Core Path · L6.3.05

Calibrating PDs

14 minCalibrationBinomial testJeffreys testLong-run averagePIT and TTCLow-default portfolios

By the end you will be able to

  1. 1Test a model's calibration in the large and across groups, and read a Hosmer-Lemeshow statistic.
  2. 2Test each grade's PD against its defaults with the binomial, Jeffreys and correlation-adjusted tests, and explain why they disagree in a bad year.
  3. 3Set a grade's PD from a long-run average, and place a PD between point in time and through the cycle.
  4. 4Bound the PD of a low-default grade and build a margin of conservatism.

Why it matters

In 2021 RiskAtoms Bank plc's grade 8 firms, rated at a PD of 4.53%, defaulted at 15.2%: 38 defaults where 11 were expected. Under independence the chance of that is 8.1 × 10⁻¹¹. Was the PD wrong? L3.3.06: Probability of default booked the history; this lesson tests it, and finds the answer depends on what the test assumes about recessions.