Chapter 11: Exercises

Bias in Financial Services and Credit

Difficulty Scale: - ⭐ Comprehension — recall and basic application - ⭐⭐ Analysis — applying concepts to new situations - ⭐⭐⭐ Synthesis — combining ideas, designing solutions - ⭐⭐⭐⭐ Evaluation — critical assessment, policy recommendation

† Exercises marked with a dagger are designated capstone exercises recommended for formal assessment or team projects. They require extended written responses, data analysis, or stakeholder simulation.

Part A: Comprehension and Recall

Exercise 1 ⭐

Match each statute or regulatory framework to its primary fair lending function:

After completing the matching, identify which of these frameworks explicitly addresses AI or machine learning models by name, and what each says about algorithmic decision-making.

Exercise 2 ⭐

Define the following terms in your own words, using no more than two sentences each:

a) Redlining (historical) b) Algorithmic redlining c) Thin-file problem d) Proxy variable e) Disparate impact f) Disparate treatment g) Four-fifths rule h) Adverse action notice

For each definition, identify one concrete example from the chapter that illustrates the concept.

Exercise 3 ⭐

List five specific applications of AI in financial services discussed in Chapter 11. For each: - Describe what the AI system does - Identify the primary bias risk - Identify the primary regulatory framework that applies

Format your answer as a table.

Exercise 4 ⭐

The FICO credit score uses five categories of data, weighted by their approximate contribution to the score. List all five categories and their weights. Then, for each category, explain whether and how that category might produce racially disparate outcomes — not because of discrimination, but because of historical patterns in credit access.

Exercise 5 ⭐

The Apple Card controversy centered on a question that Goldman Sachs and regulators could not definitively answer. State that question precisely. Then explain what features of the situation made it impossible to answer under current law.

Part B: Analysis and Application

Exercise 6 ⭐⭐

A regional bank is building an auto loan underwriting model. The data science team proposes including the following variables:

• Credit score
• Annual income
• Employment type (salaried / hourly / self-employed)
• Years at current employer
• Vehicle make, model, and year
• Loan-to-value ratio (loan amount / vehicle value)
• Debt-to-income ratio
• Zip code of the dealership where the vehicle was purchased
• Distance between applicant's home address and the dealership
• Whether the dealership is classified as serving a "luxury" or "standard" market

For each variable: a) Assess whether it might serve as a proxy for race or ethnicity b) Identify a plausible business justification (legitimate credit risk rationale) for including it c) Recommend whether to include, exclude, or include with modifications, and explain your reasoning

Exercise 7 ⭐⭐

Read the following scenario and answer the questions below.

Scenario: ClearPath Mortgage, a fintech lender, has deployed an AI-driven mortgage underwriting system. A CFPB examination team reviews its HMDA data and finds the following approval rates for a recent 12-month period:

a) Calculate the four-fifths rule disparate impact ratio for each group, using Asian applicants as the reference group (highest approval rate).

b) Which groups fail the four-fifths rule?

c) ClearPath argues that the disparities are fully explained by credit score and DTI differences across groups, which are not shown in the HMDA data. How should the CFPB examiner evaluate this argument? What additional information would help?

d) If ClearPath's argument is correct — that the disparities reflect legitimate credit risk differences — does that mean the outcomes are fair? Why or why not?

Exercise 8 ⭐⭐ †

HMDA Data Interpretation Exercise

The Consumer Financial Protection Bureau publishes HMDA data annually at https://ffiec.cfpb.gov/. Access the most recent available HMDA data snapshot and answer the following questions:

a) For your state (or a state of your choice), identify the five largest mortgage lenders by application volume. For each lender, find the approval rate for Black applicants and white applicants. What is the racial gap in approval rates?

b) Calculate the four-fifths rule disparate impact ratio for Black applicants vs. white applicants for each of your five lenders.

c) Do the disparities you find vary significantly across lender type (bank vs. credit union vs. independent mortgage company vs. fintech)? What might explain the variation?

d) The HMDA data includes the automated underwriting system (AUS) recommendation. What share of Black applicants received a "Refer" or "Refer with Caution" recommendation from the AUS, vs. white applicants? What does this tell you about where in the process the disparities arise?

e) Write a 200-word summary of your findings, in the style of a regulatory examination memo, identifying which lenders warrant enhanced fair lending examination.

