Towards Better Credit Scores: How a Stale Industry Is Leaving Money on the Table

Credit scores have remained essentially unchanged for decades despite the emergence of Big Data and revolutionary improvements in machine learning—indeed, they are largely the same as they were before the widespread use of the internet. Not only do these technologies offer better ways of building credit scores, but they make it possible to tailor scores to specific loan types.

This failure to innovate is partly the consequence of regulatory concerns, which explains why companies like Experian have expanded their scoring using an opt-in model: Experian Boost™ allows you to include additional items in your credit score if you so choose, for example paying your rent, building in a defense against political complaints; the individual chose to use a new scoring method, having had its pros and cons fully explained to him (though whether those explanations were understood is another matter).  That said, FICO seems unwilling to explore models that require data beyond the credit report itself, almost as if they fear losing what is a virtual credit score monopoly by demonstrating effective alternatives.

FICO style credit scores focus on a handful of items:

1) Amount currently owed (also called credit utilization) – 35% of your total score. This uses the total amount you owe (at the time the lenders happen to report) as a percentage of all your outstanding credit available.

2) Payment history – 30%; This includes both payments made and payments missed; however, there are a great many entities that report to the credit bureaus only when you fail to pay. Your landlord, for example, likely reports only your missed payments.

3) Length of your ‘credit history’—15%; This is calculated using both the age of your longest open credit line and the average age of your credit lines. It is a stand-in for the borrower’s age, but luckily for FICO the regulators are not smart enough to realize it or are subtly winking at it knowing it favors the middle-aged and seniors. This is why paying off a credit card and canceling it can actually damage your score. Generally, you should make an effort to keep your longest-running credit line open, even if you do so by retaining only a tiny balance. If you do decide to close a credit line, close the youngest one you can unless there is some compelling reason not to such as its offering a superior interest rate, etc.

4) New credit – 10%; This considers both the number of new credit lines you have and the number of so-called “hard” credit inquiries that have been made—where hard refers to an inquiry initiated by the borrower for the purposes of taking out a new line of credit. (So, credit inquiries by prospective employers or for pre-approved credit offers do not count.)

5) Credit mix – 10%; Namely do you have both credit cards and installment loans.

So, what ideas are being left on the table?

1) Credit card payment habits

Some of you may wonder how this differs from payment history. FICO looks at whether you have defaulted on a payment. However, that is not the only information available when you look at the details of people’s credit balances and payment amounts. You can determine a lot by looking at how people pay off their cards.  There is a huge difference between someone who has a high credit card balance but pays that balance off every month and someone who is using a credit card as a long-term debt instrument, carrying over the same balance month after month. Given the high-interest rate on credit cards, the individual who uses one to borrow money long term, rather than just to facilitate transactions, is likely either financially illiterate or under real economic stress which makes them a bigger credit risk.  Transactors (i.e. people who pay off their cards in full each month) are much better credit risks than “revolvers” (i.e. people who pay off their balances slowly) even if they use the same percentage of their available credit.

2) Using net income to determine the debt-to-income ratio (indeed, even using the debt-to-income ratio as most credit scores ignore it entirely)

FICO has been reluctant to use any information that is not on the credit report; that is why Fannie Mae has developed its own credit score, the FMCA, which incorporates information Fannie can access but that FICO cannot, most importantly yearly income. Even here, however, Fannie has been reluctant to break with tradition: Net income, i.e. your take-home pay, is far more useful for making credit decisions than gross income (your pre-tax income), yet Fannie still uses gross income because MBS purchasers refuse to revise their prepayment models. When people buy mortgage-backed securities, they want to have a sense of how many borrowers are going to pay off their loans early. That is because if the security offers you a better interest rate than the prevailing one, you want people to keep paying you interest, which they will not do if they pay their loan off early. Prepayment models determine the likelihood of borrowers paying off their mortgages early.  Interestingly, institutional investors refuse to revise these prepayment models even though net income works better for this purpose as well. After all, money that goes to the IRS is not money you can use to pay off your debts—prematurely or on time: It was never ideal to make credit decisions or to model prepayments using gross income. Fannie only did so because it was easier than using net income: All you needed to do was ask for the borrower’s W-2 along with a few pay stubs and you were done. However, improvements in networking technology make it easy to share your checking accounts with Fannie and underwrite using take-home pay even though in the past this would have involved bringing at least a year’s worth of pay stubs back to the lender.

