How does the Gaussian copula model default time correlation, and why was it controversial?

The Gaussian copula model was the industry standard for pricing CDOs and other correlation-dependent credit products before the 2008 crisis. Understanding both its mechanics and its flaws is heavily tested in FRM Part II.

The Core Idea

Each borrower i has a latent variable ("asset value") that drives default:

X_i = sqrt(rho) x Z + sqrt(1-rho) x epsilon_i

Where:

• Z = common systematic factor (economy-wide, standard normal)
• epsilon_i = idiosyncratic factor specific to borrower i (standard normal)
• rho = asset correlation parameter (the KEY input)
• X_i = latent variable; borrower i defaults if X_i < default threshold C_i

The threshold C_i is calibrated so that the marginal default probability matches the CDS-implied or rating-implied PD:

C_i = N^{-1}(PD_i) where N^{-1} is the inverse standard normal CDF.

How Correlation Enters

The parameter rho determines how much defaults move together:

• rho = 0: Defaults are independent. Diversification is maximum.
• rho = 1: All borrowers default or survive together. No diversification.
• rho = 0.3 (typical): Moderate correlation. Enough clustering to create fat tails in the portfolio loss distribution.

CDO Tranche Pricing

The Gaussian copula was used to price CDO tranches by:

1. Simulating many scenarios of correlated defaults using the factor model
2. Computing losses for the reference portfolio in each scenario
3. Allocating losses to tranches (equity takes first loss, senior takes last)
4. Discounting expected cash flows to each tranche

What Went Wrong

1. Single correlation parameter: The model used one rho for the entire portfolio. In reality, correlations are heterogeneous and regime-dependent.
2. Gaussian tails too thin: The normal distribution underestimates the probability of extreme joint defaults. When markets crashed, far more borrowers defaulted simultaneously than the model predicted.
3. Correlation was not constant: Rho spiked during the crisis — the very time when accurate correlation estimates mattered most.
4. False precision: Market participants treated the model-implied "correlation smile" as a reliable pricing tool, ignoring the fact that the underlying assumptions were fragile.
5. Calibration to CDS spreads: When CDS spreads diverged from reality (due to illiquidity), the model inherited those distortions.

Alternatives

• Student-t copula — fatter tails, captures extreme co-movement better
• Clayton copula — asymmetric, captures lower-tail dependence
• Stochastic correlation models — rho itself is a random variable

For more on structured credit and copula models, explore our FRM Part II course.