Our client is a direct marketer of home improvement goods, offering merchandise through catalogs and on the internet. The client currently uses sophisticated RFM segments to make circulation decisions about which customers should receive catalogs. There are hundreds of RFM segments, developed over time, defined by combinations of recency, frequency, sales, channel, product, and more.
Clario Analytics built predictive models, based on a series of three historical catalog mailings. Using Clario, we put together the historical mailed samples, computed historical attributes, pulled sales results, built, validated and scored the models.
We built three sub-models, corresponding to 0-12 month, 13-36 month and 37+ month recency groups.
We built a response model and an average order size model for each of these groups, and combined them into three sales/mailing models. Logistic regression was used for the response model and ordinary least squares regression for the average order size model. The final sales/mailing model rank orders customers according to their expected sales/mailing for upcoming catalogs.
We measured results by comparing the model ranking vs. the RFM ranking on each of the three mailings in the historical sample. Results are shown in the graphs, which correspond to the three recency groups. In each case, you can see that the model segments rank order Sales Per Mailing better than RFM segments. The blue line shows actual performance of a historical mailing, by segments ranked according to model score. The red line shows actual performance of the same mailing, by segments ranked according to RFM.
The goal is to find higher performing customer in the top segments, and lower performing customers in the bottom segments. The model is a much better tool to differentiate who should be mailed and who should not be mailed. For example, mailing the same depth (60%) of 13-36 customers produces 17% more sales using the model than using RFM.
This analysis shows where the models can be used to eliminate unprofitable mailings. The models can also potentially be used to find additional customers to mail (who were not mailed in the past). We will update the model scores on a regular basis, each time a new catalog goes out, to use in selecting the appropriate customers to receive the upcoming catalog.