| Course | FI311 FinTech Law and Ethics |
|---|---|
| Unit | Unit 7 |
| Paper type | Algorithmic fairness memo |
| Length | About 760 words, 5 pages double-spaced |
| Format | APA 7 student paper |
| School | Purdue University Global |
| Program | BS in Finance |
| Updated | October 2026 |
Free sample paper for FI311 Unit 7
A Model That Never Sees Age but Sorts by It: A Unit 7 Algorithmic Fairness Memo to a Lender's Credit Committee
[Student Name]
Purdue University Global
FI311: FinTech Law and Ethics
Unit 7 Algorithmic Fairness Memo
[Instructor Name]
[Date]
The lender, its model and every figure are composite; the statute, regulation and articles cited are real. Not legal advice.
Memorandum
To: Credit Committee From: Fair Lending Analytics Date: March 3, 2026 Re: Approval disparities in the personal-loan model and a recommended replacement
Recommendation
Approve removing three features from the current underwriting model and adopt the alternative model tested in January. The change raises the approval ratio for applicants 62 and older from 0.79 to 0.91 of the younger rate and costs 0.4 points of predictive accuracy, which our loss forecasts absorb without changing pricing.
Findings
Our quarterly fair lending review compared approval rates across 214,000 applications from the past year. Applicants aged 62 and older were approved at 0.79 times the rate for applicants aged 25 to 61. Applicants living in ZIP codes where most residents are Hispanic, identified through a proxy method because we do not collect ethnicity for these loans, were approved at 0.83 times the rate for other applicants. We use a ratio of 0.80, borrowed from employment testing, as a screening threshold: below it we investigate, and the age result falls below it.
The model does not use age, ethnicity or ZIP code directly. A driver analysis traced most of the age gap to three features. The age of the device used to apply carries the most weight; older applicants use older phones and computers. Second, whether the phone number given is a landline. Third, whether the email address comes from an internet service provider rather than a free web mail service. Each of these predicts repayment slightly in our data, and each is strongly associated with age.
None of the three features was chosen to identify older applicants, but together they let the model find age without being told it, which is how a model can comply with a rule against using a characteristic and still produce the outcome the rule exists to prevent.
The Legal Standard
The Equal Credit Opportunity Act prohibits discrimination in any aspect of a credit transaction on the basis of age, national origin and other characteristics (Equal Credit Opportunity Act, 15 U.S.C. § 1691, 2018). Regulation B allows a statistically sound scoring system to consider age, but it may not assign a negative value to applicants 62 or older (Equal Credit Opportunity (Regulation B), 12 C.F.R. pt. 1002, 2024). Our model does not consider age directly, so that rule is not violated on its face.
The risk lies elsewhere. Regulators and courts have recognized that a neutral practice with a disproportionate adverse effect on a protected group can be unlawful unless it meets a legitimate business need that cannot reasonably be met by a less discriminatory alternative. The existence of an alternative model that performs nearly as well is exactly the fact that makes this risk serious.
Why This Matters Beyond the Law
Research on machine learning in credit has found that more flexible models can widen differences in outcomes between groups, even as they improve overall accuracy, because they find combinations of variables that track group membership (Fuster et al., 2022). Scholars of data-driven decisions have warned that such outcomes can arise without any intent, simply from patterns in historical data (Barocas & Selbst, 2016). At the same time, research comparing algorithmic and face-to-face lenders found that algorithms reduced but did not eliminate pricing differences for Latinx and African American borrowers (Bartlett et al., 2022). Models are not automatically fairer or less fair; they have to be tested.
The Alternative Model
The January alternative removes device age, landline status and email domain, and adds two features: the number of on-time rent or utility payments verified through bank data, and the stability of income deposits over twelve months. Its accuracy, measured by the area under the curve, falls from 0.742 to 0.738. Its approval ratio for applicants 62 and older rises to 0.91, and for applicants in predominantly Hispanic ZIP codes to 0.88. Projected losses rise by about 0.1 percentage point, within our pricing margin.
Remaining Questions
The Hispanic ZIP code ratio, at 0.88, passes our screen but remains a gap. We recommend a separate review of the features driving it. We also recommend that every new feature be tested for correlation with age and with neighborhood demographics before it enters any model.
References
Barocas, S., & Selbst, A. D. (2016). Big data's disparate impact. California Law Review, 104(3), 671-732. https://doi.org/10.15779/Z38BG31
Bartlett, R., Morse, A., Stanton, R., & Wallace, N. (2022). Consumer-lending discrimination in the FinTech era. Journal of Financial Economics, 143(1), 30-56. https://doi.org/10.1016/j.jfineco.2021.05.047
Equal Credit Opportunity Act, 15 U.S.C. § 1691 (2018).
Equal Credit Opportunity (Regulation B), 12 C.F.R. pt. 1002 (2024).
