The NCAA men’s basketball tournament is one of the most engaging sporting events in the world. The American Gaming Association says that Americans will fill out almost 150 million brackets this year. I know they are not aware of any of my family’s brackets – none of us bet - so my guess is the number is two or three times that. I play several brackets every year using a variety of approaches. The brackets are defined by seeds giving everyone “expert insight” into which team should win each game. A team’s seed places it in one of four quadrants of the brackets and are based on season records, strength of schedule, and success in the conference playoffs. Lower seeds are better, 1 being best and 16 being the lowest. But here’s the thing – Seeds are not predictions of tournament outcomes; they are a reflection of how a team performed in the past. Brackets predict outcomes.
Sound familiar? If you’re a credit analyst or involved in underwriting, you might be thinking about credit scores and how they influence predictions of payment delinquency. In the same way a 15 seed can beat a 2 seed in the NCAA tournament, e.g., Saint Peter’s vs Kentucky this year, a customer with a sub-500 FICO score can make all their payments over a five-year lease. I wonder how many underwriters fill out their brackets strictly according to the seeds? The Final Four has had all #1 seeds only once in 36 years, so it’s statistically not the best way to fill out a winning bracket. But underwriters are smart, so just as seeds alone are not the best way to do brackets, credit scores alone is not the way most predict delinquency.
|Game||Seeds||Winning FICO Score|
|Iowa vs. Richmond||5 vs 12||743 vs 557|
|St. Peters’ vs Kentucky||15 vs 2||477 vs 823|
|UConn vs New Mexico St.||5 vs 12||743 vs 557|
|Notre Dame vs Alabama||11 vs 6||583 vs 717|
|Tennessee vs Michigan||3 vs 11||797 vs 583|
“Credit Score Bracketology” NCAA Men’s Basketball 2022
(1-16 Seeds as 450-850 FICO. Getting into the tournament counts for something!)
Everyone who fills out a bracket considers the seeds, but then they bring their own “special sauce” to predicting winners. My father-in-law played in the Big Ten, so my mother-in-law always picks Big Ten winners. I have a sister-in-law who believes there are dominant colors – red always beats blue. Some like to consider cage matches between mascots – not sure if Blue Devils (Duke) beat up on Bulldogs (Gonzaga), but there are those who believe they know. My approach involves asking experts – my son and wife watch almost every NCAA game televised - and then combine their advice with my numbers method -- 12 seeds upset 5 seeds. Brackets – game predictions – involve data sources not considered by the NCAA Seeding Committee.
I found a great article on alternative data sources for brackets in Smithsonian Magazine. Professor Tim Chartier, an applied mathematics and computer science professor at Davidson College, motivates a team of students to try new mathematical approaches, new data sources, to bracket predictions each year. By machine learning standards many of the variables considered - points scored, margin of victory, early season vs late season wins – used in a weighted scoring method is quite simple. But inclusion of the right variables in a simple model can make a difference. One student reached the 99.9th percentile of ESPN brackets when she factored her data by home and away games. Makes sense to me that a team the performs well on away games will do better than one that needs a home crowd.
Similarly, lessors define “scorecards” that add various customer characteristics to credit bureau scores. They want a method that works better than others: prices better, avoids losses better, makes decisions easier and quicker. We have seen underwriting scorecards that include special insights like home ownership, is the owner a founder or 2nd generation, is the business located in a flood zone, distance from lender, and “provided Harley Davidson as collateral.” Everyone has their own methods of predicting customer behavior. That’s the goal. Like tournament seeds, credit scores reflect past performance and provide insight over the long term that will be statistically relevant. But just like every game has new circumstances and the opportunity for special efforts by players and coaches, every business year has a plethora of new opportunities, constraints, and circumstances that might change how a business performs and behaves. Extra data sources help make better predictions.
Now let’s consider adding the tools of machine learning and artificial intelligence. Not surprisingly, there are a wealth of academic papers on using ML and AI tools to try to predict the perfect bracket. The odds of a perfect bracket, by the way, when one uses simple random selection for each game, are 1 in 263, or about 1 in 9.2 quintillion. That’s effectively zero. But no one just uses random tools. The challenge for NCAA Bracket AI is that reliable new data streams that cover the key inputs are hard to find. How does one gather data on all the injuries, rivalry motivations, player match ups, and pre-game speeches necessary to build a predictive model?
Fortunately, finding data sources for building delinquency and funding predictors in leasing is not as complicated as deciding basketball game outcomes. The front end predictors we train at Tamarack start with 35 standard variables and can expand as needed for specific business niches or equipment types. Often during training, the importance of some variables fades below relevant, i.e., the model is equally accurate with and without that particular variable. But the machine learning tools, the prediction machines that we build, have no trouble with this kind of complexity. When we see “outliers”, outcomes that do not align with the predictions by a large margin, we often find new variables or circumstances. Each of these events presents the opportunity to include new data streams to improve the applicability and/or accuracy of the model.
So, if you are practicing “bracketology” this season consider how you make your decisions and how you try to capture an edge. Think about how you convert seeds to bracket predictions each time you circle a winner. Then think about how that process could inform your credit and underwriting. What might you add to your prediction machine?
We are not doing NCAA Bracket predictors, but if better underwriting predictions seem like a possibility, give us a call.