CreditMetrics and Portfolio Credit Modelling

Add up the expected loss on each bond in a portfolio separately and you get the portfolio's average loss right, but you learn nothing about how bad a genuinely bad year could be. Bonds in the same industry, or in the same economy, tend to get downgraded together, not independently — a recession doesn't hit one company at a time. CreditMetrics, developed by J.P. Morgan in 1997, was the first widely adopted framework to simulate a whole credit portfolio's outcomes jointly rather than bond by bond.

How the simulation works

For each bond, CreditMetrics starts from a historical rating transition matrix that gives the probability of ending the year at every possible rating, including default. Rather than drawing each bond's outcome independently, it models an unobserved "asset value" for each issuer that drives the migration, and correlates those asset values across issuers who share industry or macro exposure — the same modeling trick used in the single-factor Vasicek default model. When a shared macro factor draws badly, correlated issuers are more likely to migrate downward together, replicating the real-world clustering of downgrades seen in recessions.

Worked example (simplified two-bond case)

A portfolio holds two BBB-rated bonds in the same industry, each with a 3% one-year probability of downgrade to junk and a further 0.3% probability of outright default, based on the transition matrix. Modeled independently, the chance both downgrade in the same year would be $3\% \times 3\% = 0.09\%$. But because CreditMetrics links both bonds to a shared industry asset-value factor with a correlation of, say, 0.4, the actual simulated joint-downgrade probability comes out closer to 0.6–0.8% — six to nine times higher than the naive independent calculation, because a bad draw on the shared factor pushes both bonds toward migration together. Running this logic across a full portfolio of a hundred bonds, correlated in various industry and regional clusters, produces a simulated loss distribution with a noticeably fatter tail than an independent-default model would show, even though the average expected loss across the two approaches looks similar.

What this means in practice

Portfolio credit VaR, economic capital, and regulatory stress tests all depend on getting this correlation structure right — a bank that models its loan book as a set of independent default risks will systematically underestimate the capital it needs to survive a correlated, industry-wide or economy-wide downturn.