Exercise 9 ⭐⭐

A credit card company sends the following adverse action notice to an applicant who was denied credit:

"Thank you for applying for the Platinum Rewards Card. After reviewing your application, we are unable to approve your request at this time. The primary reasons for this decision are: (1) too many inquiries in the past 12 months; (2) insufficient credit history; (3) high ratio of balances to credit limits."

a) The company's underwriting model uses a gradient boosting algorithm with 150 input variables. The adverse action notice was generated by selecting the three most commonly cited reason codes from a standardized list, without verifying that these codes reflect the model's actual decision for this applicant.

Identify the specific ECOA / Regulation B provision this may violate. What does the provision require?

b) Using the CFPB's 2022 adverse action guidance, assess whether this notice is compliant. What specific problems does the guidance identify with this approach?

c) The applicant believes she was denied because she is a recent immigrant with a shorter U.S. credit history, even though she has an extensive credit history in her country of origin. What ECOA provision might protect her? What information would you need to evaluate her claim?

d) Rewrite the adverse action notice to be compliant with the CFPB's 2022 guidance. Note: you may need to specify what additional information from the model would be required.

Exercise 10 ⭐⭐

The following is a stylized representation of matched-pair testing results for a mortgage lender:

200 matched pairs were constructed: Black applicants and white applicants with credit scores within 15 points of each other, DTI ratios within 3 percentage points, LTV ratios within 5 percentage points, and income within 10%.

a) Calculate the differential treatment rate (the rate at which the Black applicant was denied while the white applicant with a similar profile was approved).

b) Is this a large, moderate, or small differential treatment rate in the context of fair lending enforcement? What benchmarks exist for evaluating this?

c) The lender argues that the matching was imperfect — the financial profiles of paired applicants were not truly identical because credit scores within 15 points could represent meaningfully different risk. How would you evaluate this argument?

d) Design an improved matched-pair methodology that addresses this concern while remaining practically implementable.

Part C: Synthesis and Design

Exercise 11 ⭐⭐⭐ †

Fair Lending Model Governance Framework Design

You have been hired as the Chief Risk Officer at Meridian Digital Bank, a mid-size bank that is preparing to deploy an AI-driven home equity loan underwriting model. The model uses 85 input variables and a random forest architecture. It was built by an external vendor and uses training data from 2015-2023.

Design a fair lending governance framework for this model. Your framework should address:

a) Pre-deployment validation: What specific tests would you require before approving the model for production use? How would you conduct demographic disparity analysis when your customer data includes race only for applicants who voluntarily disclose it (approximately 60% of applications)?

b) Vendor due diligence: Draft five specific questions you would require the model vendor to answer in writing before you would consider using the model. What contractual protections would you require?

c) Adverse action notice infrastructure: Given that the model uses 85 variables and a nonlinear architecture, how would you generate accurate adverse action notices compliant with Regulation B? What technical methods would you use? What would the internal process look like?

d) Ongoing monitoring: After the model is deployed, what ongoing monitoring would you conduct? At what frequency? What thresholds would trigger escalation to your Fair Lending Committee?

e) Incident response: If your quarterly monitoring reveals that Black applicants are being denied at a DIR of 0.74 — below the four-fifths threshold — what is your response protocol? Walk through the specific steps, stakeholders, and timeline.

Format your response as an internal policy document.

Exercise 12 ⭐⭐⭐

The Apple Card controversy revealed that Goldman Sachs's algorithm produced gendered credit limit disparities that regulators could not prove were illegal. Suppose you are a policy analyst at the CFPB and you have been asked to draft a proposal for regulatory reform to address this gap.

a) Identify the specific gap in current ECOA doctrine that the Apple Card case illustrates.

b) Draft two distinct regulatory approaches to closing this gap: - Approach A: Process-based regulation — requiring specific governance processes (bias testing, documentation, oversight) before algorithmic credit models are deployed - Approach B: Outcome-based regulation — requiring lenders to meet specific demographic outcome standards (e.g., approval rates within a specified range)

c) Evaluate each approach on four criteria: effectiveness in reducing bias, cost of implementation, regulatory administrability, and risk of unintended consequences.

d) Which approach would you recommend, and why? Are there elements of both that could be combined?

