Feature Article
September 2001

Customer Attrition Vs. Value Attrition

BY STEVE GALLANT

[ Go Right To: Canary In A Data Mine ]


In the tightening economy, most companies feel the pain associated with losing customers. However, it is more important to focus retention efforts upon preserving profitability rather than preserving customers. This is the "Value Attrition" problem. To illustrate the difference, consider that if a customer who spends $10 per year leaves, you won't feel a financial loss. The customer was probably costing you money anyway. Conversely, if a $1 million relationship gets cut in half, you'll certainly feel the impact. To improve the bottom line, companies need to focus on the value lost as opposed to the number of customers lost.

Value Attrition
The better alternative to customer attrition models is Value Attrition. Value Attrition models predict customer value at risk several months into the future. For example, if we know which customers have the highest value at risk in three months, we can approach these customers with an effective retention offer. Even better, we can spend more for these offers, because we are not targeting every customer -- just those customers at risk for significantly reduced customer value.

It's Not So Easy
Creating value attrition models is tricky for several reasons. First, the most natural target variable is continuous, not binary. In other words, we are building a model to predict a decrease in dollar amounts, not whether or not a person is likely to terminate his or her relationship. This makes some kinds of models (e.g., neural networks, regression) better suited than other kinds of models (e.g., decision trees).

Second, customer value must be present in the data. Otherwise, we need to create a suitable derived variable, and use this variable as a surrogate target for model building. Fortunately, it is usually easier to get a good target variable with value attrition than it is with conventional attrition. With conventional attrition models, it can be challenging to define an attriter because the simplest measure of attrition, account closure, lags too far behind to be useful for model building. Said differently, a customer will reduce his or her activity or balances when attriting, and may not get around to formally terminating his or her relationship for months or even years. By then, it is too late for retention activities.

Time periods present another facet that requires attention. We need to know how far into the future the model should predict, in order to allow sufficient time for effective retention offers. Also, receipt of data is typically delayed for several weeks. When the data finally arrives, we want to predict attriters a month or so further into the future to give retention efforts sufficient time to be effective. For example, imagine that the attrition model is built using December through February data to predict the last half of March and April, where April is the most recent data available. Then the resulting model is applied to the most recent data, February through April, to predict future value attriters in the last half of May and June.

There are also subtleties in "rolling up" data from the appropriate time periods, and in creating appropriate held-out data for model validation. For example, we may want to average several months' worth of balances to smooth the data with respect to random balance fluctuations of account levels measured at the end of the month.

Another issue is how to correct for biases in the sample population. For example, suppose we send a test solicitation to customers with a predicted high value attrition score and also send a solicitation to a smaller random sample of customers. How can we take into account that these two groups are different populations when building a revised value attrition model? The answer, contained in work that won the 2000 Nobel Prize in Economics for James Heckman, allows us to use response data from both populations in a principled manner.

Campaign frequency also requires attention. Will the campaign be run only once, every month, or when triggered by an event? In the ideal CRM system, the infrastructure would be in place to immediately receive and act upon new data from any channel, including inbound calls. This is a difficult "plumbing" requirement for an enterprise IS system, and only recently has this problem been addressed. Achieving instantaneous all-channel customer knowledge satisfies the customers' expectations that the company has a unified view of them across all channels. It also permits models to be built with a shorter lead time, predicting less far into the future. Usually we get best performance by predicting value attrition as close to the present as possible, yet far enough into the future to permit effective retention efforts.

We hope these subtleties have engendered sufficient caution from the reader that he or she will turn a very skeptical eye toward claims of ready-made, out-of-the-box attrition solutions. It's not, "Hey, anybody can build a model!" It's more like, "Don't try this at home, kids!" Experience and know-how are indispensable for building value attrition models.

Finally, the big question: Is the retention campaign worth running at all, and how much do we expect to realize from the campaign? Answering this question can require a complex spreadsheet analysis.

The discipline of estimating campaign value is particularly useful in avoiding some common mistakes. For example, there was a case where a marketing department argued, "We need a retention campaign for these customers, because we spent a lot to acquire them." In this case, most of the customers in question were projected to lose money for this organization, so a classical retention effort would have been nothing less than throwing good money after bad.

Do You Care About the Bottom Line?
Value attrition is attrition for those who care about the bottom line. Although there are subtleties to building good models and running campaigns, the returns can be quite large with value attrition campaigns.

According to Gartner, predictive modeling (such as value attrition models) shows great promise for enabling banks to pinpoint real cross-sell and retention opportunities. Companies that fail to use information derived from predictive modeling at the touchpoint will face a competitive disadvantage from providers that do.

Case in point: Xchange compared the top 5 percent value attrition customers with standard attriters for DDA products at a large international bank. The results were striking. The amount at risk for attriters averaged $4,125, while the amount at risk for top Value Attriters was $20,498 -- nearly a five-fold difference!

Steve Gallant is director of analytic services at Xchange in Boston. Exchange Applications, Inc., now doing business as Xchange, Inc., helps companies focus their resources on customers who represent the best long-term profitability.

[ Return To The September 2001 Table Of Contents ]


Daniel Callahan

Canary In A Data Mine

BY DANIEL CALLAHAN


After mining through the mountains of information on this subject, here are a few nuggets of news.

Data integration is one of the most important issues facing enterprises today. The Gartner Group has estimated that IT organizations spend about $40 billion annually on integration software, and that the average enterprise has a median of 14 databases. IT professionals have realized the need for easier (and less expensive) access to information resources within the enterprise.

With the introduction of the Integration Server, Callixa has developed the first Enterprise Information Integration platform that helps companies create a standard data access layer in the enterprise. The solution allows companies to map data across multiple disparate data sources, and uses a common metadata-based data structure that supports any application in the enterprise. The Integration Server provides dynamic, bi-directional access to, and modification of, data located in multiple sources, enabling users to create unified views of enterprise data regardless of the data's native source.

Siebel Systems has selected the Fonix text-to-speech technology to voice-enable its Siebel eBusiness applications. This voice interface will allow Siebel eBusiness applications users to gain real-time access to corporate information from any telephone.

Sonexis has announced plans to deliver voice-driven solutions to enterprise customers utilizing the Jasmine Portal from Computer Associates. These new solutions offer enterprise customers a single, unified access point to corporate data. Through its Solution Services Group, Sonexis will be using its Show N Tel and ActiveCall software development environments to develop and deploy these Jasmine Portal solutions.

digiMine has announced its Enterprise Analytic Services, enabling organizations to manage, interpret, and act on high volumes of data from within specific business units or across the enterprise. The service includes a scalable, hosted data warehouse architecture, multiple business unit reporting, digiMine CRA (a marketing campaign analysis module), advanced data mining, View Builder (an intuitive analysis tool), digiMine Data Slurper (a software tool to automatically extract, encrypt, and deliver data to the digiMine data center), and professional consulting services.

On the consumer-centric, self-service end of the market, TroubleTree Software has announced its flagship product, TroubleTree, an interactive solution for inbound contact centers designed to answer common customer queries. By targeting the 80 percent of questions asked over and over, TroubleTree allows the customer to describe the problem step by step in an interactive manner. A decision tree is created by this series of diagnostic questions and responses, and may be scrolled backward to enter a different response at any time. Patent-pending technology utilizes a relational database, and TroubleTree is also customized to client products upon installation and integrates seamlessly into their Web site.

[ Return To The September 2001 Table Of Contents ]