Leveraging the Information on Your Customer DatabaseThe following article, written by Rhonda Drake, appeared in DM News, July 27, 1998. It is a useful article and provides direct marketers with hints and tips on how to maximize the value of their customer information.
Collecting, updating and maintaining customer information on your database is only part of the picture. Knowing how to use and exploit that information to your advantage is the real challenge.
In my experience working for and consulting to direct response marketing firms, I have found that the majority of these organizations are not properly leveraging their customer data. Being able to comb through, manipulate and leverage the information on a customer database is a unique skill, maybe even an art form. So, it shouldn't surprise anyone that relatively few direct response marketers are actually applying sophisticated "data manipulation" techniques in an effort to exploit their customer database.
Most direct response marketers fail to spend the time and resources required to get close to the data in order to gain an understanding of their underlying relationships. Additionally, most direct response marketers fail to properly answer two important questions: "What information will predict my customers' behavior?" and "How can I derive these predictors using the information I have on my database?"
At a high level, the answer to the first question is straightforward. I like to think of the types of information that predict a customer's behavior as falling into four broad categories: (1) the specifics of the current offer relative to other offers either accepted or not accepted by a customer, (2) the relationship of the product being offered to prior products offered/purchased, (3) past promotion and purchase history, also known as recency/ frequency/monetary measures, and (4) demographic and lifestyle measures.
Having identified the four main predictor categories, you can begin to aggressively tackle each one in order to maximize the predictive value of your data. Below I give some fairly simple examples of different ways to manipulate and exploit customer information by predictor category.
OfferFor each offer made, try categorizing them by the "type" of offer. That is, create a variable on your database that indicates, for each order, what the offer was. Examples might include: "half price offers," "sweeps driven offers," "installment billing," "free trial offers," "discount offers," etc. This information will allow you to determine, for an upcoming campaign, who may or may not respond to the offer based on their reaction to the same type of offer extended them in the past. You may want to segment your mailing universe based on your customers past "offer" behavior. Doing so will allow you to give different customers different offers based on their past "offer" behavior. In other words, don't give every customer an installment billing option, only give it to the customers that you know need such an enticement to order.
Product AffinitiesA customer who has purchased a similar product to the one you're currently promoting will be more likely to order than a customer who has never purchased a similar product, all else being equal. It's that simple. So capture every product purchased on your database. Doing so will allow you to effectively cross-sell related products. You can further strengthen the use of past "product purchases" in your response models by grouping purchases together based on their similarity. For example, if you have a catalogue with a variety of product offerings, you may end up with the following affinity groupings: kitchen, clothing, toys, jewelry, etc. Grouping product purchases based on their similarity will allow you to measure the degree of interest a customer has within each "affinity category" by examining the exact number of purchases they made in the categories. Additionally, tracking these affinities can allow you to characterize customers based on their loyalty to an affinity or their breadth of purchases from many. Most companies analyze profits by affinity in order to know which products/product lines are contributing the most to the bottom line. But most do not know which affinities attract new customers or how a customer may migrate from one affinity to another. Understanding these dynamics in customer purchasing behavior allows for efficiencies in establishing your marketing strategies.
If you don't know how to categorize certain products because they are either unusual or are all closely related, consider multivariate statistical techniques to help you gain insight.
Customer Purchase and Promotion HistoryKnowing what your customer did in the past is going to be your strongest predictor of what they will do in the future. So capture the specifics of all promotions sent to a customer regardless of whether of not they ordered the product. Some interesting examples of how to look at and combine recency/frequency/monetary measures are shown below. The key to maximizing the use of your customer information is to think out of the box. There are a lot of other options that will be much more predictive of customer behavior than those plain vanilla "roll-up" fields.
Examine the elapsed time between eventsWe all know that the more recent a customer's action, whether it be an order or a payment, the more likely they are to make another purchase in the future. In fact, most direct marketers segment their customer file based on recency. With a little creativity, you can make your recency variables much more powerful. For example, consider building some "elapsed time" variables. One of the strongest "elapsed time" variables you can build is "the elapsed time that has occurred since a customer's most recent paid order." In other words, how long has it been since that customer last paid for an order. This variable combines a measure of recency with that of a payment. It will not only qualify customers regarding the recency of their order but will also evaluate the paid order status. This new data element should prove to be a strong predictor in your response model even if you have already segmented the file based on recency.
There are many other strong predictors you can build by looking at the elapsed time between events. Ask yourself the question, "How can I make this variable or data even more meaningful and predictive?" For example, "Can I add in an element of recency, frequency and/or payment?" "Could this variable gain predictive power if I create various versions specific to each product line/profit center or affinity grouping?"
Create some ratiosConsider creating the ratios of "roll-up" fields such as "the total number of products ordered" to "the total number of promotions sent." This particular ratio is quite interesting in that it is basically calculating a customers overall likelihood to order. You can even add in an element of recency by simply restricting the calculations of this variable to include only orders and promotions within a certain time period (e.g., within the last 12 months). The possibilities are only limited by your ability to get close to the data in order to gain an understanding of their underlying relationships. But remember, as mentioned above, in order to create such a variable, you must capture the specifics of every promotion sent to your customers.
Think longitudinallyCapturing several of the most recent sequential actions of a customer in one variable will also help you predict how a customer will respond to a future promotion. The concept is very similar to that of time-series modeling which looks at past trends in the data to predict future observations. The same concept can be easily applied in direct mail modeling. For example, someone whose responses to your last three promotions were "order, order, order" (the first order being the most recent of the three) will be much stronger than someone whose last three responses were "order, no response, no response," all else being equal. In this example, if we would have only looked at the most recent response for these two customers (an order in both cases), we may have considered them to be very similar in terms of their likelihood to order a future product - which is clearly not the case!
Longitudinal variables can be very strong for predicting response and payment but they can be complex to implement. Keep in mind that the more sequential actions or data points you try to look at, the more complicated the variable will be.
Demographic Enhancement DataIn general, demographic/enhancement data will pale in comparison to your own product and promotion data regarding its predictive ability. However, demographic data will play more of a major role for lists where your own product and promotion data are sparse or non-existent (as in the case of acquisition or new customer mailings). In addition, if you have enhancement data on your file, you may want to try using it to correlate such things as age and income to your offers and products.
Leveraging the information on your database through data investigation and manipulation can be a complex and time-consuming project. The examples above are only the tip of the iceberg regarding what one can do to fully exploit the information contained on their database. The most important thing to remember is to give yourself ample time and adequate resources so that you can (1) get to know your data, and (2) clearly identify your specific data needs. This project could be one of the most profitable initiatives you take on this year. You most likely have the data, all you need to do now is make it work a bit harder for you!