Core Path · L6.4.03

Estimating default correlation

14 minDefault correlationAsset correlationMaximum likelihoodMethod of momentsProfile likelihoodSmall-sample bias

By the end you will be able to

  1. 1Estimate the asset correlation from yearly default counts by the method of moments and the Vasicek density, and say what each assumes.
  2. 2Fit the binomial-mixture likelihood by grade and pooled, with standard errors, profile intervals and a likelihood-ratio test.
  3. 3Explain the small-sample bias and why ten years of one bank's defaults rarely pin ρ down.
  4. 4Tell asset correlation from default correlation, compute the implied default correlation of two names, and describe the equity route.

Why it matters

Every loss distribution in this module rests on one number: the corporates' asset correlation of 0.2. The bank has ten years of its own defaults, 598 of them in 22,800 obligor-years. Fit the model to them and ρ comes out at 0.065, and the data reject 0.2. Is the booked figure wrong, or is the estimate? Both questions have honest answers, and the answers are the lesson.