Originally published in ABA Risk and Compliance, January/February 2026

Introduction

In the world of fair lending analytics, data is everything — except when it is not. While financial institutions collect mountains of transactional and behavioral data, they often lack one crucial piece of the puzzle: customers’ race, ethnicity, and sex. Outside of mortgage lending, where the Home Mortgage Disclosure Act (HMDA) requires collection and disclosure of certain demographic information for government monitoring purposes, demographic data is frequently missing. Even for mortgage data, purchased loans and loan servicing systems may lack demographic information, often referred to as government monitoring information (GMI). Indeed, Regulation B generally prohibits creditors from requesting GMI outside of narrow exceptions for applications for dwelling-secured loans, self-tests, and Special Purpose Credit Programs. Although Section 1071 of the Dodd-Frank Act requires lenders to collect demographic data on small business loans, that requirement is not currently in effect. 

Since a robust fair lending monitoring and testing program encompasses products and services beyond those for which GMI collection is permitted, the absence of GMI creates a significant gap in the ability to conduct fair lending monitoring and testing. What is a lender to do?  

Many lenders, as well as banking regulators, have turned to proxy methodologies to broaden the scope of their fair lending monitoring, testing, and examinations. Proxy methods use available information, such as name, address, and age, to infer race, ethnicity, and sex.

Common Race and Ethnicity Proxy Methodologies

Researchers, regulators, and fair lending officers have used several methods to infer demographics. Census data serve as the basis for most proxy methods. Still, they vary in the component of Census data used and the exact calculations employed in classifying a person’s likely race and ethnicity. We discuss several examples of proxy methodologies below.

A single data point, such as surname or census tract demographics, is the basis of the simplest race and ethnicity proxy methods. The benefit of these approaches is their simplicity and ease of implementation. More complex techniques, which combine multiple data points, such as surname, first name, and tract demographics, typically outperform simpler methods in classification accuracy.

Surname-Only

First, consider surname-only (SO) proxy imputation. SO proxies typically are based on the racial and ethnic distribution of surnames in Census data, although other sources, such as voter registration data, are sometimes used. For example, a person whose last name is “Garcia” is statistically likely to be Hispanic, as “Garcia” is one of the most common Hispanic surnames. Similarly, someone with a surname of “Nguyen” is statistically more likely to be Asian than someone with a surname of “Smith.” However, a surname-only proxy is less accurate when racial and ethnic groups share common surnames. As an example, “Smith” is a common surname for White, African American, Native American, and Multiracial persons. 

The Federal Reserve publishes a list of Hispanic surnames based on decennial Census data, which can serve as the basis for a publicly sourced method for identifying Hispanic applicants. This approach has the benefit of being promulgated by a federal banking regulator, but it cannot classify racial minorities in addition to ethnic minorities. 

Geography-Only

Another single data point approach is to proxy for race and ethnicity based solely on geography. In geography-only (GO) methods, the distribution of races and ethnicities at the Census tract level is used to infer probable racial and ethnic classification. Like surname-based classification, this method is relatively simple to implement. GO proxy methods are also used when names are unavailable, as may be the case with business borrowers. However, geography-only proxy methods are less accurate in diverse or gentrifying neighborhoods. In practice, geography-only methods are most effective in classifying applicants between “likely minority” or “likely not minority,” as relatively few tracts have concentrated populations of a single race or ethnicity. In the 2020 Census, less than 20 percent of tracts had a majority population of a single race or ethnicity, and Asian, Native American, or Native Hawaiian/Pacific Islander were the majority in less than one percent of census tracts each.

Bayesian Improved Surname Geocoding (BISG)

BISG combines surnames and geographic information using Bayes’ theorem to produce probabilistic estimates of race and ethnicity that are more accurate than proxies based on either surname or geography alone. RAND, a nonprofit, nonpartisan research organization, originally developed BISG in a healthcare context. The Consumer Financial Protection Bureau (CFPB) later adopted BISG for fair lending examinations when self-reported race and ethnicity data were not available. 

Benefits of BISG include its use by federal banking regulators and industry participants. BISG is also more accurate than surname-only and geography-only methods. However, BISG is less accurate for African American and Hispanic populations than for Asian and White populations. Additionally, the inclusion of geographic information results in less accuracy in diverse neighborhoods. A recent survey by Asurity Advisors found that BISG is the most commonly used proxy methodology in banking. 

Bayesian Improved First Name Surname Geocoding (BIFSG)

BIFSG considers given, or first names, as well as surnames and geographic information to produce a probabilistic estimate of race and ethnicity. Like BISG, BIFSG is more accurate than surname-only or geography-only proxy methods. Additionally, one study using BIFSG on 2010 Census data found this methodology provided equal or better accuracy than BISG for all major racial and ethnic groups. Accuracy improvements were most notable for groups with less distinctive racial and ethnic associations with surnames, such as African Americans, although there were also some gains for Hispanic and Asian/Pacific Islander populations. There is some evidence that BIFSG may modestly improve the ability to classify Native Americans accurately, but performance remains poor for both American Indian and Alaskan Native groups.  Similarly, the ability to distinguish between Asian and Native Hawaiian or Pacific Islander populations is poor for all proxy methods. 

The addition of given names may also enhance the ability to classify residents of diverse Census tracts accurately. However, no banking regulator has adopted this methodology, although it has been used by academics and federal contractors, albeit outside of banking applications. 

Machine-Learning Proxies

Machine learning models, including Long Short-Term Memory recurrent neural networks (LSTM), transformer architecture, random forests (RF), natural language processing (NLP), and XGBoost (XGB) ensembles, also hold promises for creating accurate proxies. Examples of these types of proxy methods include pyethnicity, raceBERT, ethnicolr, and rethnicity. 