Fuster, A., Goldsmith-Pinkham, P., Ramadorai, T., & Walther, A. (2022). Predictably unequal? The effects of machine learning on credit markets. The Journal of Finance, 77(1), 5-47. https://doi.org/10.1111/jofi.13090
The Purdue Global FI311 Unit 7 instructions
Unit 7 of FI311 asks for a memo assessing an algorithm for unfair outcomes. Use memo headings and lead with your recommendation and its cost. Report the findings with the groups compared, the measure used and the threshold that triggers concern. Explain how you traced the outcome to specific features, since a model can produce a group effect without using the characteristic. State the governing law narrowly, including any special rules for the protected group, and explain the disparate impact theory and why an alternative matters. Bring in research on machine learning and discrimination. Describe the alternative with its accuracy cost and its effect on each group. Close with gaps that remain. Cite the statute, regulation and research in APA 7.
How the FI 311 Unit 7 example is put together
The sample's memo, dated March 3, 2026, recommends dropping three features and adopting a January alternative. Findings cover 214,000 applications: an age ratio of 0.79 and a ratio of 0.83 for predominantly Hispanic ZIP codes identified by proxy, against a 0.80 screen. A driver analysis traces the age gap to device age, landline status and internet-provider email. The highlighted sentence explains how neutral features rebuild age. The legal section notes Regulation B's bar on scoring applicants 62 and older negatively and the role of less discriminatory alternatives. The alternative adds verified rent payments and deposit stability, moves accuracy from 0.742 to 0.738 and lifts the ratios to 0.91 and 0.88.
FI311 Unit 7 rubric: what earns full points
Fairness memos in FI311 earn points for evidence, legal precision and a recommendation with its cost stated. Graders check that outcomes are measured by group, that the threshold is explained, that features driving a gap are identified and that the legal standard is stated accurately, including special rules for older applicants. The rubric rewards recognizing proxy discrimination, weighing an alternative's accuracy cost against its fairness gain and acknowledging remaining gaps. Memos that inspect inputs only, assume a model is fair because it omits protected traits or present the four-fifths ratio as law lose points. Research on algorithmic lending, used for what it found, supports the reasoning.
Common FI311 Unit 7 mistakes, and how to avoid them
Weak Unit 7 memos declare a model fair because it does not use age or race. Others find a gap but never trace it to features, so no fix is possible. Many overstate the law, treating a screening ratio as a legal line, or understate it, ignoring disparate impact. The alternative's accuracy cost is often left out, which makes the recommendation look free. Remaining disparities go unmentioned once the main one improves. When the model in your prompt sets insurance prices, screens tenants or ranks job applicants, the memo can follow the same outcome-first method under the law that applies there. The memo should also say who will monitor the new model after launch and how often.
Get an FI311 Unit 7 example written to your instructions
Share the model, the lender and any test results your Unit 7 prompt supplies, with the FI311 memo instructions and rubric. Who is disadvantaged gets measured first, then the responsible features are traced, the governing rule stated narrowly and each option costed. It remains a course memo rather than a fair lending review of any actual lender. First custom sample free, back within 24-48h. The paper above is an original model document written by our desk, not a submitted student paper and not an official Purdue University Global document.
More FI311 papers and BS in Finance samples
- FI311 Unit 1: Discussion Post on Fintech Complaints
- FI311 Unit 2: Fintech Regulatory Map Brief
- FI311 Unit 3: Wage Advance Compliance Analysis
- FI311 Unit 4: Fintech Data Privacy Case Study
- FI311 Unit 5: Pay-in-Four Consumer Protection Paper
- FI311 Unit 6: Fintech Enforcement Seminar Reflection
- FI311 Unit 8: Cross-border Regulation Analysis
- FI311 Unit 9: Ethics Scenario Response
- FI311 Unit 10: Regulatory Position Paper
- FI310 Unit 7: Instant Payments Risk Memo
FI311 Unit 7 questions, answered
How can a model discriminate by age if age is not an input?
Through proxies. Variables that look neutral, such as an older phone or a landline number, are far more common among older applicants, so a model that penalizes them penalizes age indirectly. The sample measures outcomes by age group rather than inspecting inputs, which is how the gap was found, and then works backward to the features producing it.
What is the four-fifths ratio, and is it a legal standard?
It compares the approval rate of one group with that of the most favored group, and a result below 0.80 is commonly treated as a signal worth investigating. It began as a rule of thumb in employment selection and is not a legal line in lending. The sample uses it as the lender's internal screening threshold and says so plainly.
Where can I find a free FI311 Unit 7 sample paper?
This page holds a full FI311 Unit 7 algorithmic fairness memo on a composite lender's model that approves older applicants less often, with the features traced, the law stated and APA 7 references.
Can a lending model consider age at all?
Under Regulation B, a statistically sound scoring system may consider age, but it may not assign a negative value to applicants 62 or older. The sample's model avoids age directly, yet its proxies produce a lower approval rate for that group.
Why adopt a model that is slightly less accurate?
Because the accuracy loss is small, losses stay within the lender's pricing margin, and the alternative closes most of the gap. When a less discriminatory alternative performs nearly as well, keeping the original model becomes hard to defend.