Credit Scoring, An Advanced Tutorial

When we think about credit scoring we mostly think about the actual number and strategies to improve that number.  However, the topic of credit scoring is so much more complex and involved.  Understanding scoring in more depth will not only provide you valuable context on the subject but also arm you with more knowledge when you take your next trip to the bank.  Just think…you’ll know more about credit scoring than most of the lenders you work with.  We covered the fundamentals of credit scoring in this blog, now let’s take some bigger bites at the apple.  Let’s just say you’re not going to hear any of this listening to Dave Ramsey.

Odds to Score Relationship

Those numbers actually have a meaning.  I think most people just understand that a 750 is a good score, better than a 700.  But what does a 750 mean as compared to a 700?  Each of those numbers tells a story about your risk and that story is expressed as odds.  Odds, in a credit scoring discussion, are generally determined by studying and understanding the number of consumers who are going to pay their bills on time relative to the one consumer who will not.  This is an EXAMPLE of how the odds could change by FICO score range…

FICO 800 = 800 goods to every 1 bad

FICO 750 = 400 goods to every 1 bad

FICO 700 = 200 goods to every 1 bad

FICO 650 = 100 goods to every 1 bad

FICO 600 = 50 goods to every 1 bad

FICO 550 = 25 good to every 1 bad

FICO 500 = 12 goods to every 1 bad

Every bank knows how much they’re going to make on a “good” and how much they’ll lose on a “bad.”  This allows them to make an informed decision regarding whether to approve applicants who have scores that fall outside of their comfort zone.  Why extend credit to a FICO range where you lose money?  On the other hand, if you knew where the “break even” score range was you could easily draw a line in the sand and say, “we’ll approve everyone with a FICO score above X and decline everyone below it.”

Performance Definition

Every credit scoring model has what’s called a Performance Definition (hereinafter “PerfDef”).  It’s the stated design objective of the model.  So, for example, the PerfDef of the FICO risk score is “to predict the likelihood of a consumer going 90 days past due or worse in the 24 months after scoring.”  This is called an “Incident Model” because it’s predicting whether or not some incident will occur.

There are other models that have very different PerDef’s.  There are scoring models that have a financial based PerfDef.  For example, some scoring models will predict the likelihood that you’ll generate some amount of income or revenue.  These are more uncommon than incident models.

You can tune your model to predict almost any level of delinquency.  Your model can predict 30-day late payments, collections and even bankruptcy.  Equifax has a model called Bankruptcy Navigator Index (or “BNI”) and it predicts the likelihood of you filing bankruptcy.  And some bankruptcy models are even more complex and actually have an idea of how much you’ll cost lenders if you do file bankruptcy.  Awesome!


A validation, in terms of scoring, is the process whereby a lender or “user” of the scoring system studies and learns the expected performance of their consumers/customers by score range.  See the chart above under “Odds to Score Relationship.”  That’s how you learn those odds, by performing a validation.

So how do you know if your model has been “validated?”  Simplistically, all you have to do is see your “good to bad” odds improving by ascending score range.  In English, you should see more goods to every one bad as you get higher in the score ranges.   If that occurs then you know your model is “ranking” consumers by risk and therefore your model has been validated.  If your model isn’t ranking properly then it hasn’t been validated and it shouldn’t be used.

There are a variety of mathematical tests to determine if, and how well, your model is ranking.  The three more common tests are the Kolmogorov-Smirnov test (or “KS”), Receiver Operating Characteristic (or “ROC”), and Divergence.  Without getting too complicated let’s just say that each of these are used to measure a model’s ability to separate “good populations” from “bad populations.”  The stronger the separation, the better the model.

So, if FICO could put all future “bads” below 500 and all future “goods” above 800 then it would be a perfect model because you’d just deny everyone below 500 and approve everyone above 800.  There’s the end of defaults and collection agencies.  Obviously, there is no such thing as a perfect model.

That’s it for now.  I think my brain is about to explode.


John Ulzheimer is the President of Consumer Education at, the credit blogger for, and the author of the “credit rating” definition on Wikipedia.  He is an expert on credit reporting, credit scoring and identity theft. Formerly of FICO, Equifax and, John is the only recognized credit expert who actually comes from the credit industry.  He has served as a credit expert witness in more than 70 cases and has been qualified to testify in both Federal and State court on the topic of consumer credit.