A Simple and Useful Metric to Supplement RFM

RFM (Recency, Frequency and Monetary) has been one of the most popular and important approaches for segmentation and target modeling in direct and database marketing for the past 50 years. There are many variations and extensions to the basic RFM. For example, we expect customers with longer tenure to have higher frequency and are likely to be more loyal. As a result, customers’ tenure is often also included as part of the RFM.

What makes RFM so effective is that its components represent key “summary statistics” of the purchasing process that generates behavioral data. This purchasing process is in turn governed by the underlying consumer decision process. In other words, RFM succinctly describes what customers have done based on their decisions, and if we assume these decisions are stationary in the short-term, we can use RFM to predict future effectively.

One of the criticisms of RFM is that it ignores some detailed information. For example, RFM is silent on “temporal” patterns of a customer’s purchase behavior. When the purchase cycle is long, this is not a big problem, since recency will be the dominant feature of any temporal patterns. But for situations where the purchase frequency is high, such as in grocery retail, the recency metric may be less variable, since many customers will shop often. In such situations, another metric – the coefficient of variation (CV) – provides additional information that is useful for marketing.

CV is the standard deviation of an individual’s interpurchase time divided by this individual’s mean interpurchase time. In general, CV is a measure of difference and is not restricted to interpurchase time. We focus on its usage in depicting the temporal pattern of interpurchase time. If we look at all the interpurchase times recorded in the database, there are two sources of variations. The first variation is the within-customer variation; that is, how a customer’s interpurchase time varies from one transaction to another. For example, you may shop once per week, but sometimes you may shop twice in a week or skip a week entirely. This is also called “temporal pattern”.

The second source of variation derives from between-customer differences. If each customer’s interpurchase times follow a distribution, how do these distributions differ among customers? In marketing, the latter variation is often called “heterogeneity”. For example, the median trips per week may be 1 across the whole customer base, but some customers may shop much more often, while others are seen relatively infrequently. Such heterogeneity often leads to 80/20 rules in marketing where 20% of the heavy buyers are responsible for 80% of the total purchases.

For the purpose of this discussion,  we define CV at an individual customer level, thus summarizing the temporal patten of a customer over all his interpurchase times.

If a customer’s interpurchase time follows an exponential distribution, then his CV equals 1, since the exponential (and its counterpart in modeling frequency, the Poisson distribution) is a one-parameter distribution with identical mean and standard deviation. In marketing science, exponential interpurchase time or Poisson purchase frequency is the benchmark of a “random customer”, random in the sense that this customer’s propensity to purchase tomorrow is constant and independent of the lapse since last purchase. The smaller the CV, the more predictable or regular a customer is. For example, if a customer always purchases on Saturday and Saturday only (such customers actually exist in grocery databases), his CV equals 0, since the standard deviation of his inter-purchase times will be 0.

We should point out that CV is supposed to supplement RFM, not to replace it. For example, we can have two equally regular customers (and hence similar CV), but with very different RFM values. But when combined with RFM, CV provides additional insights. Here are some selected examples of how Fulcrum uses CV in analyzing high-frequency purchasing patterns:

1. CV is used as an leading indicator of customer attrition. Unlike banking or telecom, a customer’s attrition in retail is not directly observed. Retailers will commonly measure attrition by looking at a reduction of purchasing frequency between two time periods (e.g. if a customer has not bought in a year, they are considered lapsed). In our work, we have noticed that using CV,  has the benefit of identifying “silent” attrition and identifying it sooner than using some arbitrary period. If  a customer’s frequency has a decreasing trend, and/or her pattern becomes less regular (i.e. CV increases), we need to pay special attention to that customer, since both are signals of a diminishing relationship that may lead to attrition.
2. CV can be used as a proxy for customer potential. Consider customers A and B, both of whom have had the same number of purchases during the past 6 months. Customer A has a smaller CV and displays much more regular shopping intervals, while customer B tends to be all over the place. Holding other factors equal, customer A is likely to be more loyal, stable and less likely to churn. On the other hand, customer B is may have more potential, because the retailer is probably capturing a smaller share of wallet.  In other words, we can assume that Customer B’s variance in this setting is due to him splitting his purchases between retailers.
3. CV can also be used for analyzing the effects of changing competition. For example, let us suppose a competitor has open a new store in one of our markets. We can assume that the RFM value for some customers will change; some may even switch all of their business to the new store. But we may also observe some variations in CV: for instance, the CV values for mid-week shoppers may be changing more than those of weekend shoppers. This may suggest that we need to introduce some mid-week promotions to try to maintain that segment.

CV only works, if there exist multiple transactions at the customer level. For high-frequency categories, such as grocery, drug, and convenience, we often have hundreds of transaction per household. Medium-frequency categories, such as entertainment (books, videos, music) and home improvement can benefit from this metric, as well. For categories, in which purchases for the majority of customers are infrequent, CV probably does not add a lot of value, but then neither does RFM.

CV is easy to define, calculate, and operationalize. We encourage anyone already using RFM as a core customer metric to add CV as another important indicator of customer behavior.

Comments are closed.