recommendations and displays follow real behaviour rather than guesses.
Tested on what actually happened
We trained the model on data up to June 2011, scored every customer in September, then checked who really ordered by December. Of the top 20% it flagged, 83% ordered — against 44% of customers overall.
Top-scored customers ordered at nearly twice the average rate: 83% against 44%
Seasonal interest visible months before the peak
Autumn sales forecast within 5.7% on average