In addition to different methodologies, these proxy methods often have unique, and sometimes proprietary, training datasets.  Many of the training datasets include voter registration data or healthcare data, as well as Census data, but developers trained some machine learning proxy models on data from a single state or a small group of states. For example, Xie developed rethnicity using Census and Florida voter registration data, while Li trained pyethnicity on U.S. voter registration data from all 50 states combined with Census data.  Machine learning proxy models also use different calculation techniques than BISG, BIFSG, SO, or  GO proxy methods. The combination of unique datasets and differing calculation techniques makes machine learning proxy methods more opaque than other methods. 

Using Florida voter registration data, Chintalpati et al. found that name-based machine learning proxy models were more accurate in out-of-sample testing than other types of name-only models. They also found full (first and last name) models more accurate than surname-only models, particularly for African Americans. In testing BISG and BIFSG against machine learning models trained on voter registration data from California, Florida, Georgia, and North Carolina, Decter-Frain found that machine learning models generally outperformed BISG. Still, results differed by both state and proxy method. 

In short, machine learning methods have the potential to create better proxies but still suffer from a lack of classification strength for American Indian/Alaskan Native populations. Additionally, the methodologies are less transparent. No federal banking regulator has publicly adopted a machine learning proxy method.

Comparative Performance Analysis

For banks subject to regulatory oversight, model validation requirements, and documentation expectations, transparent and easily replicable proxy methodologies based on public datasets, such as surname-only, geography-only, BISG, and BIFSG, may be preferable to more opaque machine learning proxy methods. Amongst these approaches, past research indicates that BISG and BIFSG hold the most promise for accurate proxies.

Researchers used 2010 Census data in most past evaluations of proxy methods. To determine which method results in the most accurate proxy results, the authors used 2020 Census data and a data file consisting of 1,206,054 single applications for HMDA-reportable loans to compare outcomes using BISG and BIFSG. Use of HMDA data permits comparison of proxy results to a mortgage applicant’s self-identified racial and ethnic classifications. The classification threshold, which is the minimum probability of group membership used to classify someone as a group member, affects proxy results. To assess the impact of the classification threshold, we tested two classification thresholds  —  50 percent and 80 percent  —  for each method. We assessed proxy accuracy using three metrics:

• True Positive Rate, which is the proportion of the members of a particular racial or ethnic group accurately classified by the proxy method;
• False Negative Rate, which is the proportion of the members of a particular racial or ethnic group misclassified by the proxy method; and 
• False Positive Rate, which is the proportion of inaccurate proxy classifications for each racial and ethnic group.

Both BISG and BIFSG, when using a 50 percent classification threshold, outperform either proxy methodology with an 80 percent classification threshold for both true positive and false negative rates. However, the best performing methodology varied by minority race and ethnicity. For borrowers that self-identified as American Indian/Alaskan Native, Asian/Pacific Islander, or African American, BIFSG had both higher true positive and lower false negative rates. For borrowers who self-identified as Hispanic or non-Hispanic White, however, BISG resulted in higher true positives and lower false negatives. For all racial and ethnic groups, BISG with an 80 percent classification threshold had the lowest false positive rate. The results of the comparison are in the tables below.

Conclusion

When it comes to proxy methods, no single method is best suited for all lenders, geographies, portfolios, and purposes. Lenders proxying their data for fair lending analyses should consider the racial and ethnic groups most prevalent in their geographic footprints, as well as the proxy usage, in conjunction with regulatory standards and common industry practices, when selecting a proxy methodology. For lenders accepting HMDA-reportable applications as well as applications for other types of credit, it may be feasible to compare the accuracy of various proxy methodologies on the lender’s portfolio. Keep in mind, however, that the racial and ethnic distribution of mortgage applicants may not be the same as the demographic distribution of applicants for other types of credit. Additionally, the racial and ethnic distribution of the population may vary by geography. 

About the authors

Lynn Woosley is a Managing Director with Asurity Advisors and a member of the Editorial Advisory Board for ABA Risk and Compliance magazine. Lynn has over 30 years of experience in risk management, spanning both financial services and regulatory environments. She is an expert in consumer protection, including fair lending, fair servicing, community reinvestment, and UDAAP. Before joining Asurity Advisors, Lynn led the fair banking practice for an advisory firm. She has also held multiple leadership positions, including Senior Vice President and Fair and Responsible Banking Officer, within the Enterprise Risk Management division of a top 10 bank. Prior to joining the private sector, Lynn served as Senior Examiner and Fair Lending Advisory Economist at the Federal Reserve Bank of Atlanta. Connect with Lynn at lwoosley@asurity.com. 

Dr. Anurag Agarwal is the Founder & President of RiskExec. He is the original architect and developer behind the RiskExec platform and has been working within the fair lending and compliance industry for more than 30 years. Anurag began working on fair lending procedure automation while getting his PhD in Operations Research, Statistics, and Management Science from the Haslam College of Business at the University of Tennessee, Knoxville. Today, he is a nationally renowned expert in fair lending, Home Mortgage Disclosure Act (HMDA), Community Reinvestment Act (CRA), and other related areas, and is a frequent speaker at ABA and MBA industry events. Anurag co-founded Compliance Solutions back in 2001, before founding Risk Management Solutions Inc. in 2006 and serving as its President and CEO until 2015. RMS was ultimately acquired by Asurity Technologies in 2015. As of January 2025, RiskExec became a standalone company. Connect with Anurag at aagarwal@riskexec.com.