Exercise 13 ⭐⭐⭐

Alternative Credit Data Evaluation

NovaCredit, a fintech company, uses the following alternative credit data variables in its personal loan underwriting model for thin-file applicants:

• Rent payment history (via landlord database)
• Utility payment history
• Mobile phone payment history
• Grocery store transaction patterns (frequency, average basket size, merchant category)
• Social media account age and activity level
• Educational institution attended and degree earned
• Employer and employment type (full-time/gig/contract)
• Zip code of current residence

a) For each variable, assess: (1) whether it is likely to reduce or increase racial disparities compared to traditional credit bureau data; (2) whether it introduces new proxy discrimination risks; and (3) whether its use raises other ethical concerns beyond proxy discrimination.

b) NovaCredit claims its model approves 40% more applicants from majority-minority census tracts than a comparable traditional credit model. What additional information would you need to evaluate this claim? How might the claim be misleading even if literally true?

c) The CFPB is deciding whether to grant NovaCredit a No-Action Letter, as it did for Upstart in 2017. What conditions, if any, should the CFPB attach to any No-Action Letter? What ongoing data reporting would you require?

d) Design a methodology for testing whether NovaCredit's model reduces or increases racial disparities compared to a FICO-based baseline model.

Exercise 14 ⭐⭐⭐

Insurance Pricing and Proxy Discrimination

ProShield Insurance uses an auto insurance pricing model that includes the following variables. For each group of variables, analyze the proxy discrimination risk and the regulatory landscape:

Group 1 — Driving behavior (via telematics): Speed, braking frequency, cornering, time of day, days per week driven, average trip length

Group 2 — Vehicle characteristics: Make, model, year, safety rating, theft frequency for that vehicle type

Group 3 — Geographic variables: Zip code, census tract claim frequency, proximity to high-traffic roads

Group 4 — Financial variables: Insurance credit score (derived from credit bureau data), prior insurance history (whether insured continuously)

a) Rank these four groups from highest to lowest proxy discrimination risk. Explain your ranking.

b) California prohibits the use of credit scores in auto insurance pricing. If ProShield must remove Group 4 variables from its model for California customers, what is the likely effect on: (i) pricing accuracy; (ii) racial disparities in premium levels; (iii) ProShield's overall loss ratio?

c) ProShield's actuary argues that Group 3 geographic variables are legitimate because they reflect actual claim frequency by location — the model is not discriminating by race, it is charging more in high-claim areas. Evaluate this argument. What information would you need to determine whether the actuary is right?

d) Design a regulatory framework for AI-driven auto insurance pricing that balances actuarial accuracy with protection against proxy discrimination. What should regulators require lenders to disclose or demonstrate before using each group of variables?

Exercise 15 ⭐⭐⭐ †

Adverse Action Notice Drafting Exercise

Maria Chen applied for a $450,000 mortgage to purchase her first home. She is a 34-year-old software engineer earning $145,000 per year. Her credit score is 698. Her debt-to-income ratio is 41%. Her loan-to-value ratio would be 88%. She has been employed at her current company for 18 months, previously working as a contractor for 3 years. She has no prior defaults. The neighborhood where she is purchasing has a neighborhood risk score of 74 (out of 100), reflecting below-average property value appreciation over the past five years.

The bank's AI model denied her application. A post-hoc explanation using SHAP values identifies the following as the top five drivers of the denial:

1. Neighborhood risk score (contribution: -0.38 on log-odds scale)
2. Loan-to-value ratio above 85% (contribution: -0.31)
3. Years at current employer below 2 (contribution: -0.22)
4. Debt-to-income ratio above 40% (contribution: -0.18)
5. Credit score below 720 (contribution: -0.15)

a) Draft a ECOA-compliant adverse action notice for Maria's application. The notice should provide the four principal reasons for the denial in consumer-understandable language, consistent with the CFPB's 2022 adverse action guidance.

b) After drafting the notice, identify any ethical concerns with the reasons you have provided. The neighborhood risk score is a significant driver of the denial — how do you handle explaining this in a way that is both accurate and compliant with fair lending law?

c) Maria calls your customer service line and asks what she can do to improve her chances of approval on a future application. Based on the SHAP values, what advice would be accurate and actionable? What advice would be misleading?

d) Maria is Asian American. Her neighborhood has a census tract racial composition of approximately 65% Asian American, 20% Hispanic, and 15% white. Does this information change your assessment of the decision? What additional analysis would you conduct?

Part D: Evaluation and Policy

Exercise 16 ⭐⭐⭐⭐

Essay: The Limits of Disparate Impact Doctrine Applied to Algorithmic Lending

Write a 1,000-word essay arguing either FOR or AGAINST the following proposition:

"The disparate impact doctrine under ECOA is an adequate legal tool for addressing racial bias in algorithmic mortgage underwriting, and no new legislation is needed — only more vigorous enforcement."

