The Office of the Comptroller of the Currency’s (OCC) revised version of the Comptroller’s Handbook on Fair Lending includes guidelines and processes for examiners. These guidlines are to evaluate the risk of fair lending and to assess compliance with fair lending regulations. It preserves the essence of the current procedures for fair lending examinations from the January 2010 version and adds important details on examination scenarios, risk factors and sound risk management practices. It also includes considerations for machine learning models.

The OCC assumed a more aggressive approach to fair lending risk in the context of models and model risk management (MRM). Firms must meet the challenge of implementing effective fair lending data analytics and MRM frameworks.

Fair lending risk management highlights

• Insufficient analysis and oversight within and across products, business lines, geographic areas, countries and legal entities can pose other fair lending risks.
• Using third parties to perform business activities (such as loan servicing) does not diminish the bank’s responsibility to comply with fair lending laws.
• Effective challenge by an independent party is necessary for a fair lending risk program.
• Banks must assess the aggregate level and direction of fair lending risk and the quality of fair lending risk management, anticipating when examiners might find it insufficient or weak.
• The OCC will expand its focus from credit risk to include compliance, credit, operational, strategic and reputational risks.

In our experience, many banks have sufficient processes to evaluate individual controls; however, they often struggle to assess the effectiveness of the overall system of fair lending risk controls.

Fair lending analyses in a model risk management framework

Data analytics and statistical analysis help identify potential risk areas and direct investigative efforts. However, in practice, we see banks fail to identify and mitigate significant data issues and modeling limitations and fail to realize the full benefits of statistical analyses.

Also, firms struggle to incorporate fair lending models in the MRM framework due to differences in construction and usage from risk models. It’s important to address data, modeling and post-analysis file review for an effective fair lending framework.

Data

In performing statistical analyses and comparative file reviews, banks should evaluate the sample size and create homogeneous segments. Comparative file reviews, which evaluate differences in outcomes for similar applicants, complement statistical analysis in fair lending assessments, providing the additional benefit of identifying data gaps and overlooked factors in models.

When considering data accuracy in fair lending reviews, banks should focus on the fields used in statistical analyses. In the case of mortgage lending, this requires going beyond the reportable fields in the Home Mortgage Disclosure Act (HMDA).6 Banks should also use demographic proxy methodologies, such as Bayesian Improved First Name Surname and Geocoding (BIFSG) to analyze material segments for which demographic information is not collected, such as credit cards.

Modeling

While regulators have issued detailed documentation, development and validation standards for modeling credit, market and other risks, there is less guidance for fair lending models and related MRM. There are several key challenges for fair lending analyses and MRM.

What constitutes a viable model and the validation process differ in the fair lending context. For example, a fair lending model aims to mimic underwriting and pricing policies, therefore, traditional variable selection approaches do not apply. When a strong model cannot be developed, the internal fair lending analytical standards should guide finding the best possible model (such as choosing the best risk-factor binning or interactions) and implementing other compensating controls.

Model assessment must consider the model’s use, such as focal point selection or  selection of files for review. MRM should not apply lower standards to fair lending models because they are only one of several controls.

Importantly, automated and validated credit scoring models present lower fair lending risk. But, complex scoring systems (such as those relying on machine learning or data seldom used for credit decisions) carry the greatest fair lending risk and the handbook instructs examiners to assess MRM practices for machine learning. Risks increase with machine learning or alternative data, as the modeled relationships are no longer explicit, and the data is more likely to obscure proxies of the protected classes. Firms and vendors should also search for less discriminatory alternatives (LDA). In the case of machine learning, it is common to find alternative models, often with the help of automated tools with similar performance but less bias.

Post-analysis file reviews

The file review process should be statistically driven. We have observed that the overall process is often poorly rationalized, and statistical analysis is not used sufficiently to determine sample sizes and files to be reviewed. We advise our clients to tie the process to the firm’s risk appetite, which also helps interpret the results and size the response when issues are found.

We also observe that fair lending analysis and file reviews are often not proactive. In fact, they are often delayed. Many institutions perform their analyses annually and skip ongoing monitoring of fair lending risks. Ongoing monitoring might yield significant benefits, including early detection of risks ahead of the annual assessment. Even simple techniques (like raw data analysis along key business-specific dimensions) can yield results.

How IBM Promontory can help

IBM Promontory has extensive experience helping clients implement effective fair lending data analytics and fair lending MRM frameworks. We help our clients develop fair lending analytical standards and procedures consistent with regulatory guidance, internal MRM standards and industry best practices.

We have extensive experience in enhancing governance frameworks, performing fair lending and redlining data analytics, and advising clients on the intersection between fair lending and other risks (such as climate risks or fraud). We have helped clients develop advanced machine-learning techniques that address fair lending risks.

IBM Promontory’s fair lending expertise goes well beyond data analytics. Skilled professionals are available with experience in conducting large-scale fair lending assessments, developing enhanced analytics-driven fair lending programs and creating analytical frameworks.

Find out more about how IBM Promontory’s experts can help address your risk management needs.