Predictive Analytics (A Risky P & C Post)

So, a few weeks ago, I attended one of our carriers' annual sales meeting. This is a fun-filled afternoon of speakers discussing topics ranging from commercial automobile polices to surplus markets, loss ratios and liability umbrellas.

Oh yeah, and 5 minutes on life insurance.

Oddly enough, I always enjoy this meeting, primarily to put faces and handshakes with voicemails and intercompany emails. And this year, my curiosity was piqued by the introduction of a new (to me, anyway) term: Predictive Analytics (PA).

This is a risk-assessment tool that enables Property and Casualty companies to further refine the underwriting process. After the meeting, I spent a few minutes with the gentleman who had discussed the topic, and he agreed to put me in touch with one of his PA experts.

A few days later, I had the opportunity to spend about a half hour on the phone with Sam (not his real name - carriers are generally skittish about speaking on the record, which I can understand). In the event, Sam was forthright and interesting. Here's the low-down:

Predictive Analytics (aka "modeling") is used primarily on underwriting commercial (and sometimes other) risks. It really began in the 90's with personal auto policies; it's an extension of a concept called "risk segmentation" that's used in addition to more traditional categories.

Basically, PA delves more deeply into the financial and demographic data of a given risk (property or business). This goes beyond, by the way, just credit scores (which are the subject of some controversy in the industry). In commercial lines insurance, this could include information from the Bureau of Labor Statistics and even the Census Bureau.

Sam stressed that PA is useful in the aggregate, but (obviously) can't predict how an individual risk would behave; it's an indication of what's "likely" to happen, not what's "going" to happen. Which seems a lot like traditional underwriting (just because you have diabetes doesn't mean you're going to lose a limb). The difference is something called "univariate" versus "multivariate" analysis.

Yikes.

Univariate analysis is generally used in traditional underwriting: things like construction (steel vs wood), protection class (is it near a fire hydrant) and occupancy.  These are looked at individually and summed up.

Mulitvariate analysis uses these, but then adds in financial, demographic and other information and - most importantly - how all of these factors interact with and affect each other.

And then there's the "secret sauce:" each carrier has its own formula for determining what weight to give each of these factors and how they interrelate: what's the propensity for a loss to which this information leads you? This will differ from carrier to carrier. That's why, for example, Company A might say "no thanks, we're not writing that" and Company B might say "hey, we'll give you a great rate!"

Sam also stressed that these models have to be constantly updated, as data and relationships change over time with the change in a carrier’s book of business. The models are (as noted above) customized for each carrier, but there's a bit of a catch to that:

There's a limited pool of Subject Matter Experts available in this field, so each carrier's models will be similar but still variable based on each carrier’s history and philosophy.

Thanks, Sam!