Cracking the code on credit card, debit card, checking and savings account transaction data, so as to gain actionable insights into individual customer situations, sits at the very heart of bank product cross-sell success. Imagine having the ability to know, at an individual customer level, the occurrence of a major life event—one that practically calls out for, say, a brokerage service, a life insurance policy or a home equity loan.
With such knowledge at hand, a bank could shift its precision marketing capabilities into high gear. Suddenly, it could present cross-sell offers that are precise with respect to customer selection, the timing of the offer and the offer itself, ultimately driving increased customer profitability and share of wallet.
Clearly, a customer’s purchase history stands to reveal biographical information that is far more accurate, extensive and valuable when compared to traditional segmentation methods that rely on sparse, often outdated and unreliable demographic data. The latter tends to paint only in the broadest of brushstrokes.
A detailed panorama of a customer’s spending pattern, in contrast, could illustrate that person’s psychographic makeup along myriad dimensions. In theory, it could speak volumes about their lifestyle, interests, behaviors, aspirations and life events (both past and future), relaying timely knowledge of specific financial needs that would otherwise be all but impossible to ascertain.
The good news is that the science of transaction analytics has evolved dramatically in recent years, with an integrated set of technologies that is rapidly keeping pace. Yet, despite the many innovations, most financial institutions tend to operate under the false assumption that they have already realized the full range of benefits from their transaction analytics. In reality, most of them have just begun to experiment—by, for example, automating certain marketing responses using simple rules-based engines that monitor individual transactions, such as the deposit of an abnormally large sum of money.
Single transactions present a severely limited view—one that is akin to looking at individual pixels on a television screen. Little can be understood about the big picture at this fragmented level. Similarly, the context for understanding customer needs starts with the creation of multi-dimensional profiles, generated from transaction data over a period of time—preferably 12 to 18 months, in order to take seasonality effects into account.
The data collection process should begin with an unambiguous data specification. Pull small data samples before pulling full-project data sets, and be generous in the number of fields being captured. Seemingly uninteresting fields can turn out to contain crucial information.
Fields that are nearly always useful include: transaction type, transaction value, balance, POS type, merchant category code, merchant ZIP code, merchant description field, transaction date and transaction time. Transaction data should be used in conjunction with other available data, including masterfile snapshots.
Next comes transaction modeling, which can open the curtain on any number of hard-to-get-at customer characteristics. For example: Is the customer a frequent traveler? An armchair investor? A recent graduate? An aspiring artist? A weekend sailor? An avid movie-goer? A gourmet diner?
Separate from discerning customers’ lifestyle characteristics is a set of algorithms that detect their life events. By identifying new or deviating data patterns, transaction modeling can announce such things as the planning of a wedding, the arrival (or impending arrival) of a new baby, a retirement, an inheritance, a child about to enter college, a medical ailment, a home improvement, and so on. A cessation of previously recurring transactions can indicate such events as a job loss, college graduation and mortgage pay-off.
Understanding the customer by understanding the merchant
To learn about individual customers is to first understand some basic facts about the merchants they patronize. Accomplishing this feat, however, is nontrivial.
Consider the fact that the merchant ID field provides virtually no useful information in this regard. The Merchant Category Code (MCC) field also proves to be unreliable, misleading, frequently irrelevant and sometimes just dead wrong.
Among its shortcomings: The MCC makes no distinction between high-end and discount merchants, and the fixed set of categories are arbitrary from a marketing perspective. A ski resort may show up as RECREATIONAL SERVICES, for example. An online merchant may be classified as SPECIALTY RETAILER.
Far more useful than the MCC is the merchant description field. This field allows for identification of the merchant name—which, in turn, can unleash a whole host of useable data sources. Unfortunately, the merchant name is usually lodged inside a textual train wreck of store ID numbers and strings of seemingly random characters. Deciphering the merchant description fields, and transforming the results into actionable insights, requires sophisticated text analysis tools, as well as some old-fashioned detective work.
The first step is to extract the merchant name. At Fair Isaac, we have developed a set of algorithms to automatically remove the superfluous content and convert the abbreviations into meaningful text. Combining the results with the MCC, it then becomes possible to identify the real merchant name.
Next, we compile characteristics on as many merchants as possible from a variety of sources, including SEC filings, merchant Web sites, and other public and private data sources. We use latent semantic indexing, a technique that uses statistical algorithms to search for patterns in unstructured data to build an understanding of the relationships among key words. For example, we can correlate the appearance of the word adventure with customer interests like skydiving, scuba diving and Peruvian archeological digs, or the word youth with Ticketmaster transactions for bands like Hoobastank.
In associating attributes with merchants, a determination needs to be made as to whether the merchant contributes to a lifestyle of the customer. At Fair Isaac, we have developed a library of variables that query the merchant description and return a value reflecting how much the merchant contributes to a lifestyle.
And what if the merchant defies categorization? For example, what do you do with a Wal-Mart transaction? In most cases, the answer is: nothing. A Wal-Mart transaction doesn’t tell you much on its own, since you cannot see the specific purchase items. And because almost everyone shops at Wal-Mart, a transaction does little to distinguish one consumer from another.
A large part of transaction analysis, therefore, is determining which transactions are irrelevant. Fortunately, when a consumer has a particular hobby or passion, they tend to pursue that interest at specialty stores. For example, a golf aficionado is unlikely to buy their clubs at Wal-Mart.
It is transactions with the specialty merchants (e.g., PRO GOLF SHOP) that can provide the most valuable insight. One merchant may contribute to multiple lifestyles. For example, a subscription to Rolling Stone magazine would contribute to both the popular music and magazine reader attributes. The lifestyle attributor variables are coded and tested manually; human judgment provides the final validation.
Of course, predictive modeling is a well-honed practice at Fair Isaac. With respect to transaction analytics, we can automatically generate derived variables based on any campaign need—e.g., money spent on recreational sports in the last 15 days or money spent on home improvement purchases in the last month. Because timing is everything, we encourage our customers to update the transaction data, along with the related models and campaigns, as frequently as possible, using a high-speed processing engine (ours runs at over 50,000 transactions per second), such that each product offer has the highest probability of success.
Malcolm Forbes once made the observation that the best vision is insight. Today, the insights that enable financial service companies to envision which customers are most likely to purchase which products, and at what point in time, can be readily generated, thanks to a new breed of transaction analytics capabilities and next-generation technology tools.