Predictive Churn Modeling: Art, Science, or……Cooking?

By Michael Lowenstein

  Michael Lowenstein, CPCM, is managing director of Customer Retention Associates, a customer and staff loyalty program development, research, and consulting firm located in Collingswood, New Jersey (www.customerloyalty.org).

  With over thirty years’ management and consulting experience in customer and staff loyalty research, CRM, loyalty program development and refinement, customer win-back, service quality, customer-driven corporate culture, and strategic marketing and planning to draw on, he is an active speaker, workshop facilitator, and trainer, and he is a regular featured contributor to three customer loyalty newsletters.  His keynote, general session speaking, and workshop facilitation assignments have been in the United States and Canada, Europe, South America, and Africa.  He also provides expert customer loyalty commentary and articles for several professional CRM sites on the Internet.  

   Michael is the author of two widely-regarded books:  Customer Retention: Keeping Your Best Customers (1995), and The Customer Loyalty Pyramid (1997).  He is also co-author of Customer WinBack: How to Recapture Lost Customers – and Keep Them Loyal (2001).   Additionally, he is a contributing author to Redefining Consumer Affairs (Society of Consumer Affairs Professionals, 1995), The Answer Book for Customer Service Managers (Bureau of Business Practice/International Customer Service Association, 2000), and Customer.Community: Unleashing the Power of Your Customer Base (Jossey-Bass, 2002)

   He has been a customer loyalty instructor for Pennsylvania State University and the American Management Association; and he holds an M.B.A. degree in marketing from the University of Pittsburgh, and a B.S. degree in economics and marketing from Villanova University.  He is listed in several international, national, and professional Who’s Who directories.  His clients include First Union, Toyota, Prudential, Westvaco, Cigna, Charles Schwab, Borg-Warner, Sygma, Comcast, Baptist Health Care, Metropolitan Life, Microsoft, Alliance of Community Health Plans (ACHP), Daimler-Chrysler, and Georgia-Pacific.  

   Customer Retention Associates specializes in helping clients optimize customer loyalty and value through customer and staff loyalty research, loyalty program development and refinement, loyalty action training for front-line staff and management, and customer save and win-back protocol development.  The company is a founding member of the CRM International Consortium (CRMIC), an affiliation of independent CRM and customer loyalty practitioners from around the world, which is based in Europe.  The mission of CRMIC is to offer leading-edge customer loyalty and value solutions.

   In Gorgias, one of Plato’s lesser-known morality plays, he stages a debate (featuring Socrates as moderator) on the nature of rhetoric, or public speaking.  The questions Socrates poses:  Is public speaking art or science?  Is it a positive or negative force?  Through Socrates, Plato concludes:  “Not an art but a knack gained by experience.”  Those who take the best of art and science are like gourmet chefs.  They know what they are doing and why it works, and they can distinguish between good and bad results.

   Churn modeling is very much like the creative alchemy that is cooking.  When done well, it has a little bit of art, a lot of science, a dash of finesse, and even a pinch of  intuition.  With rates of customer defection reaching epidemic levels in industries like retail, travel, healthcare, and banking, predicting turnover has become significantly more important to business in recent years.  Having reviewed material on many churn models, in a multitude of industries, prepared by individual companies and specialized consulting organizations, some basic and some sophisticated, we’ve concluded that perhaps no industries have more predictive modeling going on than telecom and financial services.  This makes sense because of their high degree of customer risk and defection.

    Let’s begin with simple predictive churn modeling.  At Wachovia Bank, headquartered in Charlotte, North Carolina, they look at demographics, especially life events (divorce, losing jobs, opening a business, graduating kids, etc.), declines in account balance levels, and the like.  They also conduct research among high value customers who had defected to look for other root causes.  These results get factored into how they approach customers who are considered, as a result, to be at high risk for defection.  Another bank, PNC in Pittsburgh, depends very heavily on analytics which come from behavioral customer research, especially activity information regarding accounts at other institutions; and, like Wachovia, they also look at account balance levels.

    It’s always interesting to have the perspectives of experts in any field of endeavor, and this is particularly true in churn prediction. A couple of years ago, I interviewed Professor Adrian Payne, of the Cranfield University School of Management in the U.K.  He’s perhaps the planet’s most knowledgeable academic on the subject of retention and turnover modeling.  Professor Payne told me that, in his view, companies first have to look at cost of acquisition, build in retention spending, and other costs such as upsell and cross-sell.  He’s seen segments being built around behavioral-based market research (similar to what the banks are doing), eventually getting down to a microsegment level. 

