Awareness of embedded bias in US institutions, both public and private has emerged again, at this point in history; perhaps the result of George Floyd’s murder and the Black Lives Matter movement; perhaps the emergence and growth of ESG (environment, social and governance) investing; perhaps, too, a shift in the view among some corporations that they have social responsibilities beyond sole responsibility to shareholders. Backlash to these trends has increased, with anti-ESG laws and regulations in numerous states, including growing government controls over what corporations, the public sector (notably school curricula, public pension investments, women’s healthcare) and individuals may and may not do in their states, based on selective ideology.

Bloomberg reporter Nic Querelo recently published an extensive article (paywall possible) on the topic of the “Black Tax”, that is, the higher cost of financing infrastructure in majority-black cities and school districts. Others approach the topic of institutional bias through analysis of economic benefits that full inclusion could reap. (See Catherine L. Mann and Dana M. Peterson, for CitiGPS, here; or Heather McGee in The Sum of Us: What Racism Costs Everyone and How We Can Prosper Together for just two examples.).

In this context, I’m publishing a set of notes from a panel on bias in credit and credit analysis that took place in Las Vegas at the National Federation of Municipal Analysts’ 2022 annual conference, May 17-20. Emmanuelle Lawrence, Director, U.S. Public Finance, Fitch Ratings, moderated; Dr. Courtney Carter, Assistant Professor of Sociology, UNLV was our presenting academic expert; John Ceffalio, Senior Research Analyst, CreditSights and myself were panelists. You can find the conference agenda and speaker descriptions here. Unfortunately, I could not participate due to a poorly-timed case of COVID. Neene Jenkins, Head of Tax Aware Research, JP Morgan Asset Management, graciously took my place. To facilitate my contribution to the discussion in abstentia, I prepared the following summary (with a few updates and resource links) for the panelists in the form of FAQ.

FAQs to Consider Bias in Credit Analysis
NFMA 2022 Annual Conference
Natalie Cohen, National Municipal Research
May 19, 2022

  1. What are some examples where analysts missed something important that can be attributed to bias?

One of the biggest examples I can think of was in Orange County, California, which most people know, filed for bankruptcy in 1994. Most analysts, including rating agency analysts, assumed that the high wealth in Orange County meant that they would pay their debts no matter what. Residents of the county were so tax averse, that the manager of the of the county’s investment pool (which was the entity that went belly-up, not the county itself), entered into reverse repo agreements with a range of banks in order to increase returns for the local government members of the pool. Those derivatives worked in the county’s favor until they didn’t. One key credit factor that wasn’t looked at was the proportion that interest earnings made up of the county’s revenue base – which was much higher when compared with other county budgets.

2. Aren’t big data, algorithms and machine learning more objective and impartial in determining credit quality?

There’s a saying from a Buddhist monk (and I’m not sure which one) that when you wear sandals the entire earth is covered in leather. This can be interpreted in a number of ways, but in this context, whatever you bring to a situation infuses the situation. At least in your own mind. That’s a softer version of the mantra for computer programmers: “garbage in/garbage out”. Standard credit scoring systems today scrub their applications of race, gender and ethnicity, in an effort to remove bias, but that’s only part of the story.

There are several studies about bias in algorithms and now, some that specifically address credit analysis, credit ratings and scoring used in other contexts (such as ESG for example). Insurance approvals, mortgage lending, personal loans, property and asset appraisals, college and university admissions as well as recruiting and hiring each use algorithms to sort through applicants for approval. There are many other examples.

These systems are data hungry. Situations where data is sparse, or non-existent, are often routinely tossed out of the data set. The input data or the program written to sift for eligible applicants may not consider a broad enough range of variables. It may well be that a borrower is perfectly willing and able to pay, but there’s just a paucity of specific kinds of data to draw a conclusion. Also, machine learning is often tested to see if the machine comes up with the same kind of conclusion or outcome that a human would have. When you think about it, there’s really no reason to assume that algorithms and the data sets used to make decisions are any less biased than we are.

