Predictive Modeling—You Mean Actuarial Wizardry?
By Shane Barnes, FCAS, Candidate Liaison Committee
As far as “hot topics” in insurance go, predictive modeling has been one of the dominating topics over the past decade. This set of analytical tools has transformed the way that many actuaries work and has ushered in a new era of rating products. Predictive modeling will continue to shape the industry and direct future analytical developments. A basic understanding of this increasingly important specialty is essential for all actuaries—even those currently practicing in non-pricing roles.
What is Predictive Modeling?
When the term “predictive modeling” is mentioned, some may think it is crazy actuarial wizardry. But I assure you, it’s far from wizardry. It is, in fact, a more sophisticated way of handling insurance data than performing simple one-way analyses.
The core definition of predictive modeling is that you are using past data to predict the probability of some future outcome. Actuaries have been performing predictive modeling exercises for decades, the difference from recent development being the statistical rigor around the analytics.
Working with insurance data poses several unique challenges. For example, only a small portion of the insured population in any given term experiences a claim, and when a claim does occur it tends to be large. A typical insurance dataset has a significant proportion of zero-dollar loss amounts with large spikes when losses do occur. Many traditional statistical methods, like simple linear regression, assume that data follows a Normal distribution. Insurance data, conversely, is far from the Normal, and benefits greatly from more advanced statistical methods.
In the actuarial world, predictive modeling has become almost synonymous with generalized linear models (GLMs). However, GLMs are only one aspect of predictive modeling. Other examples are statistical clustering (particularly with geographic variables), Classification and Regression Trees (C&RT) analysis, price optimization, or even actuarial rate or reserving indications.
Are You Sure it’s not Actuarial Wizardry?
Many people that I talk to are unsure about predictive modeling and how it can relate to their everyday work. Predictive modeling often brings them back to the heavy mathematical manipulation they needed to perform on exams. Is there complex math being performed within predictive modeling? Absolutely, but luckily there are a handful of statistical software packages that will perform the calculations (such as R or SAS). You will need to be able to understand the statistical tests, but these tests are similar to what you would have learned in your college statistics course. However, I think it is important to understand some of the math that goes into predictive modeling. This will allow you to better understand the assumptions and results.
Predictive modeling is just as much of an art as it is as a science. Building a predictive model and interpreting the results is not just a statistical exercise; rather, to be a successful modeler (and actuary), you should strive to understand the relationships within your model and relate them to real-world business problems.
How Do I Become Proficient in Predictive Modeling?
There are several things you can do to increase your knowledge of predictive modeling. GLMs are by far the most widely used form of predictive models in the actuarial world, so that is a good place to start. Start reading some of the popular texts that are referenced at the end of the article. It’s important to understand the theory underlying predictive modeling (plus you’ll need them for the Advanced Ratemaking exam).
Another valuable approach is to talk with colleagues who work on predictive modeling. When I was first learning how to build predictive models, I had a few mentors who showed me the ropes, and their help was invaluable. Additionally, the Ratemaking and Product Manager Seminar that the CAS hosts every spring is a great meeting for those interested in predictive modeling.
Of course, the best way to learn predictive modeling is to actually build predictive models such as GLMs and test some models. There is a plethora of real-world business problems that can be addressed using predictive modeling. Many people associate GLMs with massive projects and year-long timelines, but in reality, such projects are a small portion of predictive modeling work. Most of the models I build take only a few hours to create and help me understand some of the everyday challenges I face as an actuary.
What’s the down side?
In my opinion, the hardest part of building predictive models is collecting good data. A model is only as good as its data. Good data is critical to obtaining the correct answer.
When you build a model, regardless of your proficiency, have someone peer review your work. This sounds like a basic concept, but you will always learn something from a peer review. A peer review may provide you with new avenues for investigation. If you second-guess your work, ask someone for help. Even if you are having troubles making a relationship work, having someone look over your work can benefit the overall model.
You also need to thoroughly understand the assumptions underlying your model. What is your dependent variable (a.k.a. response variable) and what is your weight? For basic predictive models, these questions are straightforward, but not every model you build will be basic. Understanding the assumptions in the model will help you build the correct model. Unfortunately, a wrong assumption can lead to an unsatisfactory model.
Use intuition when building a model. My college stats professor once told me that if you can’t explain a model result to a non-math person, then either the relationship doesn’t exist or you need to refine your explanation. Insurance is all about relationships. If you cannot think of a logical reason why something should be in a model, talk with a colleague and decide if it is truly predictive, or is behaving as a proxy for another variable. Conversely, don’t over-complicate a model just for the sake of complexity. As actuaries, we can show vulnerability in our work product if we don’t clearly articulate and explain our analytics.
Lastly, challenge the status quo. If you are updating a model that someone else built, make sure you agree with their assumptions and their inputs. Always think about ways to improve the model. In order to advance the science we need to come up with better ways to perform and improve the analysis.
Predictive modeling represents the past, present, and future of actuarial science. Hopefully you are eager to incorporate it into your own work now that it looks less like actuarial wizardry and more like actuarial science.
I would like to acknowledge other members of the CLC committee who provided valuable feedback and thoughts in writing this article, Shira Jacobson, FCAS, and Dan Tevet, FCAS.