## Statistical Segmentation in the Wild

Some may recall from an introductory statistics course that to calculate the average of an average is a statistical error.  Yet, it’s done all the time

Let’s look to nature for some examples to illustrate this issue.

While the duck-billed platypus is a mammal, it is one of very few remaining primitive mammals, termed monotremes, that lay eggs. The baby platypus arrives 10 days after the egg is laid.  Another water loving mammal, the hippopotamus has a 243 day gestation period, while the elephant’s can stretch to 660 days.  Obviously, it would be an error to say, the average time for a mammal to be born is 335 days (10 days for platypus and 660 days for elephant).

What if used a weighted average: there are approximately 100,000 platypus (I promise that platypus is both singular and plural!) in the world and 50,000 elephants.  Therefore, the weighted average given the population of 150,000 combined mammals is 226 days.  Perhaps one is emboldened because 226 days is pretty close to the 243 day hippopotamus average.  But it’s nonsensical, right?!

We can all agree that neither mathematical result provides accurate information from which to make an informed decision, but I see this type of analysis done all the time.

Like the elephant, hippopotamus, and platypus, credit union member segments vary significantly. Consider the behavioral pattern of CD-only vs. indirect auto vs. engaged-checking members.  Even within checking account member segments, a digital member vs. traditional member vs. executive member are all quite different.  The revelations of how member’s choose to engage via products, services, and transactions can be viewed in statistical clusters, yet they are distinctly different.

• Dozens of intuitive actions are revealed as one reviews member engagement segments and the variations within segments. Engagement varies even with different checking account segments.  To view the number of debit card, ACH, monthly deposits, branch transactions, fee income and average balances results in easily targeted actions.
• Attrition models should be constructed within each segment. Clearly attrition-alerts for members within the traditionalist segment (engages via branch and check transactions) versus a digitalist (mobile logins, debit card and ACH transactions) will be based on different components.
• Wouldn’t you agree that the Next Best Action, the Next Best Product, and the Ideal Product/Transaction/Service Mix would vary by segment? And how to measure success of Actions without a quantitatively proven Engagement Score?
• Finally, the ultimate goal is to have dozens of targeted prescriptive-actions (statistically proven from past data and trials resulting in prescribed actions via measured-likelihood percentages). The action results will further impact segmentation results or suggest clear sub-segments for future actions.

In conclusion, engagement, next action, and next product offer should vary by segment. To use the average of all members, or average of all members with checking likely produces an offer or action that doesn’t meet the needs of any particular segment. The ideal-bundle of products, transactions and services is quite different for no engagement, deposit, loan, digitalists, traditionalists, and executives. To have a single action plan, algorithm, or model for all members makes as much sense as saying platypus to elephant average is proven because it’s close to the average for hippos.