QuarcHome
Model catalogueBack to the model index

CreditMetrics

Simulates ratings migrations and portfolio repricing to estimate mark-to-market credit losses.

What the model measures

CreditMetrics estimates how the value of a credit portfolio could change over a chosen horizon, commonly one year. A bond can lose value when its issuer is downgraded and the spread widens, even if every scheduled payment is made. The output is a distribution of horizon portfolio values. Set a clear reference value before converting it to losses or quoting value at risk.

From rating migrations to portfolio values

Start with a transition matrix. The row for an issuer’s current rating gives the probabilities of the possible horizon ratings, including default. For each surviving rating, value the remaining cash flows with a forward discount curve for that rating at the horizon. Include coupons received before the horizon according to the chosen holding-period convention. In default, use an explicit recovery rule appropriate to the claim’s seniority and specify how recovery timing is handled.

For multiple issuers, simulate ratings jointly. In the original framework a standardised latent asset return crosses thresholds implied by each rating row. Correlating the latent returns makes migrations cluster. All claims on one issuer share its migration state, but each instrument keeps its own exposure, seniority and recovery assumptions. Sum the marked values within each scenario before taking quantiles. Averaging individual VaRs would lose information about concentration and diversification.

Worked example · downgrade losses without default

Take a synthetic bond with a reference value of 100 at the horizon. The marks below already reflect each scenario's credit spread or recovery. For clarity the example ignores carry, intervening coupons and market-rate shocks.

Horizon stateProbabilityValueLoss against 100
Upgrade5%103−3
Unchanged88%1000
Downgrade6%928
Default1%4060

The probability-weighted value is 0.05 × 103 + 0.88 × 100 + 0.06 × 92 + 0.01 × 40 = 99.07. The first two states bring cumulative probability to 93%. Adding the downgrade brings it to 99%, so 95% loss VaR is 8. The worst 5% consists of the 1% default state and four percentage points from the downgrade state. Expected Shortfall is (0.01 × 60 + 0.04 × 8) ÷ 0.05 = 18.40. If loss were recorded only on default in this example, 95% VaR would be zero.

Which assumptions to challenge

  • Check that each transition row sums to one and document treatment of withdrawn and missing ratings. Segment observations by period or sector if a pooled matrix hides materially different migration behaviour.
  • Validate both migration frequencies and horizon repricing. Look separately at spread movements within a rating. A single curve per rating can miss material difference among issuers carrying the same label.
  • Stress transition probabilities, recoveries and dependence together. Using two otherwise identical bonds, doubling one issuer exposure and adding a second issuer give different tails even when each bond has the same marginal probabilities.
  • Check quantile and tail-mean convergence across simulation runs. Keep the loss baseline and sign convention fixed. A small simulated standard error does not test whether correlations or recovery assumptions are realistic.

Further study

Chapter 2 of the original CreditMetrics technical document works through stand-alone migration and valuation. Chapters 6–8 develop probability, recovery and dependence inputs. Compare Jarrow–Lando–Turnbull for rating-state pricing, CreditPortfolioView for macro-sensitive transitions and CreditRisk+ for a default-loss portfolio model.

Research sources