Building Propensity Models

Recently, I read a promotional email from one of my favorite market research vendors. The email’s first message was about propensity models, one of my favorite tools. I wondered if they really have the sample sizes necessary to create accurate scoring; in my experience, it requires very large samples that most firms cannot afford or execute.

Next, the email announced Geo-Demographic Modeling as a new marketing resource from their organization. It just so happens that a couple of my professional colleagues spent years working to perfect this approach.

Sam Barton at Claritas and George Moore at National Decisions Systems (a company we acquired while I was at Equifax).   Both were extremely competent and well-educated practitioners. The targeting logic that both of these gentlemen employed was very exciting to me.  So, of course, I attempted to utilize their approach in my insurance marketing efforts, such as Medicare for AARP/Prudential, term life for JC Penney Life Insurance, and others.

The results were never satisfying to me. Part of the reason why is explained in an excellent white paper named Geodemographic Segmentation: Do Birds of a Feather Flock Together? (Nelson, Wake 2003).

I learned the hard way that even though I could gain a strategic view of my market potential using these cluster method schemes, I could not effectively execute one-to-one marketing with them. For the consumer-centric campaigns, we instead utilized to logistic regression with individual, household and address level variables.  Then we achieved success. Today, I do not necessarily have to use regression to achieve strong marketing ROI, but I still do not use cluster statistics to apply generalized geo-coded statistics to individuals.

My first exposure to analytics after graduate school was oriented more toward macroeconomics.   Texas Instruments instructed me to generate the global demand for our consumer electronics line by product.   Now that was a frustrating exercise. Next, Equifax allowed me to pursue more of a microeconomics approach. For example, we posed empirically provable questions like, “what impact does your credit record have on your credit card usage/payments, your claims history for automobile insurance, your response to a life insurance offer?”

Now after 35 years of working to create predictive models for marketing to individuals, I still find success to be elusive and difficult.   DDG has just concluded a 4-month exercise to identify the best Hispanic prospects for a final expense insurance policy.   It was hard work that required multiple iterations of modeling on top of an extensive EDA.

I still remember my first predictive modeling meeting.  The VP of Sales stated that the model did not work.  The head of data analytics displayed the lift curve and declared the model was excellent.  The VP of Sales then asked, “Then why doesn’t it beat the control audience?”  The answer my friends was blowing in the wind—a bad sample.

Data Science is critical to marketer’s success, and it requires the study and training of specialists in predictive analytics to do it.  But I must note that the real-world data projects and the academic classroom exercises are very different.   OJT is often the best “grad school” for those seeking to build successful models—beat the control and predict accurately versus produce pretty graphs or PowerPoint slides that are intuitively attractive but not statistically correct.