Breaking Bad Modelling

Nearly five years ago, the iconic series Breaking Bad aired its final episode. Since that time, I have been desperately searching for something new to binge-watch. The show chronicled the descent into darkness of a man named Walter White, who suddenly learns he is dying of lung cancer. Over the course of the first few seasons, he transforms from a mild-mannered high school chemistry teacher into a dangerous drug kingpin. As the series progresses, we learn that Walter is a genius in the field of chemistry and was formerly a university research professor whose groundbreaking work was stolen by his close colleague.

The series was very intense and often rather erratic. One moment, Walter would be ruthless and violent, and the next he would be a thoughtful intellectual. I found it absolutely hilarious when he would come unhinged at derelicts cooking methamphetamine and would correct them on their understanding of chemical compounding.

I can somewhat identify with this character, as the introduction of user-friendly, GUI-based modeling software has made it so that any back-alley numbers junkie can crank out a credit scoring model or other sophisticated-looking predictive tool. For analysts with little background in the field of statistical modeling, these programs can make them look and sound as if they know what they are doing. For others with advanced training, the software makes it all too easy to skip over some of the more mundane steps in the model development process.

Over my two decades in auto finance, I have hired and trained many quantitative analysts. A large part of my job has revolved around breaking bad. I don’t use that phrase in the sense of Walter White’s behavior, where it means to raise hell or to turn to crime. I am referring to breaking bad habits that frequently beset modelers, particularly in the area of credit scoring. With regard to scoring models, careless calibration won’t kill you, like it could in a meth lab – but it just might kill your portfolio. For this reason, managers and statisticians alike should pay attention to the most common mistakes analysts make when building credit scorecards. What follows are the five most common, presented not in order of frequency, but in order of their risk to the company.

Number Five: Bypassing Tedium

Only about 10 percent of the scorecard development process involves the use of highly sophisticated quantitative algorithms, i.e., the sexy stuff. The other 90 percent of the process is made up of data scrubbing, univariate analysis and a number of other tedious activities that, while extremely important to building sound models, are simply not a lot of fun.