One example that has begun to be studied is the issue of title on physical property among black farmers and homeowners. Property that has been passed on from one generation to the next, bills paid and upkeep maintained, but the property lacks the standard documentation required by banks and insurers. These properties may not be included in property assessments. Lack of title documentation is prevalent in the south, and the Pew Charitable Trusts recently focused on “tangled title” in Philadelphia. Such properties provide community stability especially when there is an intergenerational transfer of property. However, these properties may not be counted or financed in the usual way for lack of documentation.

3. How can we reconsider data inputs?

There’s a study from authors Laura Blattner and Scott Nelson (see citation below) about how flawed data aggravates inequality in credit that looked at whether people who were rejected for home loans would have been likely to default on mortgages if they had been approved. They examined all kinds of other loans, such as car loans, which correlate very closely with how likely a person might default on a mortgage, according to the study. They found that credit scores were less accurate for low-income and minority borrowers than others. Some of this is because low-income and minority borrowers have less data, or data that “checks the box”. For these reasons, certain population groups have been unable to develop strong credit histories. Quoting from the report:

“That creates a misallocation of credit. It can also perpetuate the inequalities because people who can’t get mortgages have less opportunity to build up solid records. They also then lose out on a crucial avenue for building wealth.”

3a. Data inputs may differ based on self-selection among certain social groupings

Take for example, electronic job recruiting. I was at a presentation for women in finance and a woman from online Softbank said that they revised their electronic hiring approach to try to attract more women. They found that women looking for employment would determine that they weren’t qualified unless their background basically met 100% or nearly 100% of the job description, whereas men would choose to apply with far less than a 100% match. Instead of having the job seeker make the decision about which jobs to apply for, they flipped the system around so that applicants input their background information and interests, and the computer would spit out potential matches. When they asked some of the women to share their reactions, they heard, “I didn’t think I was qualified for that!” and ultimately ended up getting many more women to join their workforce.

3b. The Persistence of Legacy Data Inputs

It’s useful for credit analysts to think about how many ways legacy data gets hard coded into decision-making. For example, property is typically appraised is by using “comps”. Aside from the size of a house, its lot size, number of bedrooms, etc. there are neighborhood demographics and amenities that figure into such comparisons. By definition, this approach perpetuates what came before. There have been congressional hearings (here and here) about bias among housing appraisers. Most readers are familiar with the practice of redlining. What might be less well known is that in the early days of the FHA/VA program, real estate agents and banks were instructed to lend only in certain neighborhoods and not in others. The Home Owners Loan Corporation spelled out in great detail, which places were considered “dangerous” for loans.

The Brookings institution and the Ashoka organization recently embarked on a project to address the under-valuation of homes in majority Black communities – called the “Economic Architecture Project”. They are offering challenge grants to innovators that are doing work to improve the valuation of homes in Black communities.

Another example of legacy inputs is higher education’s long-lived policy to accept the children of alumni. This has been the practice of some of the top, elite schools in the country. This approach sustains a demographic from one generation to the next. Some colleges and universities, notably Amherst College in Massachusetts are changing their admission criteria to eliminate legacy preferences.

4. What are some of the consequences of Legacy Inputs?

Finally, and perhaps the largest legacy issue is intergenerational wealth transfer. During the periods after the Great Depression and World War II, federal government policies and programs supported home-ownership, access to college education and money towards business. During the 1950’s and 1960’s these programs spurred suburban development and boosted GDP growth. Of course, greater affordability and availability of the automobile must be factored in as well as federal and state programs to build highways and other infrastructure. However, Blacks were explicitly excluded from these benefits in statute and practices. Federal mortgage guarantees such as FHA/VA, college tuition and small business loans through the GI Bill helped to set a baseline for (white) middle class growth along with property and intellectual assets to transfer to the next generation.

LivingCities.org Associate Director Jeff Ferguson recently offered a concise exploration of the impact of inequities in intergenerational wealth transfer between two hypothetical families, one black, one white.

