A large DIY retailer had a simple problem: predicting when a customer was about to start a large project. The client recognised they couldn’t influence if or when a customer might start such a project, however they felt they could maximise their share of the project spend if they were front of mind at the right time. Several agencies had already attempted to solve this problem and failed. While the complexity of the business question was a factor in this, it was also due to dirty data stored in an antiquated IT system which made analysis and operational delivery difficult.
As strong data scientists we trusted ourselves to build a predictive model that was robust, however we also knew we needed to be mindful of how we were going to make this work for this business. Initially, we extracted and cleaned data from a variety of systems and, using advanced machine learning techniques (gradient boosting methods), we modelled customer behaviour. Once we had a strong model, our plan was to write the SQL/R scripts that would pull and clean the data, and then automatically produce the various mailing files. Data science is often seen as being delivered by software – this is not true. Understanding the business problem and translating this into the model is critical and makes modelling just as much an art as a science. In this case, the critical factor that drove success was the fact we felt that the number of items in consumers' baskets might be important.
We discovered that if a customer only bought a packet of birdseed, they were also likely to buy a kitchen or a bathroom. In fact, we found about 160 different baskets that only held one item showing the same correlation. The big question was, why?
Two overlays helped give us clarity. The first was the understanding that the DIY retailer was a destination store – people do not go to a store on a retail park just to buy birdseed, they are doing something else in store. The second was looking at the value of the purchase – kitchens and bathrooms are expensive so there is a long consideration period. After customers had finished browsing, they were picking up sundries as they exited the store. Simple!
Setting up a triggered mailing campaign program within the antiquated IT environment was challenging. Initially we chose approximately 10 key behavioural triggers linked to a single type of project – a new kitchen. We then built a process that would support a weekly mailing providing an offer to these people. Once this worked we gradually expanded the process out to the full 160 triggers across multiple project types. We then looked to improve the trigger response to be within days rather than a week and introduced other channels.
The performance was impressive. By year two the incremental revenue was £20m and eventually peaked at £36m. The PoC success formed the basis of a business case for a new marketing datamart, campaign management tool and increased headcount, all of which were implemented in year two.