RFM, or ‘recency frequency monetary value’ to give it its full name, has long been the targeting tool of choice for the home shopping industry; so we decided to give it a challenge by building as an alternative a propensity model using exactly the same data set.
RFM is in reality a form of segmentation – customers are classified into a number of groups, usually in descending order of importance, according to their previous purchasing activities.
Propensity models can be developed in many different ways but they always use historic behavior to classify people into two groups – usually those who make a purchase as a result of being made an offer and those who do not. At this point common information known about people in each group is used to build a model to predict which group people are most likely to fall into.
An RFM score will describe the overall strength of the relationship between a business and a customer, but the question is whether we can improve on that by building a propensity score targeted at a specific purchase activity or category.
A great advantage of RFM scores is that because they are not proposition specific, they can be used across a wide range of applications; however, if the scale of any actual marketing selection is substantial enough, then the extra resource required to build the propensity score may be justified.
In addition, a propensity model can take into account not only RFM based information, but also factors like age, gender and other demographic information that might be available on customers.
In this example we are dealing with data from a home shopping company with over 1m customers, and a large number of merchandise categories. We first used cluster analysis to group the merchandise categories into six high level merchandise groups.
The RFM score was then built on customers buying across all six merchandise groups whereas the propensity model was developed for one specific group. We used those who had purchased from the specific merchandise group in the previous three months as the target variable for the propensity model.
In order to compare the two targeting approaches, we selected deciles within the customer base by each method, and then looked at the proportion of actual buyers that we found within each decile.
In reality most home shopping companies will only select the top one or two deciles for campaigns, and for this example the benefit of using a propensity model over an RFM is 49% for the top decile, and 43% for the top two deciles combined.
Another way of looking at the impact of different model techniques is to build a Gains Chart; this gives a graphical illustration of the difference between the two techniques.
Our conclusion is that where you have sufficient mailing volume to justify the not great cost of building a propensity model, then that is usually the right route to adopt.
To put all this into a commercial context, we built a table to show the difference in net value obtained from using propensity model scoring versus RFM, on different sizes of file.
The volume is the number of customers in each decile, so a customer base of 100,000 will have deciles of 10,000.
When the customer file is say 50,000 in this example, the gap between the orders value using the two different techniques is £15625 less £10875 or £4750, which is much more than the usual cost of developing a propensity model.
There are no hard and fast rules about by how much a propensity model will outperform an RFM segmentation, but generally it will perform better when there is something specific to target beyond general overall historic purchase behaviour.