David Raab Article Archive

In today’s tight-fisted business environment, grand visions take a back seat to immediate results. For many marketers, this means less interest in the indistinct (though real) benefits of enterprise-wide customer management and more interest in solutions yielding specific, discernable improvements. One result is increased demand for predictive modeling, a reliable way to improve the results of nearly any marketing process.

But use of modeling has traditionally been limited by the amount of help that marketers need to apply it. Systems experts are needed both to access corporate data for model-building and to distribute final scores to operations. Developing the models themselves usually requires a statistician. While automated modeling software has attempted to reduce the statisticians’ role, the results have been inconsistent and support from the systems group is still needed. Large companies face the additional challenge of processing the masses of detailed transaction data needed as model inputs.

Quadstone System (Quadstone, 617-753-7393, www.quadstone.com) makes a determined attempt at giving marketing departments as much control over modeling as possible, especially in very high volume situations. Its strategy is to target business analysts, who have greater database and statistical skills than most marketers but are still far from technical experts. It gives these analysts a graphical interface to extract, model and score enterprise data without understanding the technical details. And it lets them apply their own business knowledge to data analysis and preparation of derived variables, the key modeling tasks that have eluded effective automation.

Quadstone works by extracting data, transforming it, and storing it in a compressed format where it is used for analysis, model building and scoring. During initial implementation, the system is connected to corporate databases–a one-time job for the systems department, for security if no other reason. The business analyst then uses the TransactionHouse module to defines analytical data sets. TransactionHouse provides a spreadsheet-like interface to select data elements and specify transformations, calculated measures, segment codes, and other processing. Once the plan is complete, the system executes it to build the data set.

Analysis and modeling are performed in the DecisionHouse module. DecisionHouse provides both conventional query functions and advanced, interactive data visualization. The system builds several types of tree-based predictive models, propensity scores and profitability projections with a minimum of user intervention. Users can also define data-driven business rules to guide customer treatments. An optional module, ScoreHouse, provides enhanced score card functions.

Quadstone’s compressed data format allows it work quickly on the entire data set without using samples or aggregates. While performance depends on data volume and hardware, the vendor says suitably-configured systems generally return visualizations and queries in a few seconds, while scoring millions of customers takes a few minutes. Time to create the original data set depends largely on the time to read data from the source systems, but is generally kept under one hour. Quadstone makes extensive use of parallel processing when needed to meet performance targets.

Completed models, score card and business rules are deployed to production systems through ActionHouse. This can either generate SQL statements that perform the necessary calculations within the original database, or do the calculations in a Quadstone data set and send back the results. The first approach works well for simpler calculations or small data volumes; the latter is appropriate when complexity or size would overwhelm a SQL-based system. The system automatically reproduces all the data preparation steps used to build the original model, as well as the scoring formulas themselves. This avoids manual recoding, which is one of the major bottlenecks of the traditional modeling process.

ActionHouse currently generates scores in offline batch processes–typically overnight. This month, Quadstone is scheduled to release functions to allow real time scoring of single records as well. This is done by transforming the scoring formulas or rules into XML objects, which can in turn be converted into C, Java, COM objects, or whatever other format is required by an external system such as a Web site or call center. These systems can feed fresh data into the scoring object, or the object can query other systems for current information. This real-time scoring will make Quadstone somewhat competitive with systems like Norkom and Data Distilleries, although it does not provide the multi-step campaign management of these other products.

Like other vendors, Quadstone has begun to create pre-built adapters to integrate with common operational systems. An existing adapter lets the system read and write to Siebel’s analytical and operational tables, and is being extended to allow real-time Quadstone scoring within Siebel scripts. An adapter for the ATG Web server lets the two systems read each other’s real-time decision rules. Another adapter lets Crystal Reports read the Quadstone data store.

Quadstone itself runs on Sun Solaris, HP-UX or Windows NT/2000 servers. Users must currently run a Windows workstation with X-Server software, although a pure Java, browser-based interface should be delivered later this year. The system can connect directly to the Oracle, Teradata and Sybase relational databases, and via ODBC to SQL Server, DB2-UDB, flat files and other sources.

In addition to its basic modules, Quadstone offers packaged applications for specific purposes such as retention and risk management. Versions of these are tailored to different industries and specific touchpoint systems. The applications include standard data models, analysis and model templates, as well as background information and guidelines for project execution. They are supported by Quadstone’s in-house consulting group.

The three major Quadstone modules are sold as a unit. Prices start at $120,000 for up to one million customers and reach $1 million for twenty million customers. ScoreHouse, a lightweight analysis module called HouseVision, and the packaged applications are additional. The company was founded in Edinburgh, Scotland in 1995 and released the first version of DecisionHouse in 1996. It entered the U.S. market in 1999 and now has about seventy customers. Most are firms with very high data volumes and are concentrated in financial services, telecommunications and retail.

* * *

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.