Disclaimer: We conducted this experiment purely for scientific curiosity. We have no intent to duplicate the FICO score or commercialize any parts of our findings.
FICO score is probably the widely know credit risk score that everyone in the industry uses. Whether it is used to grant credit, approval for a rental agreement, there is no avoiding the infamous credit score.
Sure, there has been challengers to FICO throughout the years. Even the credit bureaus have banded together to unseat the champ. But FICO remains the undisputed champion of banking and lending industry.
In other countries such as China, credit scoring and national credit bureaus aren’t widely available. Folks at Tencent, for example, got very creative and started using activities in WeChat (a population Chinese chat program) to assess employability of a potential borrower. If suddenly the borrower got disinvited from a work chatting group, that could be a sign of pending unemployment.
Recently, our team ran an experiment to see if we can figure out the guts of a FICO score. Like Coca Cola, the secret recipe is probably guarded with utmost security. Even if you somehow managed to get the actual formula of the FICO score, it will be difficult for someone to replicate it due to lack of data.
We all know that FICO is made up of these categories of behavior:
- Debt Ratio
- Average Age of Credit Life
- Mixture of Credit Tradelines
- Number of inquiries
But is the actual makeup of FICO really as simple as what the above pie charts says. E.g. Delinquencies accounts for exactly 35% of the FICO score? Let’s find out.
Our team trained FICO score as a dependent variable (call me for a crash course on statistics, no time to explain, we have 30 minutes) and fixed over 500 credit attributes as independent variables. After a few tries of stepwise regression analysis. We actually came pretty darn close.
The Y-axis is our “Predicted FICO” and X-axis is the actual FICO score. Grade school math tells us that if the value of the x-coordinates is exactly the same as the value of the y-coordinates, then X (actual FICO) and Y (predicted FICO) is exactly the same.
Well, we aren’t that lucky. However we did get a R-Square of 0.8703 (R-Squared is a goodness of fit test and a value of 1 means a perfect correlation between two distributions.
Head hurting yet? No problem… here are the selected sample of possible FICO ingredients. Use it at your discretion.
Until next time – Timothy Li, CEO of MaxDecisions, Inc.