Exams in Real Life: MAS-I
Have you ever been studying for an exam and thought, "Why am I learning this stuff? When am I ever going to use any of it?" If so, you are in luck! This is the first in a series of articles examining how content from CAS exams are used in real life. First up is Modern Actuarial Statistics-I (MAS-I).
The magic letters are PA. No, it's not Pennsylvania, but predictive analytics, which has been a pretty hot topic in the actuarial profession over the last several years — enough so that the Certified Specialist in Predictive Analytics (CSPA) credential was the first offered by The CAS Institute (iCAS).
One of the core components of predictive analytics is modeling. And, at half of the total weight for the latest MAS-I sitting (Section C – Extended Linear Models), predictive analytics is kind of a big deal. If you were to search through the CAS 2019 Spring Meeting program guide (as of late April), approximately 20% of the sessions offered discuss models or predictive analytics. In addition, the CAS Ratemaking Committee just released a paper titled "Predictive Models: A Practical Guide for Practitioners and Regulators." Modeling seems to be all over the place, but does that mean that MAS-I gives you the tools needed to construct a model? Let's try and build our own model and see which steps, if any, are mentioned on the MAS-I syllabus. Since generalized linear models (GLMs) are some of the most commonly accepted models by regulators (and therefore widely used), let's use one of those. It's also conveniently one of the learning objectives in the MAS-I syllabus.
Assuming we have a quality set of data, here is a list of a few possible considerations to make when building our model:
- Breaking up the data into training and testing datasets. We might consider using cross validation, especially for smaller amounts of data.
- Determining what we'll be modeling. Should we model frequency and severity? If so, perhaps the best way will be using a Poisson and gamma distribution, respectively. Or maybe we should use a Tweedie distribution to model pure premium instead.
- Selecting variables to include in the model. Should we transform any variables? Maybe use splines? Perhaps we should introduce interaction terms. If, on the other hand, we have too many variables, we might want to reduce the number through regularization and the use of either a lasso or ridge regression. We should watch for aliasing as we add or remove variables.
- Evaluating the model. We can compare different iterations of the model against each other using deviance or other tests, as well as a host of other statistics and metrics.
- Testing the model. We can use the model on our test dataset and determine whether we're overfitting the model.
Lucky for us, all of the terms in bold have knowledge statements on the MAS-I syllabus! This was a very high-level look at the process of creating a model — there are many additional steps to take and considerations to make if modeling in practice (and some of these, too, are still on the syllabus). MAS-I doesn't necessarily provide you with all of the nuts and bolts necessary to design a successful model. In fact, many of the things covered by this exam are done behind the scenes by the computer programs that have allowed models like GLMs to become prominent in the insurance industry. The strength of this exam is in the foundational knowledge behind the modeling process, as well as introducing candidates to the variety of tools and tests available for building and evaluating models. This knowledge will assist the analyst in determining the correct inputs for the modeling tools, as well as correctly interpreting the output from those tools. Deeper understanding will come with experience, as one gains intuition for the "art" of modeling, as well as while studying for Exam 8 (which will be discussed in a future newsletter).
Even though modeling and its supporting sections can take up a large part of this exam, not every analyst is required to build a model. There is another section of this exam which deals with a topic that may be more related to the average analyst's everyday work. According to the CAS Statement of Principles Regarding Property and Casualty Insurance Ratemaking: "Consideration should be given to past and prospective changes in claim costs, claim frequencies, exposures, expenses and premiums." Of course, we are talking about trends, one of the four main types of movements along with seasonality, cycles and random fluctuations in time series, which is the last section in the MAS-I syllabus. You may have additional concern with the other time series movements depending on the line of business you're working on. For example, seasonality might be more important for a commercial policy covering a retail store affected by holiday shopping periods. Or perhaps you might be analyzing data for personal lines that are influenced by colder weather, such as motorcycle or boat. The coverage on the MAS-I syllabus will give you a good background on some of the considerations behind analyzing data over time, which is likely to come up at some point in an analyst's career. You might even read more about the use of trending in ratemaking in the article discussing Exam 5 – make sure to check future newsletters so you don't miss it