Fairplay
Fairplay Technical Briefing
Published March 7, 2023
Supported by IowaEDA and Brale
Overview
Fairplay founder and CEO Kareem Saleh walks through "fairness as a service" for any algorithm making high-stakes decisions. The briefing demos customer composition and demographic imputation, proxy detection on a bias-vs-predictive-power matrix, fair lending analysis across women, Black, Hispanic, API, and American Indian applicants with Shapley-based driver attribution, census-tract-level geographic maps, a Fairness Optimizer that adds $1M of profit while closing Black-applicant approval gaps, SR 11-7 model risk documentation, and second-look models that surface thousands of additional approvals from declined applications.
0:00 Fairness as a service
Kareem Saleh introduces Fairplay, "the world's first fairness-as-a-service company." Tools for any high-stakes decision algorithm to answer five questions: is it fair, why not, can it be fairer, what's the economic impact, and did we reconsider declines. Customers include Figure, Octane, Happy Money, Splash Financial, and LendingPoint.
1:18 Getting started and data upload
fairplay.ai login with GCP, Azure, or AWS integrations (or direct accounts). Upload a non-mortgage, mortgage, or marketing dataset — inputs, outputs, and outcomes — to unlock five product functions on the dashboard.
2:35 Customer composition and demographic imputation
Using first name, last name, and street address to approximate race, gender, and age for every applicant — replacing "fairness through blindness" with actual measurement. A top-10 bank took nine months to build this internally; Fairplay exposes it as a one-click function.
3:27 Proxy detection: bias vs predictive power
Every variable in the algorithm plotted on a 2D matrix — bias (predictiveness of protected status) on the x-axis, predictive power on the y-axis. Credit score is highly predictive of default but also of being Black or Hispanic; lower-bias substitutes are surfaced.
6:04 Fair lending analysis
Evaluating fairness of underwriting (or pricing, line assignment, fraud, marketing) decisions. Summary cards per protected group show adverse-impact ratios against current, previous, and credit-score-only baselines — with notes on the different fairness definitions preferred by CFPB, OCC, and FDIC.
7:22 Protected groups and Shapley-based drivers
Per-group views for women, Black, Hispanic, API, and American Indian applicants. "Drivers of disparity" uses Shapley values — e.g. FICO explains 18%, borrower income vs MSA median 12%, and co-applicant Vantage score 8% of the disparity exhibited by Black applicants.
9:07 Geographic fairness mapping
Census-tract-level maps of mortgage fairness for every U.S. originator. Example: Wells Fargo in Miami-Dade census tract 114.01 is approving Black applicants at 43% the rate of white applicants.
10:28 Executive presentations and Fairness Optimizer
Downloadable executive reports designed by Seth Feaster (ex-NYT data viz lead), almost all visual. Fairness Optimizer plots alternative models on a profitability-vs-fairness chart — finds an algorithm that approves more Black applicants and adds $1M in profit while holding the loan book constant.
13:07 SR 11-7 model risk documentation
In-time and out-of-time stress tests, modeling technique, and variables-used reports — the full SR 11-7-style model risk management documentation required by federal financial regulators, generated with one click.
13:33 Second-look models
Re-underwriting declined applications with a fairness-tuned second-look model. Accepting a 3% accuracy diminution surfaces ~5,000 additional loans, grows the book by $100M, and lifts profitability ~11% — with population-characteristics views revealing higher-DTI, lower-FICO, higher-LTV applicants the incumbent missed.
Presented by Kareem Saleh — Fairplay
Topics: Compliance & Regulation, Lending
