Searching for Answers With Cluster Analysis

Is cluster analysis dead? Over the past handful of years, several developments have contributed to clustering’s fall from favor among elite marketers. The advent of the Internet as the first personalizable marketing medium, the development of robust data and content management applications, and the teachings of a vocal group of marketing thinkers have all coincided to influence the current conventional wisdom, which points to a one-to-one future, in which marketers reach individuals with uniquely tailored offers.

On the surface, the notion of putting people into discrete preference groups based on certain attributes, such as where they live, seems suspect. Even the names that Claritas’s PRIZM database assigns to its geoclusters seem hopelessly out of date, such as “Furs & Station Wagons” and “Blue Blood Estates.”

Seeking to reconcile the one-to-one and clustering disciplines, I turned this week to Pat Moore, an expert data analyst who has developed cutting-edge marketing techniques at places such as American Express and Vision Marketing. Moore is full of useful observations, which is, of course, the perfect trait for someone who gets paid to find answers.

Moore points out that individualized models are often not actionable because offers can’t be tailored for each customer. Clustering usually makes sense because of the limited number of available offers. If a marketer has 20 possible offers, optimizing for the individual doesn’t make much sense.

Though the limits on the number of potential offers certainly would seem to dictate a clustering methodology, the goal of one-to-one marketing can still be attained within these limited possibilities, according to Moore.

“One of the first principles of one-to-one marketing is to try and identify the customer’s need,” Moore notes, “and that is not addressed by clustering based on demographics.” Transactional data, whether it’s purchased from third-party sources or found internally within a marketer’s customer databases, is far more useful.

“Transactions reveal the needs of the customers,” Moore notes. “They need to buy within categories. They don’t reveal their needs through things like their age or geographic location. The goal is to identify those needs and then address them.” For example, Moore continues, two households may show “presence of children,” but if one tends to buy expensive diaper-changing tables and another buys secondhand, this difference in needs won’t show up with mere reliance on demographics.

Moore’s recent work has led him to conclude that clustering is far from dead. It’s just built on the wrong data. Clustering built on demographic databases seems much less predictive than clustering based on transactional data.

All transactional data is not predictive, however. To be useful, it must be fresh. Moore has seen transactional data from third-party providers age very quickly, and he can quantify this degradation. The largest third-party transactional databases, such as Abacus, take time to compile. By the time a new update is available, it can often be several weeks old. Marketers who are used to seeing their models capture more than 70 percent of responders in the top five deciles might only see such success in the first few weeks after acquiring the latest third-party data. Degradation quickly sets in, making predictive models built on this data steeply decline in effectiveness long before the next update.

As a provider of analytic services for several marketers, Moore sees some of his best success when clients enable him to use their internal transactional data. Often, marketers don’t even bother handing over deep profile data on their customers when asking outside analysts to find new prospects. This is a mistake. Marketers should at least include data that can be correlated to third-party data, such as total purchase amounts and number of purchases.

Moore’s prescriptions may seem obvious, in retrospect, but they are rarely followed. Clustering still works, but it works a lot better if you use the technique with fresh transactional data.

Related reading

tencent_emily-ma_featured-image
kenneth_ning_emarsys_featured-image
bounce-370x229
site search hp
<