In “Predictive Modeling Pitfalls: When to Be Cautious,” published by Carrier Management last month, Ira Robbin from AIG presents a compelling case for exhibiting care when developing and implementing predictive models. His advice on using appropriate sample size and proper treatment of missing data, for example, identifies two very common issues for modelers today. The eight pitfalls he describes are critical examples of real-world problems with predictive modeling that can be avoided through rigorous adherence to modeling best practices.
Executive Summary
Predictive modeling challenges can also be opportunities for modelers, says actuary Bret Shroyer in response to the "Predictive Modeling Pitfalls" highlighted by Ira Robbin last month. In particular, he discusses how contributory data can be leveraged for commercial lines carriers.Robbin highlights the particular suitability of personal lines to predictive modeling with its “large number of policyholders and extensive, reliable information on their attributes.” There’s no arguing this point; the early entrants into the property/casualty predictive modeling space were all in the private passenger auto and homeowners lines of business.
Several of the pitfalls and limitations noted by Robbin can also be interpreted as opportunities for innovation today. In particular, he asserts of commercial lines that the “uniqueness of risks in these lines, the large number of relevant attributes and the relatively small number of such risks all pose challenges in extending predictive model pricing applications into the large risk and specialty markets.” Today, modelers are rising to that challenge.
