@llecs wrote:Disclaimer: I didn't read the entire PDF and missed the methodology probably, but I'm curious how they obtained the data. If a lender pulled a FICO, they wouldn't both pull an industry-specific score and a Classic FICO at the same time. There's no need to do that and it costs them money....unless the study paid for these 200k pulls. Without reading more, and on the surface, I have some red flags.
They actually did pull all the scores from a CRA at the same time. So they have 200,000 identity filtered CRs with VantageScore, FAKO, FICO, and two industry specific FICOs pulled on each consumer report at the same time.
However, these 200,000 are random samples of each CRA's entired database. They are different sets of individuals so there is no way to compare score consistency on the same individual across CRAs nor was that the studies goal.
The study was mandated by Congress and targeted only to scoring variation from one algorithm to another on the same consumer report.
As it is clear that "correlation" behaviors aren't generally known, Just a note of clarification.
Correlation values of .99 do not mean that 99% of the scores match or that two scores will only be, on average 1% off. It's a statistical measure and one that leaves a lot of room for variation in FICO scores. It's also kind of weird in that as it increased it tends to overstate what one would expect.
For instance if a correlation of .92 produces a typical variation of 40 points (using FICO scaling and taking a SWAG) around the 620 mark, a correlation of .98 would only reduce the expected variation to 20 points. Alternately, if 1 out of 4 people had no history of autoloans and took a 40 point hit as a result but otherwise matched the generic FICO that would also produce a correlation of .98