FIN 480 Week 2 AI, Big Data and Automation in Finance Example

Lending on Cash Flow, Not Credit Scores: How ShiftPay Built a Machine Learning Model for a Small-Dollar Loan, Tested It for Fairness and Explained Its Decisions to Applicants

[Student Name]

University of Phoenix

FIN/480: FinTech and DeFi

Week 2 Assignment

[Instructor Name]

[Date]

ShiftPay, its data and its model results are composites written for a model paper; methods, legal requirements and research findings come from the sources listed and are stated generally.

Many ShiftPay users, the composite Texas earned-wage startup's hourly workers, need more than an advance on a few days' pay: a $1,200 car repair, a security deposit, a medical bill. Many have thin or damaged credit files, so traditional lenders either decline them or charge high rates. ShiftPay decided to offer installment loans of $500 to $2,000, repaid over six to twelve months through payroll deduction, with its sponsor bank as lender. The startup believed its payroll data could predict repayment better than a credit score, but believing it and proving it fairly are different tasks. This paper describes how it built and tested the model.

The Data

ShiftPay holds data no credit bureau has: each user's hourly wage, hours worked each week, tenure with the employer, schedule stability and repayment of past advances. With consent, applicants can also connect their bank accounts, adding information on income deposits, recurring bills, overdrafts and balances. Credit bureau reports provide traditional data. Some variables, such as zip code, were excluded from the start because of their close link to race.

Two Models

A traditional scorecard, built with logistic regression on 14 variables, gives each applicant points for each characteristic, such as tenure above one year, and approves those above a cutoff. A gradient-boosted tree model, a common machine learning method, combines hundreds of small decision trees to capture interactions, such as the way irregular hours matter more for short-tenure workers. Both were trained on 38,000 past advance users who had taken a pilot loan, using 70 percent of the data to build and 30 percent held out to test.

Measuring Accuracy

On the held-out data, the scorecard achieved an area under the curve, a measure of how well a model ranks good and bad borrowers, of 0.71. The machine learning model reached 0.78. At the same approval rate, the machine learning model would have cut the default rate from 9.4 percent to 7.1 percent, or, at the same default rate, approved about 18 percent more applicants.

What Research Shows

Berg et al. (2020) found that even simple digital footprints, such as the device and email provider an online shopper used, predicted default as well as a credit score and added information when combined with it. Fuster et al. (2022) studied machine learning in mortgage lending and found that it increased overall accuracy but that the gains were unevenly distributed, with some minority groups less likely to benefit than white and Asian borrowers. Their finding warns that a more accurate model can still widen gaps.

Testing for Fair Outcomes

Federal fair lending rules bar lenders from treating applicants differently because of race, sex, national origin or other protected traits, and they reach policies whose effects fall disproportionately on a group without a business justification. ShiftPay does not collect race for its loans, so its analysts estimated it using a standard method that combines surname and geography, applied only for testing. Approval rates under the machine learning model were 64 percent for white applicants, 58 percent for Hispanic applicants and 55 percent for Black applicants, compared with 57, 49 and 46 percent under the scorecard. The new model raised approvals for every group and narrowed the gap slightly. Among approved borrowers, default rates were similar across groups, suggesting the model was not systematically wrong for any group.

A Variable Removed

The analysts found that one bank account variable, frequency of payments to check-cashing stores, strongly affected decisions and was concentrated in some neighborhoods. Removing it reduced accuracy only slightly while narrowing the approval gap by two points, so it was dropped. Bartlett et al. (2022) found that FinTech mortgage lenders discriminated less in pricing than face-to-face lenders but that algorithmic discrimination remained, which supports this kind of testing.

Explaining Decisions

Federal law requires lenders to tell denied applicants the specific principal reasons for the decision. The Consumer Financial Protection Bureau has stated that this requirement applies even when lenders use complex algorithms, and that a lender cannot use a model whose decisions it cannot explain. ShiftPay uses a method that measures each variable's contribution to an applicant's score, then maps the largest negative contributions to plain reasons, such as hours worked varied widely in the last eight weeks or less than six months with the current employer.

Automation and Human Review

The model approves or declines most applications instantly. Applications near the cutoff, about 8 percent, go to a trained underwriter who can consider information the model cannot, such as a recent promotion. Any applicant can ask for a manual review.

Governance

A validation team independent of the model builders tested the model before launch. Performance, approval rates and default rates by estimated group are monitored monthly, and the model is retrained no more than twice a year so that changes can be reviewed. A written record documents each variable, the reasons for including it and the fairness tests performed.

What Could Go Wrong Later

A model that is fair and accurate at launch can drift. If an employer client changes its scheduling software, hours data may shift in ways the model misreads; if a recession lengthens unemployment spells, past repayment patterns may no longer hold. ShiftPay set triggers for an early review: a rise of more than two points in the default rate, a change of more than three points in any group's approval rate or a new data source from an employer. Monitoring dashboards show those figures weekly to the credit risk team. Planning for drift is part of responsible automation, since an unattended model can harm borrowers long before anyone notices.

Privacy

Bank account data are used only with consent, only for the loan decision and are deleted after a set period if the applicant does not proceed.

Conclusion

ShiftPay's machine learning model, built on payroll and cash flow data, ranked borrowers more accurately than a traditional scorecard and expanded approvals for every group tested. Fairness testing led to removing a variable, explanation methods meet the requirement to give specific reasons and governance keeps people responsible for the model's decisions. Accuracy and fairness were treated as two parts of one design.

References

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

Berg, T., Burg, V., Gombović, A., & Puri, M. (2020). On the rise of FinTechs: Credit scoring using digital footprints. The Review of Financial Studies, 33(7), 2845-2897. https://doi.org/10.1093/rfs/hhz099

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