Core Path · L6.3.06

Estimating migration matrices

14 minMigration matrixCohort methodDuration methodGeneratorCredit cycleLifetime PD

By the end you will be able to

  1. 1Estimate a one-year migration matrix from a rated panel by the cohort method and by the duration method, and explain why their PDs differ (M-C13: Markov chains and transition matrices).
  2. 2Turn a generator into a matrix for any horizon, test a cohort matrix for a valid generator (the embedding problem) and repair it.
  3. 3Shift a pooled matrix to a year's credit cycle with one factor Z, and fit Z to each year of the panel.
  4. 4Build lifetime PDs from matrix powers and from the generator, with bootstrap bands, beside the booked master scale.

Why it matters

L3.3.07: Rating migration raised a one-year matrix to the fifth power and read off five-year PDs. The matrix was handed down. Now estimate one from the bank's own panel of 22,800 obligor-years, and two honest methods disagree: grade 10's one-year PD is 32.1% by one and 42.3% by the other, and the best grade's is zero or 0.4 in a million. Lifetime provisions, loan pricing and portfolio models all start from such a matrix, so which one you build is not a detail.