Risk tiers
Loans with a fraud probability of 0.2 or more
Flag rate by state
Share of loans with probability 0.5 or more
Score distribution
Loans per 0.025 band of fraud probability, log scale; dashed line is the review threshold
Why loans were flagged
Red-flag rules fired among flagged loans (a loan can fire several)
Synthetic fraud test
Share of 300 planted fraudulent loans per type ranked in the top 1% of real loans. * no rule targets this type
Key findings
- The payroll formula is the strongest public signal. 3.3% of loans under $150k exceed the $20,833 per job ceiling the program allowed.
- Jobs data has lender-level gaps. Wells Fargo, Regions and Celtic Bank left jobs blank or zero on 80 to 100% of loans; the model treats those zeros as unreported rather than as no employees.
- Well-disguised fraud is invisible in public data. Planted loans that copy a normal profile scored near the bottom; payroll, identity and bank data are needed to catch them.
- Stable and calibrated. The top 1% of the queue overlaps 90% across random seeds, and the model stays on the label model's probability scale.
What would make it better
- Confirmed outcomes from the SBA (hold-code resolutions, OIG and DOJ dispositions) keyed by loan number, to validate and calibrate the score.
- IRS payroll and EIN checks (Forms 941 and W-2, Schedule C, EIN assignment date) to test claimed payroll directly.
- Identity and network data: Do Not Pay, SSA verification, and EIDL bank, IP and device links to expose fraud rings.
- A secure enclave with computer matching agreements, to use these data lawfully.
Leads, not findings. A score measures how closely a loan resembles known fraud patterns using public
data alone. Every flagged loan needs human review against source records. Demographic fields are never model inputs;
a fairness review of flag-rate gaps is required before operational use. Details are in the full report, ppp_fraud_report.pdf.