DECISION TECHNOLOGIES IN DATABASE MARKETING: PART XI
by Gene M. Ferruzza, Senior VP, Decision Technologies

Model Recalibration (Living Model)

As discussed above, parametric and non-parametric modeling technologies are data-driven processes; the data are used to tune the parameters in the model. Today, the process of collecting and storing data is highly efficient. Any active database is likely to be changing at least monthly, if not daily, hourly, or even in real time. This means that the customer data at any given moment differ from the data that were used to develop the model -- as do the customers themselves (as a result of customer acquisition and attrition).

Do changes in data affect the model's performance? Yes--as do many other changes, such as communications with the customer, micro- and macro-economic effects, and changes in the company. The model was developed at one time, whereas the world is in constant change over time. Therefore, the performance of any model steadily decreases over time. There are two ways to deal with declining model performance.

The first (and in the past, the most popular) way to deal with declining performance has been to let the model decay. If the model's performance is monitored over time, the decay can be converted into loss of return on investment (ROI). When this loss in ROI is greater than the cost of redeveloping the model, the model is redeveloped, bringing its performance back to what ever level the current data can support.

The other approach is what I often call the "living model." By this approach, the original model is developed in the same manner as always, but, in addition, the processes that were followed during model development, including data sampling, data transformations and representation after EDA, and model tuning algorithms, are automated.

As new data populate the data mart, the automated processes are performed, usually with fewer man-hours than were needed for initial model development. The model recalibration schedule is set up to coordinate with the data mart's data refresh schedule. The result of this living-model process is a model that never decays in performance, and sometimes improves.

CONCLUSION

Data mining is most appropriately viewed as a rapidly developing discipline encompassing the full range of scholarship and methodologies involved in deriving knowledge from customer databases for use in database-marketing programs. Data mining encompasses both basic approaches, based on application of domain expertise, and complex approaches, based on the use of data-driven modeling. Data-mining can range from activities as basic as name and address matching to application of the most advanced modeling techniques to predict customer responses to marketing campaigns.

Data mining is a professional practice, as well, through which modelers provide expert services to business users, to assist them in developing and maintaining data marts, developing and executing marketing programs, and evaluating the effectiveness of campaigns. Professionals in this field select their approaches and techniques from an array of available methodologies, which they continually refine and develop through experience and research. But however sophisticated the methodologies they use, the aim of data-mining practitioners must be to provide the business user with clearly defined, actionable, application-specific knowledge, delivered in an automated, user-friendly environment.

Gene Ferruzza may be contacted at: gmf@cmsnet.com