After graduating in 1989, I found my way to Rose-Hulman,
where I spent the first quarter needlessly worrying about flunking out, and the remainder
letting the grades take care of themselves. My junior year abroad at the University
of Limerick in Ireland also turned out to be a wonderful experience, as did some summer
REU programs at UT-Knoxville and at FSU.
From Rose-Hulman I moved on to Georgia Tech, where I
picked up my Master’s Degree in Mathematics in 1995. After a couple more
years at Tech, I decided to instead enter the work force and see how I could use my
mathematical background in the “real world” (see below).
After seven or eight years in the workforce, I found myself
coming back full-circle, as the probabilisitic aspects of my work focused my attention
on the real beauty of the subject, and brought me back to Georgia Tech to finish my
Ph.D. in probability theory. Although I intend to pursue an academic career following
graduation (I love teaching), I will always maintain a strong connection to the non-academic
world as well.
As mentioned in the bio, after obtaining my MS, I decided
to move on to the “real world”. First, as a well-educated mathematical
student, I had to prove that it existed. It did. Having done that, uniqueness
followed easily (after slashing a few cosmologies with Occam’s Razor). The
hard part was finding what would actually interest me.
Initially, I started down the actuarial path, but no sooner
had I passed the basic prob and stats exam than I found a job at a small credit scoring
consulting firm in Atlanta, namely Scoring Solutions, Inc. My knowledge of that
industry was practically nil (0 < epsilon << 1) at that time.
But that is the point. Precisely because I had the
technical background in mathematics (and a little statistics), I could learn what I
needed about the industry on the job in order to do the statistical modeling the position
called for. The firm, which had been started by a couple women who had statistics
backgrounds themselves, understood that fully.
What did I actually do there? Well, the main purpose
of credit scoring is of course to produce a credit score, a number typically between
1 and 999, which characterizes how risky a potential customer is apt to be. Generically,
when people refer to “the credit score”, they are talking about
the FICO score, produced by the eponymous industry pioneers. Our firm was and
is a competitor with FICO.
That there is competition is understandable. Since
some banks or retailers might be targeting a particular segment of the population,
a general score might not be as helpful as a score created on the basis of their own
portfolios. That’s where a firm like ours could help.
Various analyses go into creating a credit score. First,
one must aggregate the data, which typically includes the applicant information
and, most importantly, data from at least one of the three credit bureaus (TransUnion,
Equifax, and Experian). Literally hundreds of descriptive variables for literally
tens of thousands of applicants are created, giving the analyst a data set often exceeding
a gigabyte in size. (Quite unlike your stats book examples, huh?!)
Using some statistical software – we used SAS, an
industry standard – the analyst next often does a segmentation analysis,
which looks for natural ways to break up the data into distinct groups for separate
consideration. Some of the tools used for this include cluster analysis, factor
analysis, and tree-based analyses, all topics one encounters in graduate statistical
Once the data segments are finalized, the analyst next
performs a regression analysis using the aggregated variables to model the
probability that a given account will go bad (the definition of “bad” having
been decided earlier). Now there’s more to life than just ordinary least-squares
regression! In particular, logistic regression makes more sense
when you’re modeling a probability, i.e., a number between 0 (“bad”)
and 1 (“good”). Once the appropriate handful of model variables have
been decided upon, each segment then has its own final model – its scorecard – based
upon the regression coefficients. A credit score is born.
Finally, one needs to determine how well the score separates
goods from bads, and to see how it will work in practice by producing forecasts, etc. All
of these steps require an understanding of mathematical and statistical concepts.
My own unique contribution to the firm, in my opinion,
lay in developing procedures (macros, etc.) which automated the more routine parts
of the model-building process, while still allowing adequate flexibility. (No
two projects are the same!) Even more satisfying was that I was able to become,
in effect, the R&D Department there, creating new procedures for novel challenges
as they arose.
What parts of my Rose-Hulman math education did I use? What
parts didn’t I use Prob and stats courses obviously were crucial, but
I also needed to recall linear algebra to understand the multivariate nature of the
data, discrete math to enumerate various data and algorithmic structures, and optimization
(linear-programming and calculus) and numerical analysis to find the optimal parameters OK,
I never used much group theory, though I did personally commute, like the other Abelians
I hope this gives you a better idea of the sorts of things
one might do with a math degree in the wider world, particular within the context of
credit risk. Good luck!
|Trevis Litherland at work