Default Correlation and the Asset Threshold Model

Two unrelated companies — a retailer and an airline — rarely default because of the same specific event. But both borrow against a backdrop of the same interest rates, the same consumer spending, the same recessions. Their defaults are not independent coin flips; they are more likely to happen in the same bad years than pure chance would suggest. The asset threshold model (also called the Merton-style single-factor or Vasicek model) is the standard way to turn that intuition into an actual number for default correlation.

The setup

Each firm $i$ has a standardized asset value:

$$X_i = \rho \, M + \sqrt{1 - \rho^2}\, \varepsilon_i$$

Here $M$ is a single common factor (think: the state of the economy) shared by every firm, $\varepsilon_i$ is a firm-specific shock independent across firms, and $\rho$ measures how much of firm $i$'s asset value is driven by the common factor versus its own idiosyncratic noise. Firm $i$ defaults if $X_i$ falls below a threshold $c_i$ set so that the unconditional probability $P(X_i < c_i)$ equals the firm's known default probability, $PD_i$.

In plain words: every firm's fortune is part shared economic weather, part company-specific luck. When $\rho$ is high, firms are mostly buffeted by the same weather and their defaults cluster; when $\rho$ is low, each firm's outcome is mostly its own idiosyncratic story and defaults look close to independent.

Worked example: computing conditional default probability

Take a firm with an unconditional annual PD of 2 percent and an asset correlation $\rho = 0.20$ — a typical value used for large corporates under Basel-style formulas. Using the standard normal inverse, the default threshold satisfies $\Phi(c) = 0.02$, giving $c \approx -2.054$.

Now condition on a bad economic state, $M = -2$ (a two-standard-deviation bad year). The conditional default probability is:

$$PD(M=-2) = \Phi\left(\frac{c - \rho M}{\sqrt{1-\rho^2}}\right) = \Phi\left(\frac{-2.054 - 0.20 \times (-2)}{\sqrt{1 - 0.04}}\right) = \Phi\left(\frac{-1.654}{0.980}\right) = \Phi(-1.688) \approx 0.046$$

So a 2 percent through-the-cycle PD becomes roughly 4.6 percent conditional on a two-standard-deviation bad year — more than double. Every firm sharing that same $\rho$ sees its PD move the same direction in the same bad year, which is exactly the mechanism that produces correlated, clustered defaults across a portfolio rather than independent ones.

Worked example: two firms, same bad year

Take two firms both with $PD = 2\%$ and $\rho = 0.20$. Their pairwise default correlation (the correlation of the 0/1 default indicators, not of the asset values themselves) works out, from the joint bivariate normal probability of both falling below threshold, to roughly 1–2 percent for typical corporate parameters — a modest-looking number that nonetheless has an outsized effect on portfolio tail risk, because correlation compounds across a large book far more than it appears to pairwise. This is the exact mechanism that, scaled up across a hundred correlated mezzanine tranches, produced the tail-risk blowup described in ABS CDOs and the 2008 Correlation Failure: individually modest pairwise correlations still translate into a fat tail of joint defaults for the portfolio as a whole.

Where you meet it in practice

The single-factor asset threshold model underlies the Basel regulatory capital formulas banks use to compute risk-weighted assets, the Vasicek portfolio loss distribution used in structured credit, and the intuition behind every "systematic versus idiosyncratic" split in credit risk. Whenever a risk report distinguishes a firm's stand-alone default probability from its behavior "in a downturn," this model is what is being used underneath.