When Not to Model (1,934 words)
Many companies are using models successfully to improve their response rates and increase profits in marketing to outside prospect lists. To use a model successfully, the marketer has to have a marketing situation that meets both of the following criteria:
1. The customer response to a promotion must be significantly determined by factors that the marketer can append to a prospect list; and
2. The lift in profits from using the model must more than pay for the cost of the appended data and the cost of building and running the model.
For example, if you are selling encyclopedias through the mail, you may rent a list of parents of high school students. Appended to this list you might be able to get the estimated household income, the PRIZM cluster code, the type of dwelling, the incidence of home ownership and a few other factors. It is possible that you could put these factors into a model and be able to prove that you can sell the encyclopedias profitably to people who: have a child in high school; have a household income over $X; are homeowners of a home worth over $X; and live in PRIZM clusters A, B, C and D.
You might find that if people have these four characteristics, your response rate is 2 percent or better. If they lack them, your response rate is below 1 percent. It could be that it cost you $Y per thousand to append this data to a rented list of parents of high school kids, and that your model costs you $35,000. If your mailing is big enough, the profits from using it to direct your mailing could pay for the cost of the model and the appended data.
Companies are doing this kind of analysis all the time. Banks use it to sell credit cards and home equity loans. Life insurance companies use it to sell life insurance policies, annuities, health insurance and retirement plans. Brokers use it to sell mutual funds. In many situations, such data appending and modeling is a profitable solution to the marketing problem.
NOT ENOUGH LIFT?
On the other hand, for most products and services, modeling does not provide sufficient lift to justify the expense. Thousands of companies have experimented with modeling and come away dissatisfied and somewhat poorer.
Why is that? In many cases, the factors that lead people to buy your product do not depend on data that you can buy to append to a prospect file. Often, modeling is not successful with the marketing of automobiles, packaged goods, fund raising, home sales, clothing, gasoline, furniture or hardware because rented data factors are less important than non-rentable factors in the decision-making process.
Buyers can be classified by two types of characteristics:
• who they are—demographics.
• what they do—behavior.
Some people buy books and read them; most people do not. Do the readers buy books because of their age, income, presence of children, or home value? Or do they do so because they bought books in the past and enjoyed them? In most cases, past behavior is a far better predictor of future behavior than any demographic factors that you can assemble about a group of prospects. That is why "mail responsive" lists are far more profitable for direct marketing than any other type of list that you can buy.
Mail Responsive Lists
To sell its $15 allergy-free air filters by mail, 3M purchased a large list of people who were known to suffer from various air-borne allergies. It asked a direct mail expert whether the project would be a success.
"No," he replied.
"Why not? These people need this filter!"
"You should use a list of mail responsive people. People who have bought something, anything, by mail that costs $15 or more."
"But these people may not need a filter," the 3M guy said.
"Trust me. I know what I am doing," the expert said.
So 3M mailed 100,000 to each of two lists: the allergy sufferers and a mail responsive list. The mail responsive list blew the other list away. Past behavior is the best predictor of future behavior.
Companies today are beginning to shift from acquisition mode to retention mode. They want to acquire new customers, but they find it more profitable to increase sales from the customers they already have. So now we get to the big question: Should you use modeling to channel the communications with your existing customer base? I say, "in most cases, no." Here's why not:
1. Once you have gained a customer and built up a database file on her, you know so much specific information about her demographics, desires, needs, family composition and purchase behavior that a model using appended data is really unnecessary.
2. Customers like to be treated as individuals, not as members of some data segment.
3. Use of simpler, less expensive techniques related to a customer's purchase history can result in higher profits than can possibly be gained through modeling.
We can base our customer communications on customer behavior that is stored in our marketing databases. You build a database that includes customer purchase history: a history of behavior. This behavior history is used directly (not using a model) to make relevant offers to your existing customers.
What Vs. When
We can classify the customer behavioral data that is stored in a customer marketing database into two types:
Product history: What specific products did she buy?
Transaction history: How much did she buy, how often and how much did she spend?
Product history is easier to understand, and widely used. The most typical use is for affinity analysis. You look at a customer's purchases and draw conclusions as to the customer's likelihood to buy other products. For example:
• Banks use checking/savings account behavior to find the best customers for credit cards and home equity loans.
• Insurance companies use behavior on life and health policies to offer mutual funds and annuities.
• Airlines study customer travel patterns to make vacation offers.