5. How did these biases play out during the pandemic?

There are some studies showing that the Paycheck Protection Program was targeted at larger employers and at least in the early months and was biased against smaller business. (See Rachel M. B. Atkins, Lisa D. Cook, Robert Seamans paper here.). Along with having less credit-related data, smaller businesses and black-owned businesses tend to be “under-banked”. PPP was initially written to flow to the country’s large banks which, for obvious reasons favored existing clients. It wasn’t until PPP was opened to CDFI’s and MFI’s which tend to serve under-banked communities, that PPP money started flowing to smaller and black-owned business. For similar reasons, there’s evidence that black-owned business tended to get loans from online lenders rather than from banks. I wonder whether those online lenders charged the same interest as the big banks, or higher interest. Common sense would dictate that automated online services are less expensive to operate, but I question whether such savings flowed through to pricing offered to borrowers. Just a hypothesis, worthy of further study.

6. Anything else to share?

6a. Size

There’s certainly a bias towards larger size in our industry. Most firms, whether buy-side, sell-side, or rating agencies, tend to focus on larger borrowers, obviously because that’s where the money is. Of course, smaller borrowers may have less financial flexibility when there’s an economic downturn. However, from a climate perspective, it’s possible that lower income, rural and black-majority communities emit less carbon than larger, more technologically advanced communities. Again, this is another hypothesis to be studied.

6b. Distribution of IIJA funds

Finally, as the Infrastructure Investment and Jobs Act money starts to flow we have concerns about political bias in the distribution of funds. Money is designed to flow to the states which will, in turn, distribute to projects to local communities. Unfortunately, in today’s hyper-partisan pre-election environment, there are several states where the political party of the governor and legislature differ from the party represented at the local level. We would not be surprised to see partisanship interfere with an objective determination of need. The Brookings Institution and other NGO’s have written about this. Sarah Quinn, in American Bonds: How Credit Markets Shaped a Nation, discussed the concept of “veto points”. Veto points reflect those places in an organizational structure that allow for a process to be stopped, slowed down or diverted (in Beltway parlance, “slow-walked”). Consider current governmental partisan gridlock. Quinn writes:

“These cleavages were the foundation for distinct economies, lifestyles, and worldviews that became an existential threat to the nation in the lead-up to the Civil War. Today, a mix of geography, economy, culture and race mark an increasingly hostile gap between ‘blue’ coastal elites and the ‘red’ heartland. Other divides cut across space: the separation of the haves and have-nots, of gender and ideology.”

Sarah Quinn, American bonds, page 8 (Kindle version)

This discussion is intended to raise questions and highlight approaches that we tend to take for granted. From a risk perspective, we may be overlooking the “elephant in the room” but we are also overlooking lots of undiscovered opportunities by only relying on existing medians and benchmarks that may have bias built in.

Selected Additional Resources:

  1. Laura Blattner, Scott Nelson “How Costly is Noise? Data and Disparities in Consumer Credit” https://arxiv.org/abs/2105.07554v1 [econ.GN]
  2. James Manyika, Jake Silberg, Btittany Presten “What Do We Do About the Biases in AI?”, Harvard Business Review, October 25, 2019. https://hbr.org/2019/10/what-do-we-do-about-the-biases-in-ai
  3. Davon Norris, “Embedding Racism: City Government Credit Ratings and the Institutionalization of Race in Markets”, Social Problems, 2021;, spab066, https://doi.org/10.1093/socpro/spab066
  4. Destin Jenkins, The Bonds of Inequality, Debt and the Making of the American City, The University of Chicago Press, Chicago 60637, 2021.
  5. Sarah L. Quinn, American Bonds: How Credit Markets Shaped a Nation, Princeton University Press, 2019. ISBN-13: 978-0691156750
    I note that this book is a must read for those interested in the history of credit markets and in credit analysis. The early pages have a bone-chilling discussion about one of the first collateralized mortgage securitizations in the US: slaves as property assets. Pools of slaves as assets were mortgaged by slave owners to attain bank loans and could be sold if the owner/borrower couldn’t pay or needed to raise additional funding