Analysis Paralysis

Although analysis tools help you target your efforts, they must tie into customer touch points.

You hear it all of the time: One-to-one marketing and customer service is key to customer acquisition and retention. Every day you absorb phrases such as “customer profiling” and “relationship management,” and every night the words “neural networks” and “tree analysis” echo through your head.

With so many retention tools on the market — each using various approaches to profiling — sorting through them all to find the solution that is best for you is a daunting task. An explanation of the major analysis methods, as well as how they tie into your customer touch points, can help you get started on the right path.

Your situation: You want to run a campaign to reach your most valued customers — female white-collar workers living on the East Coast. A variety of customer-analysis approaches show how to find and retain these people effectively, as well as to market, cross-sell and up-sell to them.

Segmentation

According to Sam Koslowsky, Harte-Hanks vice president of modeling solutions and CRM analytics (www7.harte-hanks.com), the most basic analysis method can be done manually and requires separating a group of people into what he calls the haves and the have nots.

“In the wireless world, the haves might be those who are valuable wireless customers vs. the have nots who are less valuable,” he said.

For instance, you know your female white-collar customers on the East Coast are your most valuable wireless customers, while male blue-collar workers living in the Midwest are less valuable. In this case, you simply segment the group to find the people who meet the desired profile by doing a search by gender, then occupation, then location.

Dave Howlett, Athene (www.athenesoft.com) vice president of marketing, said segmentation entails using anywhere from two to 10 variables that you know about your customer base, finding a correlation among those variables, then grouping like individuals together and marketing to them in a more efficient way. Although segmentation offers some efficiency, it is limited in accuracy and broad in scope.

“It is really not predictive at all in terms of understanding customers' wants and needs; it just says that because you and I drive the same car and are close in age that we might behave the same way, and, therefore, whenever I accept a new product, there is a higher probability that you will take it as well,” he said.

Koslowsky agreed that this method does not present a full picture because instead of narrowing down the search with each new variable, segmentation looks at one variable at a time. In this example, all the females could hold white-collar jobs, so instead of narrowing down the search, you are finding the same people twice. However, the upside to this simplest form of segmenting is that anyone can do it, and no special software or statisticians are needed.

Profiling

Howlett describes profiling as segmentation on steroids. Good for small data sets, profiling looks at 11 to 20 variables and finds correlations among them, offering more granularity than segmentation allows. Profiling analyzes usage patterns, such as how often your targeted women are using their wireless phones. However, it is intuitive based and is limited in the number of variables it can manage. Like segmentation, profiling is not predictive, Howlett said.

Classic Statistical Approaches

Classic statistical approaches are multivariant approaches that allow you to be more discriminating. Variables fall into two categories: categorical and quantitative.

“Categorical is: I have 50 states, and I might find a better relationship between location and churn if I reduce that to three regions, but quantitative would be minutes of use,” said Jerome Nadel, SLP InfoWare vice president of worldwide marketing (www.slpinfoware.com).

In this approach, correlation breaks a group into core categories, and software will tell you what those categories should be. Perhaps you might want to break the group into equal age brackets. The software's statistical engine, on the other hand, finds more effective breakdowns. For instance, the engine might break it down as: 15 and younger, 16 to 19, 20 to 36, and 37 to 39.

Tree Analysis

Like doing a search on the Internet for Web sites with the word “female,” then narrowing the search with the phrases “white collar” and “East Coast,” a tree analysis searches the database for females first. From that group, it selects white-collar workers. Then, from that segment, it picks out people living on the East Coast. Unlike segmentation, which looks at one characteristic at a time, tree analysis looks for multiple variables simultaneously, Koslowsky said.

Collaborative Filtering

Howlett said that the Internet brought about the advent of collaborative filtering, a predictive technology suited to larger data sets. For example, when someone purchases a book from Amazon.com, the site recommends a list of 10 books that person might enjoy based on the purchase. As the user purchases more books, the technology becomes smarter.

The problem with collaborative filtering is that it can steer in the wrong direction if, for instance, someone buys a gift. Then, it begins to serve up recommendations that are relevant to the person for whom the user bought the gift. One of your female customers may purchase a book about business strategies over her wireless device because that is her main interest. But during the holidays, she buys several books for children, nieces and nephews. Suddenly, her recommendations switch from Atlas Shrugged to Rugrats videos.