Your essay should: - Accurately describe how disparate impact doctrine currently applies to algorithmic lending - Engage with the strongest counterarguments to your position - Cite specific cases, investigations, or regulatory actions discussed in Chapter 11 - Conclude with a specific, actionable recommendation

Exercise 17 ⭐⭐⭐⭐ †

Policy Simulation: Regulatory Rulemaking

You are a senior staff member at the CFPB preparing to draft a Notice of Proposed Rulemaking (NPRM) on "Algorithmic Fair Lending Requirements for AI-Driven Credit Models." The proposed rule would require:

1. Pre-deployment disparate impact testing with disclosure to regulators
2. Ongoing quarterly demographic monitoring with reporting thresholds
3. Explainability requirements for adverse action notices using AI models
4. Vendor due diligence standards for purchased AI models
5. Consumer remediation requirements when significant disparate impact is identified

For each of the five proposed requirements, prepare:

a) Regulatory justification: What statutory authority supports this requirement? What evidence of harm justifies regulatory intervention?

b) Expected industry objections: What objections would industry stakeholders likely raise? How would you respond?

c) Benefit-cost analysis: What benefits does the requirement provide, in quantitative terms if possible? What compliance costs would it impose?

d) Draft regulatory text: Write 2-3 sentences of actual regulatory text for the requirement, in the style of federal administrative regulations.

Note: This exercise is designed for a team of 4-5 students, with each team member responsible for one proposed requirement.

Exercise 18 ⭐⭐⭐⭐

Critical Evaluation: The Fintech Democratization Narrative

The fintech industry has consistently marketed itself as democratizing finance — using technology to expand credit access to underserved populations. Evaluate this narrative critically.

a) What is the strongest version of the fintech democratization argument? What evidence supports it?

b) What does the empirical evidence actually show about racial disparities in fintech lending compared to traditional banking? Summarize the research findings discussed in Chapter 11 and any additional sources you can identify.

c) Is there a structural reason why fintech lending might produce more racial equity than traditional banking? Is there a structural reason why it might produce less? Evaluate both possibilities.

d) The fintech regulatory gap — lighter supervision for non-bank fintechs — is sometimes justified as enabling innovation. How should regulators balance innovation promotion against fair lending protection? What is the cost of the regulatory gap if fintech lending is systematically biased?

e) Write a 400-word executive summary of your findings for a CFPB director who is deciding whether to expand supervision of fintech lenders.

Exercise 19 ⭐⭐⭐⭐

Case Comparison: Apple Card vs. Algorithmic Redlining

Chapter 11 presents two major case studies of algorithmic discrimination in financial services: the Apple Card gender controversy (individual-level disparate treatment) and the algorithmic redlining documented by The Markup (geographic/demographic disparate impact).

Write a comparative analysis addressing:

a) What are the key similarities in how bias arises in each case? What are the key differences?

b) What are the evidentiary challenges regulators face in each case? Why is illegal discrimination harder to prove in one case vs. the other?

c) What remedies are available in each case? How effective are they likely to be?

d) Which case presents a more serious problem for consumers? For the financial system? For fair lending enforcement? (You may reach different answers for each of these three questions.)

e) Design a unified regulatory approach that would address both types of algorithmic discrimination without requiring fundamentally different legal frameworks for each.

Exercise 20 ⭐⭐⭐⭐ †

Capstone Simulation: Fair Lending Committee Review

This exercise requires a group of 5 students playing the following roles: - Chief Risk Officer (chair) - Chief Compliance Officer (fair lending lead) - Chief Data Scientist (model owner) - General Counsel (legal risk) - Consumer Advocate (community development officer)

The scenario:

First Community Bank is preparing to launch an AI-driven credit card product targeting "emerging credit" consumers — people with credit scores between 580 and 680. The model uses 120 variables including income, employment, credit bureau data, bank transaction history, and a proprietary "financial resilience score" developed by the vendor.

Pre-deployment testing shows:

• Overall model AUC: 0.81 (good predictive accuracy)
• Approval rate, white applicants: 58%
• Approval rate, Black applicants: 41%
• Approval rate, Hispanic applicants: 44%
• Disparate impact ratio (Black): 0.71 (below the four-fifths threshold)
• The "financial resilience score" is the third-highest importance variable (7.2% importance) and is correlated with applicant zip code (r = -0.42)
• Removing the financial resilience score reduces the AUC to 0.78 and increases the disparate impact ratio for Black applicants to 0.77 (still below 0.80)
• The vendor declines to disclose the proprietary components of the financial resilience score

Each student should: a) Prepare a 5-minute oral statement from their role's perspective on whether First Community Bank should deploy the model in its current form, with the financial resilience score removed, or not at all.

b) After all statements, the committee must reach a consensus decision. Document the decision, the key arguments that prevailed, and any conditions attached to the decision.

c) Write a 300-word memo to the Board of Directors summarizing the Fair Lending Committee's decision and reasoning.

