David Raab Article Archive

Mainframe computers are like orga­nized religion: they come in and out of style but never really go away. Today, mainframes are back in fashion as compa­nies realize they are still the best tool for certain tasks–such as batch processing huge masses of data. For direct market­ers, this means that mainframes are often still the best way to generate statements and make complicated selec­tions from very large customer lists.

But even mainframe software con­tinues to change. Systems must present increasingly sophis­ti­cated selection strate­gies in a format that users can compre­hend. One solution is a graphical flow chart, which can easily show the over-all selection structure and still let users drill down on an indi­vidual segment to review its contents.

List Selection System (Syllogy Corpora­tion, 201-343-8900) was built for very large direct market­ing companies, where tens of millions of customers with hundreds of millions of transactions had to be segmented for multiple promotions simul­taneously. Syllogy’s solution was to process one customer at a time, first gathering all that customer’s data and then running it through the entire selection tree. This meant the system could work directly from existing production files, so long as these were sorted or indexed by customer, instead of spending time and money to merge re­cords from the production sys­tems into a separate database.

Syllogy combines the customer data by mapping each source file to a single “micro data­base”. Each source is accessed through a custom-written driver, so input can include flat files, rela­tion­al databases, and other formats. The data is present­ed to users in a “terms dictionary” that includes the actual fields, predefined functions such as mathemati­cal and logical operations, predefined actions such as selecting a record for output, and custom algorithms–typically written by the corpo­rate MIS staff–for standard concepts such as “paid customer”. Standard definitions help to ensure consistency among selec­tions by different people.

Once terms are defined, users can apply them to build the actual segmentation tree on an OS/2-based personal computer. LSS displays the tree as boxes connected by lines–similar to a typi­cal organization chart. Users can collapse a several boxes into their parent for an overview or open an individual box to define the selection logic and actions taken for selected records. The color of each box indicates whether records selected in that box are also passed on to later nodes in the tree. Boxes can also be linked directly to other boxes outside of the normal sequence. Once a segmenta­tion has been run, each box shows the number of records that passed into the box and the number selected. Sections of the tree can be copied or moved with a graphical cut-and-paste, and tree segments can be built separately and later combined.

Selections within a box can draw on any data in the micro database and can include complex rules written in the system’s own selection language or implemented by passing control to an external pro­gram. The system can limit the total number of customers selected for a segment, perform Nth or random selections, or distribute records into equal-sized subgroups. It can also load all customers belonging to a single household and evaluate these as a group. LSS can select the top 1,000 records based on a model score, although this takes two passes through the data. The system can output selected records to a new file or add new data–say, a model score–to either the original file or a new one.

After a tree is defined, LSS generates Assembly-language computer code that is executed on the mainframe. Assembly code runs very quickly, so speed usually depends on how much data is involved and how quickly it can be read from the source files. One client with 12 input files and 7,000 bytes per customer processes a 100-cell select at a rate of one million customers per hour. Clients wanting a quick count typically run tests against a sample file.

The system has very limited standard report­ing, although it can provide counts by segment and can generate samples of the records selected for each group. Users can take advantage of calculation and record-writing abilities to generate more sophisticated statistics that are then analyzed with third-party reporting or statistical tools.

LSS has one installation in addition to the client where the techniques were originally devel­oped. The system is now being marketed jointly with Group 1 Software and is priced at $100,000 for the software and basic installation including simple drivers. Complicated customization or drivers would be additional. The company is working on a version to generate C code that could run on non-mainframe systems but has not announced a release date.

TRIAD(tm) (Fair, Isaac, 800-999-2955) also uses a tree-based interface to segment sequentially-processed mainframe data. But TRIAD grew out of Fair, Isaac’s credit scoring business, and was origi­nally developed to allow score-based segmentation within operational processes such as preparing credit card statements. This meant that TRIAD segmenta­tions were executed as a record was processed in another system, not during an independent selection procedure. But the software can work either way, and some TRIAD systems now run independently against a separate customer database.

TRIAD installation begins with consulting and custom programming to determine how TRIAD will interact with existing systems and what data elements will be available for segmentations. Data elements can be individual fields, behavior scores or the result of complex calculations. Users can also define up to ten types of actions–such as sending a letter or raising a credit limit–that TRIAD can assign to selected records.

Once the system is set up, users build strategy trees on a PC running Microsoft Windows. The first step is to define the conditions (in terms of existing data elements) that will qualify records to enter the tree. The user then selects up to ten data elements to use within the tree. The tree building screen lists these elements down the left side, essen­tially assigning a horizontal band to each element. Starting at the top, the user defines the value ranges of the first element that will segment the incoming records; each range gets its own box. Each box then gets its own set of value ranges on the second row, generating more boxes; each of these boxes is as­signed its own ranges on the third row, and so on. The number of boxes on the lowest row can reach into the hundreds. Each final box can be assigned one or more actions, which can include both the generic action type and a specific code such as a telephone script ID.

Portions of the tree can be copied to help speed data entry, but the system cannot link boxes that derive from different origins. Users cannot select a set of non-sequential values (such as state codes) or use calculations based on the data, although these could be defined in advance through custom programming. There is no support for random or Nth selections within a tree or for limiting the num­ber of records assigned to a certain result. An “estimator” does count records that would be as­signed to each output once the tree is complete, and can help users manually adjust range boundaries to achieve specific quantities when needed. But the system does not provide samples of the selected records.

The completed tree is loaded onto the mainframe computer. During operation, TRIAD checks each record as it is processed in the host system, runs qualified records through the appropriate tree, and passes back the action codes. To help implement “champion/challenger” tests of marketing strategies, the system automatically assigns new accounts to one of 100 equal-sized groups and lets the user assign a different groups to different strategy trees. This ensures that strategies can be tested without accounts migrating from one group to the other over time.

TRIAD reports can count the actions gener­ated by each tree, estimate the profitability of an account portfolio, compare actual to expected score­card performance, and show the results of different strategies. The system can feed data to SAS for more advanced analysis.

TRIAD was introduced in 1991, based on a predecessor system from 1985. Over 100 firms use the system, either through in-house installations or at a service bureau. Versions exist for collections, credit line management, transaction authorization, fraud intercept, reissue decisions, pricing and market communications such as cross selling. These may run at posting, at billing or on demand as appropri­ate. Pricing is upward of $300,000 depending on the products and services purchased and includes on-going strategic consulting.

David M. Raab is a Principal at Raab Associates Inc., a consultancy specializing in marketing technology and analytics. He can be reached at draab@raabassociates.com.