Continuing on the theme of segmentation, RFM Analysis is another tool for understanding and identifying different types of customers.  RFM stands for recency, frequency and monetary value.  This tool will help you:

1. understand customer value quickly when limited data are available (e.g., just purchase data)
2. develop a basic value segmentation that can be used to determine if your customer strategy is optimal
3. find untapped markets if there are segments which are not targeted
4. gain insight into gaps that might exist between accepted wisdom about the customer base and actual purchase behavior

The name suggests that recency is the most important factor for determining a customer’s value followed by frequency and monetary value.  However, you can set different priorities.  For one of my clients, monetary value was more important than recency and frequency.  Thus, their analysis was driven by monetary value first, recency and finally frequency.  It all depends on your product and the typical buying cycle.

The actual analysis involves calculating the R, F, and M dimensions, specifically:

1. creating a reasonable number of categories based on the date of most recent purchase (e.g., date was within the last month, within most recent 2 to 6 months, within prior 7 to 12 months, etc.)
2. breaking the number of purchases into a reasonable number of categories similar to recency
3. summing all revenue and creating a reasonable number of categories similar to recency

The number of categories you create depends on how you intend to implement the RFM analysis and should be guided by the means and standard deviations of the variables.

The fun part comes when you bring all of this together.  You first need to decide which dimension is most important and which is the least important.  Next, you need to determine the number of segments you want.  Will it be high, medium and low or 1 through 10?  If there are too few segments, then the segmentation will not be very targeted.  If there are too many segments, it may become a burden to implement and may ultimately be considered too complicated to use.  Business judgement and knowledge of the customers’ behavior should drive the creation of the segments. 

Once the segments have been decided, business rules or code can be written so that the segments are applied to your customer base on a regular basis.  This has the advantage of identifying new best customers or up and comers that can then be targeted with a special welcome communication.    Further, the segmentation can be used with other tools to drive marketing messages and campaigns.  However, you may need to revisit your RFM segments from time to time as your business changes significantly.   For example, if you raise or lower prices significantly after the segments are put into production, you will want to reassess the original recency categories.

Cluster segmentation is a descriptive, multivariate technique that creates distinct, homogeneous groups within your customer base.  The goal of cluster segmentation is to classify consumers or businesses based on behaviors, demographics or firmographics, and/or attitudes.  In this way, you can develop more targeted programs and tailor messages based on the needs and characteristics of specific groups.  One client reorganized their marketing department as a result of a segmentation project I worked on, assigning one marketer to each segment so that consistent messaging and product offers could be employed against each customer group.  Further,the segments that are developed can be combined with models or other segmentation schemes to identify the best customers to target for particular campaign or offer. 

Determining what methodology to use for clustering depends on many factors including your clustering software, the type of data you have, and the number of consumers or businesses available for segmentation.  You should also consider the optimal number of segments to meet the business objective and which behaviors or other factors are most important in defining customers.    

Regardless the methodology chosen, you will need to do data prep.  You typically start with data summarized to the household level for B2C analysis and establishment or enterprise level for B2B analysis.  You might also need to do missing value substitution, transform categorical variables to binary or scaled variables, weight variables to drive preferred ones into the solution,  and standardize continuous variables.

Data reduction might also be necessary if you have many variables.  Tools for data reduction include correlation analysis, principal components and factor analysis.  

Once that is complete, you can create your segmentation schemes.  I run many more segmentation solutions than I show to a client because I want segments that are actionable within the client’s marketing plans and that are intuitive as well as not overly complicated.  In addition, I test the validity of my cluster solutions through goodness of fit statistical measurements and by replicating my results on a hold-out sample.   The end result is that a company can align its marketing efforts against segments, taking a customer-centered approach rather than treating every customer the same.  Cluster segmentation can be a tool for giving the right message at the right time to the right person.

Modeling is a powerful tool that is worth considering when determining how best to spend your marketing dollar.  At its simplest, modeling looks for patterns in data to predict future behavior.  That data could be past behavior.  If someone bought diapers last week, it is very likely they will buy them again this week.  It could also include demographics such as age and gender or, in a B2B context firmographics, the number of employees and annual sales volume.  Attitudinal information, such as willingness to purchase a product, could also be used in a model.  The power of modeling comes from the fact that it weighs all of the factors and results in a unique algorithm that predicts future behavior.  Instead of the usual “spray and pray” approach, modeling enables you to focus your dollars where they will have the most effect.

Two articles in the Wall Street Journal last week offered real life examples of how models can solve business problems.  I have seen clients use attrition models and proportional hazard models to determine which customers are likely to leave.  Google is building an attrition model to identify which of its employees are most likely to leave the company for another opportunity.  Presumably Google will target those employees most likely to leave and be able to retain valuable talent that might otherwise walk out the door.

Chrysler’s digital agency has designed a media modeling system according to the Wall Street Journal.  It sounds like a marketing mix model and is being used to allocate Chrysler’s marketing dollars.  At a basic level, this model tells Chrysler how much money needs to be spent on marketing to drive a certain number of vehicle sales based on the web traffic generated.  By monitoring online activity and tying it to their marketing campaigns, Chrysler has determined how many web visits translate into sales.  The media modeling system, including enhancements based on the ongoing performance of television advertisements, has helped Chrysler determine how to structure their marketing campaign and tweak marketing in real time to drive results.

These two examples may not fit your exact situation but they highlight the power and value of modeling.

