 |
|
RFM:?
Is it "Kudzu" or is it Gold? |
|
|
In the July 15 issue of DM News, Jim Wheaton wrote a very interesting article, "RFM Cells: The "Kudzu" of Segmentation" He said that "more sophisticated statistics-based predictive models" are better than Recency, Frequency, Monetary analysis (RFM). Statisticians today, he pointed out, "generally agree that RFM should be relegated to history’s dust bin, to paraphrase a famous 19th century analyst". Jim Wheaton is an expert on modeling. He knows more about it than I will ever know. I have attended his lectures and read his papers. He knows his stuff. He is a good analyst and a good guy. But in this article, I think that he went a little overboard. I certainly agree with him that modeling can, and usually does, produce higher response rates than pure RFM. Whether it is more profitable in all situations, however, which his article clearly suggests, is another question. In this response, I would like to discuss modeling and RFM. I will show where modeling is best, and where RFM is best. Let’s start with a little background on RFM. How RFM Works RFM has been in active use in direct marketing for more than 40 years. It can be used only for customer files that contain purchase history. There are two methods of doing the coding. The traditional method, which Jim criticized, is called "hard coding" in which the categories are divided by exact values (0-3 months, 4-6 months, 7-9 months, etc.) The other method, which I favor, is called "exact quintiles". You sort your entire customer file by most recent date and divide the file into five exactly equal segments from most recent to most ancient. You do the same thing with frequency (number of times the customer has purchased something from you) and monetary (total dollars that she has spent with you). If you have exact quintiles, then you have 125 RFM cells of equal size (5 X 5 X 5). I number them from 555 (most recent, most frequent and highest monetary value) down to 111 (most ancient, least frequent, and lowest dollars). Why RFM Is Popular The reason why RFM has been so popular for so many years, is that it is a very inexpensive, simple and useful way to classify customers by behavior (as opposed to demographics). It is often an extremely accurate way to predict how a large number of customers will react to a particular promotion or offer, based on the response of a small test group. If your whole customer base is divided into 100 or more RFM "cells", and you mail an offer to a small representative sample of each cell, you can learn the response rates of each cell in the test. The members of the corresponding cell in your unmailed large customer database should be almost identical to those in the test. This can be powerful knowledge. Modeling works the same way. A predictive model of a customer file uses the same input data as RFM: most recent date, frequency of purchase, and total dollars spent. But, models also often include demographic information as well, such as family income, presence of children, type of house, length of residence, etc. Instead of grouping customers into a fixed number of RFM cells, modelers produce an algorithm which can be used to create a score for each household. The score indicates how likely it is that the household will respond to a particular offer. Using a three digit score for each household usually results in better predictive ability with a model than with the less sophisticated RFM cells. So, Jim is correct. Modeling is almost always better than crude RFM, because modeling includes the basic ingredients of RFM but then adds other data and sophisticated CHAID or regression analysis. Where I disagree with Jim is in the issue of economics. The reason for using RFM or predictive modeling is to increase profits. RFM is crude, but very cheap to do. To do RFM you do not need to hire an outside expert. Any marketer can do it. The coding of a file to create RFM cells is a simple matter for your programmers, or you may use an inexpensive PC package like RFM for Windows. The analysis is done on a spread sheet. Modeling, on the other hand, is sophisticated, usually takes a few weeks, cannot be done by an untrained marketer, and is usually somewhat expensive. A Practical Example Let’s take a practical example to see the difference. Suppose that a marketer has 200,000 customers in a database, complete with previous purchase history. He codes his database by RFM, dividing it into 125 "cells" with approximately 1,600 customers in each. He does a test mailing to 40,000 of these cells. He is selling a product priced at $100, which yields a $40 net profit on every unit sold. Let’s suppose that it costs him $0.50 for every outgoing piece mailed. His total sales are 600 units (1.5% response rate). The test mailing cost him $20,000 (40,000 X $0.50). The profit from sales was $24,000 (600 X $40). He made a net profit of $4,000. At this rate, if he had mailed the whole 200,000, he would have sold 3,000 units and made a net profit of $20,000.