3) Purchase history

Some purchases are simply more responsible than others, esp. if those purchases are being made with long-term credit. For example, no one would think blowing borrowed money on clothing is as responsible as obtaining a mortgage to purchase a house. Credit card companies can easily look at the purchases someone is financing with his credit card and determine how necessary the expenditure was. A person making a necessary purchase, even if he is using his credit card rather than some lower-interest loan, is a better credit risk than someone making a frivolous one.

4) Scores tailored to certain loan types

Of course, as I mentioned above, some entities have built credit scores to suit their specific needs. I have already mentioned one example, Fannie Mae’s FMCA. That said, there is still room for more innovation here. There are also FICO scores that focus on auto loans. However, almost all of these specialized FICO scores confine themselves to information that is readily available on the credit report. Automotive loan-specific credit scores, for example, ignore the lack of a credit card history: They often throw information away that is considered irrelevant. Few branch out into other data sources. Credit card companies, ostensibly, have access to their own customers’ “payment habits,” as I discussed above, and could be doing more to extract information from their own customers’ payment habits.

However, there are other loan products for which new credit scores could be built. Originators of HELOC’s might care about how well the home in question is being maintained and the probability that the property contains a certain amount of equity. Kukun’s permit data and property condition score can give you insight into the condition of a borrower’s home. Furthermore, it is possible to create a tailored AVM that offers not only a value estimate but also a “probability of equity” that could allow HELOC originators to offer their borrowers substantially more credit while taking on less risk. In essence, you could cut down on risk if you built an AVM that, in addition to returning a value estimate, looks at the outstanding amount owed on the property and asks, What is the probability that the equity in the home covers this loan amount? In essence, you build a confidence interval on the fly that covers the price range above the outstanding loan amounts plus the amount of the HELOC itself. (If you are interested in a bespoke AVM product for your underwriting needs, I suggest you reach out to us by contacting me or Raf Howery on LinkedIn. You can also reach us through our website, mykukun.com)

5) Using machine learning

Regulators and consumers alike insist that credit scores be transparent: If someone has a bad score, FICO must be able to point to the exact item that is dragging it down. Machine learning techniques, which have performed miraculously in many domains, are often difficult to interpret: In a particular case, it can often be difficult to determine why the algorithm disliked a candidate.   While there has been some interesting work done on interpreting machine learning models, the industry is unlikely to abandon the simplicity of traditional logit models any time soon. (I highly recommend Christoph Molnar’s book Interpretable Machine Learning for a detailed treatment of the topic.) However, machine learning can still provide guidance to model developers: For example, you can analyze the tree structure of a boosting model to automate both variable selection and the bucketing of dummy variables. When building models, you may have to lump together certain things, like credit utilization or debt-to-income ratios, rather than treating them as continuous variables. There is no reason to waste precious man-hours using trial and error to determine how this bucketing should be done, or which variables should be investigated to add lift to your model when this process can be automated. A friend of mine used XGBoost to automate the bucketing of DTI ratios by analyzing the resulting tree structure, for example. One can use machine learning to streamline model development even if the old modeling techniques are still used to preserve interpretability.

Machine learning also opens the door to “supplementary models” which can be used to find hidden business opportunities by abandoning the requirement of monotonicity. Credit scores are expected to improve when positive factors increase and to decrease when negative factors become worse, a model requirement called monotonicity. (Strictly speaking, monotonicity allows the score to stay the same as something gets “better” as well.) However, machine learning models can find exceptions, islands within a range of variables, where certain combinations of “bad features” are actually better than some combination that is ostensibly better. Lenders can combine monotonic credit scores with non-monotonic machine learning models to find customers that are, in fact, lower credit risks than they appear at first glance.

New data sources and modeling techniques promise to revolutionize credit scoring: If only businesses will take the time to research these new possibilities and government grants them the freedom to do so. We here at Kukun are happy to produce custom credit products for you. Our real estate data and analytics can enhance your credit models in unexpected ways, and our team of economists have nearly as much experience with credit as we do with AVM’s.

Read more: Agent based models

Franklin Carroll

Franklin Carroll is Kukun's VP of Modelling and Analytics; After finishing his studies at the University of Chicago, he worked at Fannie Mae where he helped build Collateral Underwriter, roll out the Value Verify initiative, and specialized in collateral related and credit related policy.

Published August 19, 2023