James Surowiecki writes in the April 20, 2009 edition of The New Yorker that during hard economic times, companies cut costs including labor, advertising spend, and research & development, as well as forego acquisitions even though prices are lower.  However, companies that remained market leaders during the 1990-1991 recession increased spending according to a McKinsey study.  Research by Bain found that recessions can be opportunities for companies to leapfrog over their competitors.

Thus, CEOs and senior management have a difficult choice.  They can choose to slash costs in order to win a war of attrition – becoming lean to survive the recession assuming that their competitors drop out of the market or business altogether.    Alternatively, they can invest in advertising, research & development and acquisitions in order to grow market share and transform themselves into market leaders.  However, there is no guarantee.  Do you risk sinking the boat in order not to miss the boat?

Tangyslice had a recent post about customer retention metrics.  To his list, I would add lifetime value.  There are many ways to calculate lifetime value but let me tell you about one way I have calculated it.

A cellular phone company was losing customers due to churn and wanted help to retain their most profitable ones.  In other words, they wanted to know the future value of their customers.  With a lifetime value model, the company could increase ROI through targeted retention and provide one-to-one marketing.

In the cellular phone industry, most customers sign up for two year contracts.  However, some customers default on their contract and others continue their service even when the contract ends.   Thus, the lifetime value model in this example consists of two parts: 

1. survival analysis which predicts survival probability, the likelihood that a customer will remain a customer
2. financial data that include revenue and costs used to determine future customer profit

The first step was estimating the tenure of each customer.   A proportional hazards model or baseline hazards model can be used to estimate tenure.  Once the tenure is determined, the next step is estimating the future value based on past average monthly spending by the customer and the cellular phone company’s costs.

The formula calculates the discounted profit.  In the calculation above, i= the cost of capital and the terminal value is an estimate of the revenue beyond the 36th month of tenure. 

Does this bring back nightmares from Finance and calculating the discounted present value  of cash flow?  The same advice applies.   Be very careful how you calculate the terminal value because it can account for a large percentage of a customer’s estimated value.

I was catching up on Tangyslice’s blog and enjoying his 5 meaningless marketing metrics post, when I thought of another meaningless metric.  Last week I was reading a presentation which described the response rate of one group as slightly greater than the control group.  What does slightly greater mean in this context?  Well, it turns out the difference was statistically significant once I did the math.  What I find meaningless is when analysts do not look for statistical significance when comparing two groups.  This is known as A/B Testing.

Conceptually A/B testing is very simple.  You are comparing Group A to Group B.  A might be a control group and B the test group.  Alternatively, A and B might be two different offers, landing pages, e-mails, direct mail lists, or landing pages.  As the name suggests, this is a test which is why A/B testing is also known as split testing.  Ultimately, you want to know if A and B differ in a way that is statistically significant.

Here’s an example to make it concrete. Let’s say that you marketed to 50,000 customers encouraging them to purchase product A and 5,000 of them responded. That is a 10% response rate. In addition, there were 5,000 customers that you could have marketed to but that you did not.  Instead, you assigned them to the control group.  They look and act just like the 50,000 customers that you mailed. The reason for the control group is that some customers might buy product A regardless of whether you market to them or not.  In this example, 450 of them or 9% purchased the product. Is the difference between 10% and 9% statistically significant?  Was the campaign successful?

In this case, we perform the two-proportion z-test for equal variances using the following formula:

and

where…

p1=10% (response rate for Group A)

p2=9% (response rate for Group B)

x1=5,000 (number of responders in Group A)

X2=450 (number of responders in Group B)

n1=50,000 (quantity mailed in Group A)

n2=5,000 (quantity mailed in Group B)

If the value of z is greater than 1.96 then the difference is significant at 95% confidence.  In this case, the z value is 2.26 so the difference is statistically significant.

In order for the test to be valid a few assumptions must be met:
1. Your control group needs to contain customers or prospects that look and behave like the treatment group
2. You need to have sufficient numbers of direct mail recipients and responders such that n₁ p₁ > 5 AND n₁(1 − p₁) > 5 and n₂ p₂ > 5 and n₂(1 − p₂) > 5 and n2>29 and the groups contain independent observations

The math might look scary but really the hard part is making sure that the test is done properly.  It is vital that the control contains a random selection of customers who are similar to the treatment group.  If not, you could end up with very strange results

I was asked recently how to prioritize new customers if you do not have demographic or firmographic data available.  In other words, what can you do with just the data from the first purchase with which to work?

To make this more concrete, let’s consider the following situation.  You are asked to call each and every new customer who has made a purchase.  The question is, how do you prioritize the calls?  You want to make the first calls to those with the greatest potential to become loyal and valuable customers.  The only data available relates to the first purchase:  total revenue generated, products purchased,  product revenue, etc. 

In this case, a linear regression could be used to help you identify the factors that predict lifetime value.  (Other types of models can be used depending on the independent and dependent variables available.)  Using your existing customer base, build a model that leverages data about the first purchase to predict lifetime spending.  You can identify the best and worst new customers using the resulting model equation.  Armed with this insight, you can test your model by calling on new customers with the best predicted lifetime revenue and a random selection of new customers regardless of predicted lifetime revenue.  In addition, you can test call back timing to determine if there is an optimal call back window. 

Even with limited data, analysis can lead to insight.  Further, there is always an opportunity to incorporate testing.  In this case, testing can validate initial findings and help you learn more about the purchase cycle.