Skip to content
All library documents

Validating Performing-Loan LGD Estimates Against Realized Losses

Article Quant Q&A · Author: I. STACK OVERFLOW
Read the original

Summary

The document asks how to validate a performing-loan Loss Given Default model by comparing its estimates with realized workout LGD. It distinguishes a model developed using a through-the-cycle sample of defaulted loans from estimates produced for a point-in-time performing sample. The author questions whether comparing those outputs directly is meaningful because the populations and time perspectives differ.

One proposed approach is to estimate LGD for performing loans and then compare those estimates with realized LGD for loans or clients that enter default during a subsequent one-year period. The document does not establish whether this proposal is methodologically sound, give a validation procedure, or provide empirical results. It is best read as a statement of a sample-alignment problem and a request for guidance. Any comparison would need to account for the fact that realized LGD is observable only for defaults, while estimates may cover a broader performing population; the text leaves that issue unresolved.

Key ideas

  • The model is described as developed on a through-the-cycle sample of defaulted loans.
  • The proposed comparison is between point-in-time estimates for performing loans and realized workout LGD.
  • The document questions whether differing samples and time perspectives make a direct comparison appropriate.
  • A proposed alternative is to compare estimates with realized outcomes for loans that default in a later period.
  • No validation method or evidence is supplied, so the proposal remains unresolved.

Tags

Fixed incomeStatisticsRisk managementBacktesting
Full text
# LGD performing model - LGD estimate vs LGD observed


# LGD performing model - LGD estimate vs LGD observed












LGD (Loss Given Default) performing model is developed on through the cycle sample which consists of loans in default.

What I want is to compare LGD estimate and LGD observed (realized). LGD observed is workout LGD calculated on all defaulters from last (for example) three years. What is LGD estimate?

I think that it is not good to calculate LGD estimate on 'point in time' performing sample and to compare it with LGD observed since LGD estimate is calculated on point in time performing sample and LGD observed (realized) is calculated on through the cycle sample which consists of only defaulted loans.

One of my ideas which is more complicated than those matched above (but I do not know if it is OK) is to use performing sample (and calculated LGD estimate on it) and to calculated LGD observed on all loans/clients who enters into default in period of 1 year.

How do you do it? Or what do you think about comparison of LGD estimate and LGD observed in validation of LGD performing model.

I will be very thankful for any suggestions. Also, if you have some articles or books which can resolve my problem, please send to me.

Also, I find a lot of articles where it is necessary to compare LGD estimate and LGD observed pa I can not find on which sample LGD estimate is calculated.

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.