When is too much data in banking not enough?
When the newly released HMDA data report information on loan pricing but not credit risk.
These half-the-story data create problems for the industry, regulators, and especially analysts evaluating the data. This was less of a concern with previous HMDA data that showed significantly higher denial rates for minorities than for whites, even when income was held constant.
Many analysts found the approach of comparing denials or even loan rates of minorities to whites’ for the same income groups limited, since income is but one risk factor. Most lenders, for example, prefer a middle-income applicant with good credit to a higher-income one with poor credit.
Despite these and other problems, the focus of traditional fair-lending HMDA analysis remained on comparing standard disparity ratios.
A lender criticized for a very high ratio would typically respond by calculating it for different areas and loan categories (for example, purchased versus refinanced home loans). The lender would hope its ratios were lower than or close to the respective aggregate disparity ratios of all HMDA filers or a relevant subset.
Some lenders with relatively high approval rates calculated what I call the “peer approval ratio” for minorities to whites. These lenders would hope that their ratio, defined as the approval rate divided by that of all peers, was higher than or close to 1.00. This meant that the lender was approving minorities at a rate greater than or equal to those of its peers.
Many lenders with high disparity ratios even compared with peers (using the standard approach) might have peer approval ratios well in excess of 1.00.
I reviewed data for one lender that had high disparity ratios, both absolutely and relatively, and very high approval rates for minorities — but even higher approval rates for whites. This resulted in peer approval ratios ranging from 1.00 to 1.50 for different areas and loan categories; a minority applicant had an equal or up to 50% better chance of loan approval with that lender than with others.
The standard approach of not looking beyond simple disparity ratios would have ignored that lender’s much-above-normal minority approval rates.
With the new HMDA pricing data, lenders now must take traditional fair-lending HMDA analysis a step further to include FICO scores and other credit- and transaction-risk information obtained from loan files. This is not easy or inexpensive, but the lack of such data in the HMDA files means that prudent lenders cognizant of reputational and regulatory risks must do it themselves.
For example, a lender that undertook a HMDA FICO risk assessment, which statistically compared median FICO scores for different applicant groups, had a median FICO score for approved minority applicants that was fully 60 points above the median for denied minority applicants.
A FICO difference of 60 points is huge when Fair Isaac Corp.’s own data show that an applicant with the lower score has a three to six times greater likelihood of defaulting.
That same lender’s median FICO score for approved white applicants was likewise 60 points above the median for denied white applicants, again implying a three to six times greater likelihood of defaulting.
Because its median FICO scores for all approvals, regardless of race, was roughly 60 points higher than all denials, again regardless of race, the lender felt it was reasonable to conclude from this analysis that it is was a color-blind lender, as the denials would have resulted in three to six times greater default risk.
An oversimplification of this finding to denied applicants, regardless of their color, might be, “Give me a FICO that is 60 points higher and your loan will be approved.”
Because most lenders consider additional transaction- and credit-risk measures besides FICO scores, and because they don’t cover all borrowers, many lenders now take their fair-lending analysis a step further with a HMDA expanded FICO risk assessment, which statistically compares non-FICO metrics of credit risk and transaction risk for different applicant groups.
Lenders conducting these fair-lending self-assessments hope there is no real (discrimination) problem. If there is, these lenders must take immediate corrective action, especially if pricing patterns in their mortgage banking are questionable or if their loan officers have considerable pricing discretion.
The best proof that traditional HMDA analysis is of limited value today comes from the Fed itself, in its recent study. It found that minorities were generally 20% more likely to get higher-priced loans than whites. That study used existing borrower data on income, loan amounts, and other HMDA factors to explain away more than two-thirds of the discrepancy.
The Fed’s traditional HMDA analysis ended up with a “target list” of about 177 lenders, or 2% of the 8,853 filers. Nonetheless the Fed, realizing how much the new HMDA data are limited, was forced to turn to an outside source to get, as Paul Harvey would say, “the rest of the story.”
To handle what the Fed understatedly called an “important limitation” of their analysis and “to provide some insight into how important” other factors might be in explaining pricing differences across borrower groups, they turned to Georgetown University’s Credit Research Center.
The CRC developed its own database of 626,000 loans from eight subprime lending subsidiaries of large financial institutions, which accounted for 22% of all higher-priced conventional or refinanced loans in 2004. It expanded the traditional HMDA data with internal credit and transaction data such as FICO scores, loan-to-value ratios, appraisal information, type of loan (fixed versus adjustable), type of documentation, whether there was a prepayment penalty, and whether the loan came from a broker.
Using aggregated data for these expanded non-HMDA factors, the Fed found that about one-third of the incidence of higher-priced lending to minorities was explained for conventional loans — and all of it was explained for refinanced loans. The non-HMDA factors also helped explain away most of the higher mean APR spreads for those loans.
The Fed had no choice but to conclude that an analysis employing such comprehensive information, currently not available with the HMDA data, “would be required to draw firm conclusions about racial or ethnic differences in pricing.”
In other words, the most useful data in the Fed study were not even from the Fed but from an outside source.
So why not just make this additional information available with HMDA data in the first place? The industry should blame itself for allowing half-the-story HMDA data to be made available, yet amid all the bad publicity it still rejects calls for making FICO or other credit information available.
Though the Fed’s analysts probably would like the additional pro-competitive data, the official Fed position supports the industry. It argues against expanded data, using some questionable arguments based on additional data-collection and reporting costs to the industry, possible privacy issues, and, oddly, even issues of competitive business strategy.
The only cogent argument that can be made against making such data available is the fact that there is no consistent credit- risk measure, including FICO, for all lenders and borrowers. Despite its limitations, FICO is the closest thing to an industry credit-risk standard, and an increasingly large portion of the public understands it more than other possible measures.
Considering the discrimination allegations and that the Fed must turn to outside data to produce a useful HMDA analysis, it is time to consider making FICO data aggregated by broad category available with 2005 HMDA data (and possibly on a retroactive basis for 2004). Other non-FICO data, similar to that in the CRC study, might also be made public over time.
Meanwhile lenders that wish to make public their median FICO scores or other non-FICO data aggregated by key 2004 HMDA categories should be allowed to do so — especially lenders targeted by the Fed.
Traditional fair-lending HMDA analysis is of limited value with the new pricing data, which demand rigorous statistical analyses of FICO and non-FICO credit- and transaction-risk data from loan files. If the regulators do not require these latter data to be included in HMDA filings, then prudent lenders must take it upon themselves to conduct their own fair-lending self-assessments.
The best public policy solution, however, would be for regulators, the industry, and consumer groups alike to focus more effort on credit availability and pricing issues for borrowers on the basis of their need — that is, their income — rather than their race.
That is why the CRA’s focus on income is better than fair-lending’s focus on race, especially since “minorities” are becoming the majority in more states — now including California and Texas — as they’ll ultimately be in the nation at large.