Exercise 21 ⭐⭐

International Comparison: EU AI Act vs. U.S. Approach

The EU's AI Act (2024) classifies AI systems used in credit scoring and creditworthiness assessment as "high-risk AI systems" requiring: - Pre-deployment conformity assessment - Technical documentation and risk management - Data governance and quality standards - Transparency and human oversight - Registration in an EU database

The U.S. framework relies primarily on: - ECOA disparate impact enforcement after-the-fact - Bank regulatory examination through SR 11-7 - CFPB supervisory authority over large nonbanks - State fair lending laws

a) What are the advantages of the EU's proactive, pre-deployment regulatory approach?

b) What are its disadvantages or risks?

c) Why might a U.S.-style post-deployment enforcement approach be preferable in some respects?

d) Design a hybrid approach that takes the best elements of each system.

Exercise 22 ⭐⭐

The Thin-File Problem: Structural Solutions

The thin-file problem — in which historically excluded groups have insufficient credit history to generate reliable scores — is documented and widely acknowledged, yet it persists. For each proposed solution below, assess its potential effectiveness, its risks, and its feasibility:

a) Including rent payment history in FICO scores universally (currently optional) b) Creating a federal alternative credit database covering utility, rent, and telecom payments c) Allowing credit scores to be ported internationally for immigrants d) Requiring lenders to offer a "credit starter" product with guaranteed approval and low limits to all adults with no credit history e) Replacing credit scores with income-based DTI assessment as the primary underwriting variable for small consumer loans

Exercise 23 ⭐⭐⭐

Fraud Detection Disparate Impact

A major bank's fraud detection system flags approximately 2.3% of all credit card transactions for secondary review or automatic blocking. An internal analysis finds the following flag rates by demographic group:

a) Calculate the disparate impact ratio for each group, using white customers as the reference.

b) What business justification might the bank offer for these disparities?

c) The fraud detection system was trained on three years of historical transaction data. What features of the training data might produce these disparities, even without any intentional discrimination?

d) Design a remediation plan: what steps should the bank take to reduce these disparities without significantly increasing its actual fraud losses?

e) What customer-facing transparency would be appropriate? What transparency would compromise fraud detection effectiveness?

Exercise 24 ⭐⭐⭐

Community Reinvestment Act in the Digital Age

The CRA requires banks to affirmatively serve the credit needs of all communities in their "assessment areas" (geographic service areas defined by branch locations). A federal court in 2024 considered whether a bank with no physical branches — all digital — had any CRA assessment area obligations.

a) What is the policy rationale for the CRA's geographic assessment area approach?

b) Does that rationale apply to a digital lender? Why or why not?

c) If a digital lender serves customers in all 50 states but its algorithm approves applications from majority-white census tracts at significantly higher rates than majority-minority census tracts, does it matter whether the CRA technically applies to it?

d) Design a CRA compliance framework for digital-only lenders that captures the law's intent while acknowledging the absence of geographic service areas.

Exercise 25 ⭐⭐⭐⭐ †

Research Project: Tracking the Markup's Findings

The Markup's 2021 investigation documented racial disparities in 2019 HMDA mortgage data. Using HMDA data from subsequent years (2020, 2021, 2022, or the most recent available), conduct a follow-up analysis.

a) Replicate The Markup's core methodology as closely as possible using the available HMDA data fields. What approval rate disparities do you find for Black, Hispanic, Asian, and Native American applicants compared to white applicants, controlling for income and DTI?

b) Have the disparities documented by The Markup decreased, increased, or remained stable in subsequent years? What factors might explain the trend you observe?

c) Did the COVID-19 pandemic years produce any notable changes in HMDA approval rate patterns by race? What hypotheses would explain any changes you find?

d) The Markup focused on conventional mortgages. Extend the analysis to FHA, VA, and USDA loans. Are racial disparities larger or smaller in government-backed lending programs? What would explain the difference?

e) Present your findings in a 1,500-word investigative report in the style of The Markup, with appropriate statistical methodology notes and caveats. Include at least one data visualization.

Note: HMDA data is publicly available at https://ffiec.cfpb.gov/. Python or R code for data processing is acceptable and encouraged.