Problem
A bank's PD model estimates default probability over one year. IFRS reserves need the probability of default over the whole life of a loan, and most small business contracts run longer than a year. VTB's small business risk team had monthly rated portfolio slices for 2015-2021 and a default database. Raw cohort curves by rating are noisy: for rare ratings the cohorts shrink fast, and the curves of a good and a bad rating cross.
Example from the article: for individual entrepreneurs, rating MC2 gets a cumulative PD of 0.48% in year 1 and 12.81% in year 10, while rating MG3 goes from 55.23% to 81.04%.
Solution in brief
Build the cumulative default rate (cDR) per rating from cohorts. Merge sparse neighbouring ratings into groups until the curves stop crossing. Fit each curve with a Weibull-type function by least squares on a linearized form. Take year-1 PD from the bank's master scale and derive years 2-10 from the fitted curve.
Questions this case answers
- How to extend a one-year default probability into a multi-year default curve for each rating when cohorts shrink and curves cross
- How to fit cumulative default curves by group from sparse cohorts with a parametric function, using a known one-year level as the anchor
- How to forecast lifetime default probability curves for a loan portfolio beyond one year
- How to build cumulative default rate curves by rating from cohorts with few observations
Tools and methods
- Cohort default rate and cumulative DR — by rating, per month since the slice
- Rating grouping — merges sparse ratings so curves do not cross
- Weibull approximation (classical and adjusted) — least squares on a linearized form, chosen by
- Master scale — gives the year-1 PD for each rating
Google or a one-time email code. No password.