Interactive model of how borrowers, bank partners, capital partners, and Upstart's AI clearing engine interact
📊 See Data & Methods for sources, data generation logic, and case studies.
Upstart: S-1 filing (Dec 2020), Q1–Q2 2022 earnings calls
Industry: Federal Reserve consumer credit, TransUnion research
Academic: XGBoost (Chen & Guestrin 2016), Fairness in ML (Barocas et al. 2021)
Full details: See Data & Methods page →
From first Google search to final payment — 9 stages, each with a different product, data requirement, and failure mode. Most borrowers don't experience all 9 in sequence; the drop-offs at each stage define the funnel that the Marketplace Optimization team owns.
100% — visit rate check page
↓ 65% — complete pre-qualification
↓ 45% — receive a rate offer (qualify)
↓ 27% — accept offer (hard pull)
↓ 24% — complete verification
↓ 23% — funded ← this is what you optimize
The simulator models Stages 03–07: AI assessment → offer → acceptance → clearing. It does not model acquisition (traffic), verification friction, or the repayment lifecycle. Those stages are equally important — just owned by different teams. The Marketplace Optimization PM owns the clearing mechanism at the center of this funnel.
Capital partners are the supply side of the marketplace. They provide the funding that makes every loan possible. There are four structural types — each with different incentives, risk tolerances, and relationship dynamics. The Marketplace Optimization team must simultaneously satisfy all of them.
Pre-committed volume. Get first allocation in waterfall. Upstart's most stable capital source.
No commitment — buy individual loans on demand. Require higher APR floor to compensate for no volume guarantee.
Lower cost of capital (deposit funded). Higher FICO requirements. CRA credit benefit for serving underbanked.
Buy rated bonds backed by loan pools. Require pool size, rating, and credit enhancement. Longer setup time.
Being routed the worst loans while better loans go to spot market buyers. This happens when waterfall routing logic is misconfigured. The fix: transparent eligibility matrices + balanced allocation across tiers. If a partner suspects adverse selection, they leave — and they tell others.
When actual default rates diverge upward from predicted. Can happen from: model drift (borrower population shifted), channel mix change (riskier acquisition), or macro shock (rate spike). EPD is the fastest signal — loans defaulting within 90 days are the strongest model quality indicator.
The bank is not optional — it's structurally required by US banking law. Understanding why, and what the bank actually does in each transaction, is essential for understanding why Upstart's marketplace is built the way it is.
In the United States, only a federally or state-chartered bank can legally originate a consumer loan. A hedge fund, investment firm, or fintech platform — no matter how sophisticated — cannot make a loan directly to a consumer. They can own loans. They cannot make them.
The second reason is even more important for marketplace economics: a national bank can "export" its home state's interest rate laws to any borrower in any state (the "valid when made" doctrine, from a 1978 Supreme Court case). Upstart partners with banks chartered in states with no usury cap (APR ceiling). This lets Upstart's AI model operate as one unified marketplace with consistent pricing across all 50 states. Without bank partners:
The bank is not a rubber stamp. It must genuinely underwrite and approve each loan, maintain fair lending documentation, and is subject to regulatory examination on its third-party lending program. It earns an origination fee (~1–3% of loan amount) and has zero credit risk — it sells the loan immediately.
Role: Legally makes the loan. Issues disclosures. Holds briefly.
Earns: Origination fee (1–3% per loan). Zero credit risk.
Regulated by: OCC, FDIC, CFPB. Subject to exam.
Cares about: Regulatory compliance, not losing charter, fair lending documentation.
Role: Buys the loan from the bank. Takes on credit risk. Earns interest over 36–60 months.
Earns: Interest income minus defaults. Net 7–11% target return.
Regulated by: SEC (if fund). Not subject to banking regulation.
Cares about: Model accuracy, adverse selection, EPD rates, consistent volume.
This is what the Marketplace Optimization PM owns. Three layers — eligibility, pricing, routing — each with distinct logic, distinct failure modes, and distinct PM levers. A loan must pass all three to clear.
Which capital partners can fund this loan? Binary checks — no ML here.
Novel: APR is fed back into the risk model as an input, not just an output.
Given a clearing APR, which capital partner gets this loan? Priority-based routing.
Model prices loan at 28% APR. Borrower's tolerance is 24%. Loan doesn't clear — borrower rejects. Demand-side failure.
Borrower's FICO (610) or loan purpose (small business) doesn't meet any partner's criteria. Supply-side failure — no price fixes this.
Eltura hits their monthly commitment cap. Loan falls to spot market (higher APR required) or balance sheet. Clearing rate doesn't drop — but quality of placement does.
Upstart model approves the loan. Bank partner's own credit policy rejects it. Creates "late decline" — worst UX (borrower thought they were approved).
Routing error sends worst-performing loans to forward-flow partners, best to spot market. Partners see EPD spike. Begin exit process. Capital supply collapses.
What this simulator models accurately, what it simplifies, and what real Upstart does differently. Knowing the gap between a model and reality is as important as knowing what the model shows.