However, since he has coded his file for RFM, he can double his profits by using the information he has learned from the test. Let’s say that after reviewing his test results, he discovers that only 50 of the 125 RFM cells "breaks even" or better. This is a typical result of RFM analysis. Break even means that the profit from sales to members of the cell just pays for the mailing to the cell members. Since only 50 cells broke even or better, it means that 75 of the cells did not break even. He lost money by mailing to them. On his rollout, he will mail only to members of the 50 successful cells. How many will he mail? Approximately 80,000 customers (50 X 1,600). If he is doing his job correctly, he will, of course, remail the 16,000 customers who were in the 50 profitable cells in the test. Only a dunce would leave them out. What will be the result of his mailing to the 80,000 customers? This can be calculated with great accuracy. Using RFM, the rollout response rates of each cell usually mirror the test response rates of the corresponding cell. Let’s say that the overall response rates of the 50 profitable cells averaged 2.5% on the test. The sales on the rollout to these cells should be similar (or slightly smaller, since rollouts seldom do as well as tests). His total sales to the 80,000 will be about 2,000 units (80,000 X 2.5%) for a net profit of $80,000 (2000 X $40). Against that we must subtract the $40,000 he spent in mailing to these 80,000 customers. His net profit from the mailing would be $40,000 – exactly double the profit he would have made if he did not use RFM, but mailed the entire 200,000. So RFM has increased his response rate from 1.5% to 2.5% and given him a net increase in profits of 100%. His ROI is 200%. For every dollar invested, he now has $2. Modeling Results In Higher Response Rates Now lets consider modeling. Suppose that he pays a professional modeler to do a predictive model of his customers using an identical test. The modeler appends demographic data to the 200,000 customer base (@ $20 per thousand) and develops a model based on the purchase history, the demographic data and the test results. Models are more sophisticated and usually do better than RFM, so lets assume that as a result of his model, in the rollout he again mails only 80,000 customers, but by better selection due to the model, he achieves a response rate of 3% instead of 2.5%. The model gives him results that are 20% better than the RFM would have produced. That is not at all unusual. The mailing results in a sale of 2,400 units (80,000 X 3%), with a net revenue of $96,000 ($40 X 2,400) after spending only $40,000 on the mailing. He has made a profit of $56,000 instead of $40,000! He is $16,000 ahead. Clearly, modeling beats RFM! But, hold on there a moment, partner. We have not yet figured in the cost of the model and the demographics. Appending the demographic data (@ $20/M) costs $4,000. His modeler takes two to three weeks and charges the usual fee of about $25,000 for doing the model. Total cost of the model and data is $29,000. This being the case, the total cost of the operation, including the mailing ($40,000), the demographic data ($4,000) and the modeling ($25,000) comes to $69,000, and his profits from the sales are only $56,000 – he has lost $13,000 by modeling! Outrageous, you say? Perhaps I have exaggerated the cost of the model. Let’s assume that our modeler works cheap, and instead of $25,000 he accepts only $12,000 for his model. We would just break even. In fact, few if any modelers will build a predictive model like this for as little as $12,000. This being the case, there is no way that you can make a profit by using a model in this situation. RFM is faster, less expensive, and produces greater profits, although with lower response rates. So why does anyone use models? For two reasons: a) modeling can be great for customer acquisition mailings. Prospect files seldom contain purchase history and therefore cannot use RFM b) modeling of customer files can be profitable if you are dealing with very large numbers of customers. If our customer file, in the above example, were 2 million instead of 200,000, then the model would win, since the larger numbers permit us to absorb the modeler’s fee of $25,000 and the cost of demographic appending. Conclusion: RFM Will Survive Good old RFM will be around for a long time to come. It is a very old, and tried and true technique, simple to use, inexpensive, and it works. You can do the coding of your customer file for RFM with a simple program developed by your own programmers, or you can use a PC package like RFM for Windows (703-644-4830). You don’t need to hire a modeler. Just use a spread sheet to keep track of your response rate by cell. You will have to agree, though, that Jim is right. Head to head, modeling usually produces better response rates than RFM. But with small customer files, RFM often produces higher profits. Which is more important: better response rates or higher profits? By the way, who was that "famous 19th century analyst" that Jim cited as the source for his quotation on what should happen to RFM? None other than Karl Marx, the founder of Communism. An interesting source for a quote on modern marketing techniques!
|
|
