Predictive modeling

[ActSciMan]

At my internship, I hear people talking about doing modeling. I know they use things like general linear modeling and I have heard phrases like predictive modeling, which I vaguely understand from doing a time series class.

Can any one who actually does modeling for a P&C company tell me more about what you do? Do you learn what you need to learn in the exam process, e.g., Exam C - Actuarial Modeling? Or, is this something you learn on your own? Or, is it based on statistics classes you took in college? And, is it all general linear modeling, or are there other types of models you use? Do you actually understand the modeling stuff or are you just using some software that does it for you? I have heard a few people talking about wanting someone who actually understands it.

Thanks for any help

[Buck]

Exams? Yes -- Exam Part 8 though; GLM/Predictive modeling is part of requirements. This article:

http://www.casact.org/pubs/dpp/dpp04/04dpp1.pdf

[Vorian Atreides]

That paper Buck links to is a good start . . . it gives a good overview of the mechanics of performing GLM (and related multivariate) analysis. Note that it assumes a strong background in linear algebra.

But to your question about what "predictive modeling" is all about, it's essentially a means of developing a "formula" with which to classify a given risk.

Put another way, predictive modeling's goal isn't so much to determine how much to charge for insurance as to develop a "score" (like a credit score) which is then used to help with underwriting decisions--like whether or not to write that risk; and if to write it, what rate plan to use. (Note that this last underwriting "decision" is presented in an extremely simplified form from what's really done in practice.)

[Ron Weasley]

I work in a non traditional research area. The current exam syllabus can help you begin to scratch the possibilities of predictive modeling, but will not make you a proficient modeler. Accomplishing that will take much more work.

[Hero3128]

Usually predictive modeling departments don't even care whether you pass exams. I work in predictive modeling and they don't even give raises or exam support for the prelims.

Actually most of our time isn't even spent modeling, it's spent understanding and processing the data.

[ActSciMan]

Originally Posted by Ron Weasley

I work in a non traditional research area. The current exam syllabus can help you begin to scratch the possibilities of predictive modeling, but will not make you a proficient modeler. Accomplishing that will take much more work.

Okay, then what sort of things do I do to get better than the exams? A lot of work in what areas? What books am I reading, what topics am I looking into?

[Whiskey]

Originally Posted by ActSciMan

Okay, then what sort of things do I do to get better than the exams? A lot of work in what areas? What books am I reading, what topics am I looking into?

In my opinion, you don't get to be a good modeler by reading articles and books, but by actually working with real data and getting used to solving the problems that it presents.

It seems to this day that 90% of any "predictive modeling" project at a company is data preparation and cleansing. I know this is shifting as many companies are on their fifth or so generation of model on certain lines. But I still think the best skills to have for someone who really wants to get into predictive modeling are day skills. The ability to manipulate data using SAS, SQL and R. And scripting language you will need to know will vary greatly from company to company.

If you really want to learn the "Science" behind what the vast majority of P&C actuators are referring to when they say Predictive Modeling, then read up on GLMs. But like I said, the literature will only take you so far.

have fun.

[Vorian Atreides]

http://www.amazon.com/s/ref=nb_sb_ss...ns%2Caps%2C261

I have the one by de Jong & Heller. It's a pretty good book if you have a solid background in Linear Algebra. (And you're likely going to find the better books making this assumption as well.)

A course on the (mathematical) theory underlying multivariate statistics can be helpful as well.

[Vorian Atreides]

I also agree with Whiskey . . . you'll need to do actual work to get a better understanding of it all. However, understanding the theory can help you develop skills in better "predicting" what will happen if you make "small" changes to your model; and understanding how to interpret the actual changes when those "small" changes are done.

[Ron Weasley]

Originally Posted by ActSciMan

Okay, then what sort of things do I do to get better than the exams? A lot of work in what areas? What books am I reading, what topics am I looking into?

Whiskey is right about most of the time being spent with obtaining, cleaning, transforming and otherwise processing data. Also, somewhat like programming, there is no better way to learn it than by actually doing it, but you have to learn something before you can start doing, right

They syllabus materials do a pretty good job of introducing GLMs. The topics that have been essential for me to delve deeper into to be able to converse intelligently with my peers who have Masters and Ph.D. level degrees in statistics, industrial psychology, etc. include things like Principle Component Analysis, Clustering, and Logistic Regression. Subjects that you'll hear mentioned in actuarial circles, but still seem to be shrouded in mystery, are Neural Networks and Decision Trees.

Logisitic Regression merits it's on comment. The actuarial liturature covers logistic regression to the extent that you learn about the logit function being one of the possible links for GLMs. It doesn't give an indication of how deep and rich the history is for statisticions using logistic regression. Before GLMs were as well developed as they are today, statisticians developed a large number of ways to squeeze imperfect data into the logistic regression mold to create working models. Even though I don't care for many of these transformations, they are a strong working part of the working knowledge for many experienced statisticians, so being familiar with them is necessary from a practical communications standpoint.

For understanding decision trees and neural networks, the machine learning course offered on Coursera was very useful. The SAS course on "Predictive Modeling Using Logistic Regression" will help understand a typical modeling workflow using clustering and principal components analysis. I've found Extending the Linear Model with R: Generalized Linear, Mixed Effects and Nonparametric Regression Models by Julian Faraway to be useful to both reinforce GLMs and to broaden my horizons in related areas.

Hope that helps.