• Booksellers learn what books interest their customers from their existing purchases.
Transaction history is also very predictive but less understood. Most marketers today break transaction history into three related behaviors: recency, frequency and monetary amount (RFM). Over the last 50 years, marketers have learned that:
• Recent buyers are more likely to respond to new promotions than less recent buyers.
• Frequent buyers respond better than less frequent buyers.
• Big spenders respond better than small spenders.
Paul Wang and I give seminars several times a year for The Database Marketing Institute. Over the years, we have worked with many companies who have been experimenting with the intelligent use of the RFM data contained in their customer databases. They have discovered some fascinating information through the simple graphing of customer RFM behavior.
A Case Study
I do consulting work with an art gallery in Amsterdam, Holland, that is using RFM software developed at the Institute. The gallery began creating a database in 1993 and have sold art works to 267,028 customers since then. It recently did a promotion to an Nth of the database, consisting of 56,239 customers. From the data in the file, the gallery is able to determine the most recent purchase date, the frequency of purchases and the average order size. Using this information, it coded the file for RFM and created 75 RFM "cells," each with 3,560 customers. There are five recency, five frequency and three monetary divisions.
The gallery recently did a test mailing to 56,239 customers. In the test, it got 2,463 sales averaging $104.20 each, for total sales of $261,560. The outgoing promotion cost about 84 cents each, for a total cost of $47,240. Average profit margin on a sale was 57 percent, so the gross profit from the test was $149,089. If we subtract mailing costs of $47,240, the gallery made a net profit of $101,848.
Gallery staff studied the RFM graphics that the RFM software created from the test results. In previous years, it had always mailed the entire customer base for promotions. Using the RFM test results, this year, the gallery decided to do a selective rollout to determine whether it could increase its profits. The RFM graphics showed that by using this test as a guide for a selective rollout to its customer base, it could increase total profits by a whopping $93,487.
Let's look in some detail at how the gallery went about creating this additional profit.
Recency was a very important factor in customer response rates. It had five recency divisions, from most recent to most ancient. Each division was of equal size: 11,248 customers. (See Chart A for how the divisions responded to the test promotion.)
How profitable were the lower recency divisions? When looked at in terms of break-even, the lower divisions should not have been mailed at all (see Chart B).
Frequency of purchase was even more dramatic in its effect on the success of the promotion. All the frequency divisions were profitable, but in terms of break-even, only the first quintile (most frequent buyers) made a significant contribution to profit (see Chart C).
Perhaps the most interesting view of customer behavior came when these same buyers were compared in terms of their prior spending behavior. The biggest spenders definitely responded better and brought in more profit—from three standpoints. First, look at total sales (see Chart D). The gallery had only three monetary divisions. (In our discussions of monetary value, we have converted Dutch Guilders to U.S. dollars at a rate of 1 Guilder equals 48 cents and calibrated the chart in Guilders.) Why did monetary amount prove to be so decisive in terms of predicting sales? There are two reasons: High dollar spenders tended to respond better, and high dollar spenders had a higher average order size (see Charts E & F).
Finally the gallery looked at total spending by RFM cell. We looked at the break-even index: which cells were profitable, and which were not. The contrast was striking: Of the 75 cells into which customers were divided, 37 lost money and 38 made money. Armed with this knowledge, it didn't mail the entire customer base for the rollout, as it had done previously. Instead, it mailed only those customers in the 38 profitable cells. And to be sure that it wasn't making a mistake, the gallery mailed 10 percent of the customers in the unprofitable cells. The selective rollout resulted in mailing to only 148,455 customers instead of 267,028. This was a savings in postage of about $100,000—a very significant savings, since customers in these dropped cells would have made very few purchases. The result of this selective rollout was a profit gain of $93,487.
Note that this gain in profit does not result from the running of a model. There is no cost for appended data. There is no modeling fee. The cost of this exercise was essentially nothing, since it already had the data and the RFM software. If a model had been used, the profit would have been reduced by the cost of the model.
External appended data and modeling is profitable in some instances for prospecting. For customer communications, it is better to use customer product-specific data combined with affinity analysis and transaction-based data with RFM analysis.
Arthur Hughes is executive vice president of ACS Inc. of Reston, VA, a database marketing company. You may reach Hughes at DBmarkets@aol.com or (703) 742-9798. The RFM software mentioned in the article is available from Jon Lowder at (703) 730-5656.