   He also felt that few companies have sufficient customer information to develop really accurate models.  That’s where the alchemy, and some intuition, often comes in, because final decisions regarding potential churn may be left up to internal staff.  Confused?  Well, he added that companies should also look at competitive intensity, the industry retention average, overperformance and underperformance on service, and the like..

   Before we return to reasonable churn prediction approaches again, we’ll first get really complicated.  At a recent wireless  telecom customer loyalty conference in London, one European telecom’s information systems ‘expert’ presented sets of predictive churn models which had as many as 240 individual qualitative and quantitative variables, and up to seven streams, or sources, of customer data.  Model characteristics included:

1. Use of local calls
2. Frequency of calls to customer service
3. Number of successful/failed calls to customer service
4. Number of international calls
5. Single/multiple line subscription
6. Intensity of Internet usage
7. Calls to land-based vs. mobile phones
8. Use of automatic checking/savings account debit

   Principally, the model was set up to predict the level of potential turnover, i.e. the likelihood that customers would migrate to another service or simply abandon their current service.  He was using information sources such as customer service data, invoice information, satisfaction surveys, claims data, informal customer feedback, competitive data, socio-demographic information, and, as he stated:  “Any data that can increase one-to-one knowledge of the customers.”

Although the telecom’s representative offered some general results of model application to marketing programs, such as significantly reducing customer contact costs among certain segments, many of his findings were labeled as ‘confidential.’  This, to me, was mystical, black-box churn prediction alchemy in its purest sense.  He was, in effect, saying:  “Trust me.  I’m an excellent chef.  The food will taste wonderful.”  To paraphrase what a character in Dickens’ Great Expectations once said, without evidence there is no proof.

   Getting back to the real world of predicting churn, there are several professional firms that stand out in their approaches to modeling.  SLP Infoware, based in France, has a model (called Churn/CPS) which tracks multiple end-user defined churn behaviors, so that clients can engage and refine their retention strategies. 

   Their clients are almost exclusively in the telecom industry.  One of them, Cellular One of Puerto Rico, equips customer service agents with predicted at-risk customer behaviors, so they can apply any of several potential scripted marketing approaches to reduce churn. 

   Using the Churn/CPS model, they’ve been able to reduce customer turnover by one-third.  Since telecom churn averages about 30 percent in the U.S. and Europe, and over 50 percent in the Asia-Pacific region, this is really significant.  SLP figures that, with their model, something like 62 percent of churn behavior can be modified if spotted in advance, so this offers clients a fair amount of flexibility in how they approach at-risk customers.

   A second noted modeling firm is U.K.- based Quadstone.  Quadstone positions itself as a holistic predictor of customer behavior; and, through their Decisionhouse software, they offer marketers the opportunity to profile and segment customers “both visually and interactively”.  Like SLP Infoware, they have taken complex statistics and mathematical algorithms and converted them to hands-on application for marketers.  Quadstone has clients in retail, telecom, and banking, helping predict churn, determine which products/offers will encourage at-risk customers to stay, and identify what it will take to win-back lost customers. 

   In the telecom industry, they estimate they’ve been able to double response rates to add-on service campaigns, reduce customer churn by over 10 percent, and reduce the costs of churn management by half.

   The third innovative churn modeling consultant is @RISK, Inc., based in Pennsylvania.  @RISK, Inc. uses advanced techniques – neural network protocols, artificial intelligence, and causal inference algorithms – to detect patterns and trends in customers’ transactions which could mark them as potential defectors.  Their Pathfindertm program ‘learns’ the stable, causally associated indicators of defection from the transaction data itself, yielding better predictive accuracy and precision. This, in conjunction with systems that can produce unique prediction equations at the microsegment level, have enabled @RISK, Inc. to identify the vast majority of  “would-be” defectors months in advance. @RISK Inc.’s clients include fund companies, brokerage firms and banks, and they have used Pathfindertm to help identify high risk customers and alternative positioning approaches.  @RISK, Inc.’s techniques have proven both leading-edge and highly effective.

   One business writer, commenting on the emergence of churn prediction in the fund industry, said:  “As this technology advances, the marketing departments of fund companies will start to look decidedly different.  You may be hiring segment managers rather than product managers into your marketing group.  And sooner than you think.” 

This opinion can apply to any industry and business where excessive defection is a concern.  In other words, marketing will be seeking the gourmet chefs – the Emeril Lagasse’s – of the predictive churn modeling world, to help create the most appetizing  loyalty menus for their customers.