Live From NCDM: Is Log Dead?
That would be log, as in logistic regression modeling, the mid-20th-century cousin to linear regression analysis. And the answer is yes, according to Doug Newell, president and CEO of Genalytics, Newburyport, MA. Newell throws linear regression and neural network analysis (“Try to get a vice president of marketing to stake his career on a model nobody understands,” he says, in dismissing the latter) into the same dustbin.
What killed them was an explosion of data collection. Storage and data capture have become cheaper, and model-processing times have been dramatically reduced. Alas, these analytic techniques are not able to scale up as well.
Why does this matter? After all, companies are still able to generate models that improve campaign performances using these techniques. According to Newell, however, the analysts writing these models are limiting their focus to relatively few variables, and are effectively throwing the rest of the data away. If a multi-million-effort marketer such as a credit card issuer uses a limited-variable model that is right 95% of the time, the 5% of the analysis that is incorrect represents money left on the table.
Furthermore, it takes time to write and test marketing model. Companies could simply hire more analysts, but PhD’s schooled in modeling techniques are at a premium right now.
As a result, most analytic activity is based on reporting the results of either in-field or past campaigns.
As one might suspect, Newell has a successor to these analytic techniques in mind. Newell advocates using a series of “genetic algorithms,” in which every variable is individually coded.
Much as Charles Darwin postulated in natural selection, when these models are run, certain variables will surface as being more highly predictive, while others will fall by the wayside. These variables are then isolated and run through models again, and after a few generations should yield the most powerful models.
The advantage of this system is that, like Darwin’s theory of evolution, it is agnostic – not influenced by outside design. It allows all available variables to be tested, rather than relying on a few selected by analysts with preconceived notions.
Newell offered his theories during a breakout session at the National Center for Database Marketing conference in San Francisco. The conference runs through Wednesday.
