Back to Projects

Upstart Lifecycle Simulator

Live

Browser-based simulator for Upstart's capital marketplace — generate borrower pipelines, run side-by-side Model 18 vs Classic clearing, compare 36-month portfolio performance across market scenarios, and walk through individual loan decisions step-by-step.

Vanilla JS Tailwind CSS Jinja2 FastAPI
GitHub Live Demo
❌

The Problem

Marketplace clearing decisions have a 90-day feedback delay — a PM changes partner eligibility today but doesn't see the EPD impact for months. The 2022 capital crunch showed what happens when this delay prevents preventive action: Upstart held hundreds of millions in loans on its balance sheet.

💡

The Approach

Build an interactive lifecycle simulator that compresses 36 months of portfolio evolution into seconds. Generate synthetic borrowers, clear them through a three-layer engine with side-by-side model comparison, then compare performance across scenarios.

✅

The Solution

A browser-based sandbox with 7 tabs: PRFAQ, User Manual, Pipeline & Clearing, Marketplace Performance, Loan Deep Dive, Personas, and Re-Application. All client-side JS — no backend computation, no database, no API calls.

What

A full lifecycle simulator for Upstart’s capital marketplace — from borrower pipeline through clearing to 36 months of portfolio performance.

Seven Tabs

1. PRFAQ Amazon-style press release, customer FAQ, and internal FAQ framing the problem, the solution, and the v2 roadmap.

2. User Manual Methodology behind the simulator (borrower generation, three-layer clearing, Markov chain lifecycle) plus step-by-step usage guide and disclaimers.

3. Pipeline & Clearing Generate 1–100 synthetic borrowers, then run them through the clearing engine. Results show a 13-column side-by-side table: FICO, Model 18 Score, amount, purpose, hidden-prime flag, then APR/outcome/partner for both Model 18 and Classic. Filter by outcome (Cleared, APR Rejected, No Partner).

4. Marketplace Performance Fixed 100-loan baseline (20 per partner) run through 36-month lifecycle under all three scenarios simultaneously. Compare Healthy Market vs Capital Crunch vs Rate Spike side-by-side with partner filtering and month scrubbing.

5. Loan Deep Dive Click “Walk” on any borrower to see their 4-step journey: eligibility matrix (with failure reasons per partner) → pricing comparison → waterfall routing → funding & payment history strip.

6. Personas Three borrower archetypes — Maria (hidden-prime), Carlos (ineligible/subprime), James (prime/balance sheet) — explaining the risk segments the simulator generates from.

7. Re-Application Product roadmap stub framing how lifecycle data can predict re-applications and lower CAC for returning borrowers.

Key Features

  • Side-by-side Model Comparison — Every loan shows Model 18 and Classic outcomes in the same row
  • Scenario Presets — Healthy Market, Capital Crunch (2022), Rate Spike — each adjusts FICO distribution and lifecycle parameters
  • Hidden-Prime Discovery — Model 18 Score = FICO + 50 for hidden-prime borrowers, visible in every table
  • NO_PARTNER Explanations — Deep Dive shows exactly why a borrower failed: FICO too low, capacity exhausted, APR mismatch

Why

Upstart’s capital marketplace is one of the most sophisticated clearing engines in fintech — but the feedback loops are invisible.

The Problem

  • 90-day feedback delay — Changes to clearing rules take months to show in portfolio performance
  • 2022 lesson — The capital crunch happened partly because decision-makers couldn’t see the compounding effect of tighter partner capacity in real-time
  • Black box clearing — Few people understand how FICO pricing → Model 18 → partner routing → marketplace economics all connect

What This Solves

  • Compress time — 36 months of portfolio evolution, instant
  • Make tradeoffs tangible — Clearing rate vs partner EPD vs balance sheet exposure, visible in one view
  • Three-stakeholder thinking — Borrower experience, platform economics, and partner returns all in one tool
  • Model 18 intuition — See exactly how APR-as-feature discovers hidden-prime borrowers and where the value is created

Who It’s For

PM candidates learning marketplace dynamics, marketplace operators stress-testing decisions, data scientists understanding the pipeline, anyone curious how lending marketplaces actually work.

How

Architecture

Client-Side Only - Entirely browser-based JavaScript — no backend, no database, no network latency - 100 borrowers × 5 partners × 36 months × 3 scenarios computed on page load - FastAPI serves the HTML template; all computation happens in the browser

Modules - borrower_generation.js — Synthetic borrower creation with FICO distributions, hidden-prime flags, seeded PRNG - clearing_engine.js — Three-layer engine: eligibility → pricing → waterfall routing - lifecycle_engine.js — 36-month Markov chain with grade-based transition probabilities - lifecycle_simulator.js — Main coordinator: pipeline controls, side-by-side rendering, deep dive, marketplace performance

User Flow

  1. Read PRFAQ — Understand the problem and context
  2. Check User Manual — Review methodology and step-by-step instructions
  3. Select Scenario — Choose Healthy Market, Capital Crunch, or Rate Spike
  4. Generate Pipeline — Create 1–100 synthetic borrowers
  5. Run Clearing — Process through both models; review side-by-side results
  6. Deep Dive — Click “Walk” on any borrower to trace their clearing journey
  7. Compare Scenarios — Switch to Marketplace Performance to see baseline comparison

Limitations

  • Illustrative, not production-grade — Markov chain is directionally correct but not vintage-calibrated
  • Small scale — 100-borrower demos show dynamics, not statistical significance
  • No persistence — Page refresh resets everything; no backend storage
  • Obfuscated partners — Names are fictional; structures inspired by public filings
  • Re-Application is a stub — Product roadmap framing, not full simulation
Back to Projects