Regression Analysis

Regression analysis uses software to search through available characteristics and identifies the most important variables, Koslowsky said. You may choose eight predictors and then assign a value, or weight, to each predictor. If the fact that a customer is female is more important to you than the fact that she is a white-collar worker, you would grant gender a weight of .5 and occupation a weight of only .255. Regression analysis then multiplies each characteristic by its weight to come up with a final score. The higher the score, the more you look like the female Easterner who might work for a Fortune 500 company and drive a $60,000 car.

“Regression gives each a different score, so it is more refined,” Koslowsky said. “You get smaller groups, and that performs better.”

Machine Learning

Howlett said machine learning, developed for very large data sets where relationships are complex and changing on the fly, is the most sophisticated prediction technology available. Using multiple statistical methods to analyze large data sets, machine learning makes predictions based on analysis of the inter-relationships of many variables.

“Machine learning is a whole family of non-linear mathematical models that allows you to predict with a high degree of accuracy what that person may look like or how they might behave,” he said.

In the wireless realm, machine learning, which includes neural networks, looks at hundreds of variables, including billing, usage and customer-service data, to capture information on every interaction and transaction that goes through that wireless carrier. Not only does machine learning look at the relationships between those 200 to 300 variables, it looks at second- or third-order effects, combinations of variables and how each variable might affect behavior.

Machine learning uncovers some trends and patterns that typically are not viewable by the human eye but are subtly manifested in the data. Therefore, you can predict more accurately how your female white-collar East-Coast dweller may behave and what problems and issues she will have in the future compared to other professional women.

Koslowsky added that the accuracy of neural networks — a type of machine learning — improves with each model. For instance, a neural-network model studies the data, finds a complex pattern, then makes a prediction. If the prediction was incorrect, it takes that data and learns from it, then makes another prediction.

Interestingly enough, although neural networks are more complex than regression analysis, the results frequently are the same. Regression is easier for marketers to understand because it places the scores in silos and explains the relationships between variables, he said.

Cluster Analysis

Koslowsky suggested that cluster analysis is the most sophisticated analysis method. Rather than other methods where you know the groups that you want to target up-front, clustering requires you to present the data to the software, which analyzes it to identify natural groupings within the data. One subgroup that the software might identify for you within your target market of professional females might be well-educated females who are under 35 and use wireless devices only during the day. Another might be females who only use wireless for emergencies. Identifying these groups enables you to market to them differently and more effectively.

“When cluster analysis works, it is really beautiful because some of the groups it finds, you never would have thought of,” he said.

---

Betsy Harter (betsyharter@aol.com) is a freelance writer based in Athens, GA.

Analysis Is Not Enough

All of the customer-analysis technologies in the world can't help you if you don't know what to do with all of the valuable data you collect, said Jerome Nadel, SLP InfoWare (www.slpinfoware.com) vice president of worldwide marketing. You must package that predictive model in a way that helps you use it in designing campaigns, as well as integrate it into the customer touch points for real-time support.

“You may build a powerful model and have tremendous success in the campaign because you retain a major percentage of people who would have left,” he said, “That does not necessarily make an effective campaign because you might offer people with lower value some incentives that are more costly than they warrant.”

He explained that SLP Infoware collects and analyzes data, then passes the results to all customer touch points, including an SMS engine, a Web interaction engine or a call center. A carrier using a call-center solution from Siebel Systems (www.siebel.com) or a Web interaction engine from Netonomy (www.netonomy.com), for example, could take the campaign to the next level. Netonomy, which enables wireless customers to manage their accounts from a carrier's Web site, then becomes more intelligent based on the analyzed data. As customers log into a site that is powered by Netonomy, they are scored with current mechanisms based on their behaviors. Then, Netonomy automatically presents offers that are specifically targeted to them. In the call center, on the other hand, a caller is scored immediately, and the Siebel or other software would present appropriate up-sell or cross-sell offers to the CSR.

For instance, Belgium's Mobistar (www.mobistar.be) uses both Netonomy and SLP Infoware, said John Hughes, Netonomy co-founder & executive vice president of business development.

“SLP gives us information on who is likely to churn, and we profile them and then offer special things to these users,” Hughes said. “We can pro-actively go out there and offer high-value customers something and keep them before they churn.”

Andrew Cole, head consultant of Adventis' (www.adventis.com) wireless practice, said such information is crucial not only to understanding the needs, wants and desires of your customers, but also to understanding their dislikes.

“Tailoring (customized) services to users means better stickiness, less churn and additional revenue streams,” Cole said.