A multi-billion-dollar speciality retailer marketed to consumers using primarily traffic-driving coupons (e.g. “20% Off Any One Regular Price Item”). However, the client had a variety of different coupons in the rotation that consumers could use in any given week. This made it difficult to forecast promotions effectively and the client turned to a team of data scientists to get an accurate picture of which coupons would work.


Many retailers use coupons to drive traffic to their stores. However, it becomes difficult to track the success of a singular promotion when many coupons are in the rotation. The client had several different coupons consumers could use on any given week, including:

  • 20%, 30%, or 40% off one regular priced item
  • 20% off your entire basket
  • 30% off all items within a department
  • 20% off all clearance items

It became increasingly difficult for the client to forecast with all the variations of these promotions in play because the client still used traditional econometric methods. While these can accurately forecast single product promotions, they are not as reliable when it comes to the “entire basket” discounts. Therefore, the client could not forecast which products customers would be inclined to buy and had a difficult time predicting the effects of different coupons. Moreover, like most retailers, the client relied on promotions around seasonal assortments and holiday spikes. If there were any changes in the prediction forecast, the client could not accurately say if it was a promotion issue or a consumer pattern change affecting all retail shopping.


A team of data scientists approached the problem differently. Instead of focusing on the coupon, they focused on the consumer, and each consumer’s buying history.

The team looked at the entire customer database and predicted the likelihood any individual would come to a given store, in a given week, given the coupon that’s offered. Because they understood each consumer’s individual purchase history, they could accurately predict the probability an individual shopper would buy various products.

While this seemed like a computationally massive problem, the team was able to crunch the entire problem in-memory – really quickly and economically.


The results were an unprecedented increase in forecast accuracy, cutting the error rates by >50% across the board. For the first time, the retailer was able to get an accurate and precise picture of what will happen and a simulator that allowed them to test the effects of different coupons and how consumers would respond.