Using Multiple Models for Databases
Predictive modeling has been a proven tool in the direct marketer’s toolbox for many years that uses information about a consumer to predict an outcome or metric of interest (such as response to a mailing or lifetime value). When desired, marketers utilize multiple predictive models to improve results above what a single predictive model could provide.
In some cases, the acquisition process involves several stages. For example, a credit card marketer sends a direct mail piece to a set of prospects. To become an active card member, a prospect must (1) respond to the mailing; (2) be approved for the product; and (3) activate and use the card.
In this example, there are three outcomes or behaviors required to get the active new account. A single predictive model that selects the most responsive prospects may produce a high response rate but a low approval rate. Or a single predictive model that selects prospects most likely to be approved may ultimately produce a low response rate. In either case, the marketer targets a set of prospects that ultimately produces few new active customers.
One way to address this problem is to build a single model that directly predicts active new customers from prospects. Essentially, this model combines all of the outcomes or behaviors (response, approval and usage, in this example) into a single outcome or behavior. Another approach is to build separate models for each outcome or behavior and then combine those models. Following the credit card marketing example, one model predicts response, another predicts approval and the third predicts usage.
So which approach is better? The answer (of course) depends on the situation.
In short, the multiple model approach allows more flexibility and is likely to produce better results. But it is more complicated than a single model approach.