Supporting note to The Sufficiency of Upstart’s Capital Is Vulnerable to a Few Unexpected Points of Additional Losses. It explains what level of tail risk bank capital standards are built to absorb, and what loss shock those standards imply for a loan book like Upstart’s.

1. The tail risk capital standards target

• Basel risk weights: a 1-in-1,000-year loss over one year. The credit risk formulas behind U.S. and international risk weights set capital at a 99.9 percent confidence level over a one-year horizon. The Basel Committee itself cautions that the high confidence level partly compensates for estimation error and for the weaker loss-absorbing capacity of Tier 2 capital, so the target is a very remote loss rather than literally a once-in-a-millennium event.
• Federal Reserve stress tests: a severe postwar recession. The Fed’s severely adverse scenario is designed to reflect conditions in severe postwar U.S. recessions, with unemployment rising 3 to 5 percentage points and to at least 10 percent. Firms must remain above regulatory minimums after the shock. Commentators place that severity around the 95th percentile of outcomes, roughly a 1-in-20 to 1-in-25-year event. In recent stress tests the scenario produced nine-quarter credit card losses of about 17 percent, roughly twice normal.
• Rating agencies: multiples of expected loss. S&P’s AAA stress for subprime auto securitizations is 2.0 to 3.5 times expected loss, with lower multiples for pools with higher expected losses. This is the securitization counterpart of a 99.9-percent-plus standard.

2. Method

The shock is computed with the same one-factor credit model that underlies Basel risk weights (the asymptotic single risk factor, or Vasicek, model). The stressed default rate at confidence level q is N[(G(PD) + √R × G(q)) / √(1 − R)], where N is the standard normal distribution, G its inverse, PD the expected default rate and R the asset correlation. Inputs for Upstart’s book: expected lifetime cumulative net loss (CNL) of 16.5 percent, loss given default of 90 percent (so a lifetime default rate of about 18.3 percent), and a weighted-average life of about 1.8 years.

Correlation is the key input. At this default rate, the Basel “other retail” formula falls to its floor of about 3 percent, on the premise that high-default borrowers default for idiosyncratic reasons. That premise is weakest for a book concentrated in near-prime and subprime unsecured borrowers, so results are also shown at 6 percent.

3. Implied loss shock

History lines up with the model. The stress-test-level shock of about 8 to 12 points matches what was observed in severe downturns: charge-offs on HSBC Finance’s unsecured personal loans roughly doubled in 2008 and 2009 (about 11 to 15 points on this base), and Upstart’s own macro index peaked 68 percent above its baseline in January 2024 (about 11 points). The Basel standard sits above anything observed, as intended.

4. What the shock means for holding-company capital

The table applies each shock to Upstart’s consolidated position using the same favorable treatment as the main post: Simplified Supervisory Formula Approach (SSFA) risk weights, about $90 million of loss per point (about $85 million on beneficial interests plus $5.5 million on loans), and beneficial-interest losses capped at Upstart’s contractual exposure of about $1.2 billion. The “CET1 needed” column is the starting Common Equity Tier 1 capital required to end the shock at the 4.5 percent CET1 minimum, the test the Fed applies after its stress scenario.

Two points follow. First, at the stress-test standard, a shock of 8 to 12 points consumes essentially all of the holding company’s current CET1, and staying above the minimum would take roughly $1.0 billion to $1.35 billion of CET1 before credit for earnings. Second, at the Basel standard, losses reach the contractual cap on the beneficial interests, which means those positions warrant capital equal to Upstart’s full exposure. That is the economic result a 1,250 percent risk weight produces, and it is why we do not think the SSFA is the right model for these positions.

5. Caveats

• These are results of the regulators’ own one-factor model, not forecasts.
• Converting a one-year standard to a lifetime figure is an approximation; a stressed year affects the loans outstanding during it, and Upstart’s 1.8-year average life limits how many vintages are exposed.
• Beyond about 4 points, the dollar figures extend Upstart’s disclosed 10 percent and 20 percent sensitivities in a straight line, up to the contractual cap.
• Upstart is far below the asset size at which the Fed’s supervisory stress tests apply. The stress-test standard is used here as the Fed’s own benchmark for a severe recession, not as a requirement that formally applies.

Sources

• Basel Committee, An Explanatory Note on the Basel II IRB Risk Weight Functions (2005)
• Federal Reserve, Basel II notice of proposed rulemaking (2006): 99.9 percent confidence level and one-year horizon
• Federal Reserve, Amendments to Policy Statement on the Scenario Design Framework for Stress Testing (2019)
• GARP, Stress Testing in Volatile Times (2021)
• S&P Global Ratings, Global Auto ABS Methodology and Assumptions
• HSBC Finance Form 10-Q, September 30, 2009
• Upstart Form 10-Q, September 30, 2026

Return to the main post