David Raab Article Archive

Some people have always considered statistical modeling to be the heart of database marketing. For traditional direct marketers in particular, the main purpose of building a database has indeed been to create models that squeeze additional profit out of existing customer and prospect files.

But direct marketers are a special case–their basic customer and prospect lists have always been readily accessible. In industries like financial services, travel, telecommunications and retail, a marketing database often provided marketers with their first opportunity to address a unified customer list. This let them build their first cross-sell, retention, profit-measurement and loyalty programs, which could be quite effective using manual segmentation schemes. For these marketers–now the bulk of the industry–modeling was a subtle refinement, offering incremental improvements that might not even be worth the considerable added effort.

This has led some observers to wonder whether modeling is anywhere near as important, or as common, as the attention paid to it suggests. But even as the question is raised, modeling is assuming an expanded role.

One reason is the very success of the first-generation database marketing efforts that did not rely on modeling for their value. With these in place, marketers now have the unified databases, messaging mechanisms and measurement systems needed to support more elaborate projects. These aim to optimize results by coordinating formerly-independent marketing efforts. This type of optimization relies heavily on statistical modeling to identify the appropriate balances among product and promotion variables. Concrete examples are projects to allocate marketing contacts in the pharmaceutical industry and to balance response, revenue and risk in credit card promotions.

The second factor is everybody’s favorite change agent, the Internet. All that delightful personalized interaction must be automated to be delivered quickly and cheaply enough to be practical. This takes lots of modeling.

Of course, the supporting technology must be available as well. Systems to build large numbers of models with a relatively low amount of user time and skill are an obvious requirement. Such products, often based on neural nets, decision trees and increasingly on multiple modeling techniques, have become almost annoyingly common over the past few years. New ones continue to appear, some of which may even be genuinely better than their predecessors. But even without new products, no one can complain that a lack of adequate model-building software has hampered their efforts in recent years.

However, deployment of large numbers of models requires more than tools to build them. It requires an infrastructure to keep track of which models are available, where they can be used, how they are performing, and when they need to be reconstructed. Perhaps most of all, it needs an efficient way to integrate models with marketing campaigns, so scores can be used as easily as conventional data elements.

CampaignOne (Nestor, Inc. 401-331-9640, www.nestor.com) combines a campaign manager with a set of model management tools. The system also includes auxiliary functions needed to make it a complete database marketing system–including a scripting language to import data and export results, selection tools to define the records that will be used in a campaign, a database to hold available marketing communications, templates for evaluation reporting, and a job scheduler to run everything automatically.

Like most of today’s newer products, CampaignOne does not include database construction tools. Instead, it works on the assumption that a database has already been assembled, either in a corporate data warehouse or specialized marketing system. Users would write scripts to load records into CampaignOne’s own database, which runs on the Microsoft SQL Server relational database. Scripts can call modules written in any programming language, allowing CampaignOne to work with any type of data sources and destinations.

The CampaignOne database can either hold lists imported for single promotions, or be updated over time as a permanent marketing database. Either way, selections for specific campaigns are defined by specifying the characteristics of the records to be included. Users also define the available communications by specifying the format and contents of the output file, a script to create and distribute the output file, and cost per record selected and per response received.

Campaigns themselves are defined in tree-like structures that segment the group of selected records. Splits can be based on data in the records, percentage distributions, absolute quantities, or participation in previous campaigns. There can be as many levels of splits as the user desires, and the system automatically eliminates records selected in one split from further consideration. However, the splitting rules cannot include user-defined calculations or some other complex relationships. Nor does the system allow the user to define time-based relationships, such as executing one node three weeks after another.

Communications are assigned to the final node on each branch. The system allows multiple communications per node, and can allocate records based on a percentage split or by setting a maximum quantity per communication. Names exceeding the specified maximum for a communication can either be discarded or made available to the next communication.

This is a fairly respectable set of campaign management capabilities, although not quite on par with today’s most powerful systems. Still, at a relatively low price of $275,000 regardless of file size or number of users, CampaignOne would be a reasonable option for many firms based on these functions alone.

Of course, the real interest of the system lies in its model management tools. Nestor, whose primary business is neural network model development and software, has defined a standard protocol that lets users import models built by any system. The protocol captures how to execute the model, how to convert its outputs to a standard scale, and what data was originally used to construct it. This latter information lets reports compare a later data set to the original inputs, so users can judge whether they are similar enough for the model to be valid. Each time a model is run, CampaignOne also stores the distribution of model results and expected performance, allowing it to highlight output changes and performance deterioration over time. This lets users identify models that are in need of redevelopment–an essential capability when many models are operating continuously.

Somewhat ironically, CampaignOne model execution is not well integrated with the campaign structures themselves. Models must be applied to the entire campaign audience before the segmentation is executed. This creates unnecessary work when a score is needed only for a particular subsegment. The problem can be significant when many models are in use and campaigns are executed frequently. While it can be avoided by breaking campaign segments into separate campaigns, this approach has its own drawbacks.

CampaignOne was released in October 1998. The system is initially being targeted at the credit card industry, where Nestor’s other products are most widely used. The $275,000 license, with 20% annual maintenance in later years, also includes two Nestor neural net models for marketing applications of the user’s choice.

The first installation is expected some time in 1999.

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.