Skip to content
All library documents

Using Vintage Losses for Credit VaR and Its Limits

Article Quant Q&A · Author: Victor123
Read the original

Summary

The discussion considers an empirical method for estimating a credit portfolio loss distribution from monthly account snapshots. The proposed procedure groups loans by origination vintage, measures defaults over a fixed horizon, applies a common recovery assumption, and aggregates losses across accounts. Repeating this across vintages creates observations that can be summarized as a loss distribution, from which a tail quantile could be used as a Credit VaR estimate.

The response notes that this is a quick empirical approximation, not a reliable standalone tail model. Overlapping monthly observations can make default rates autocorrelated, while a five-year history may contain too few extreme events to estimate tail quantiles well. Credit losses also cluster across economic cycles, contradicting an assumption of independent defaults. A Vasicek loss model is offered as an alternative: its default probability and correlation parameters can be calibrated to history to produce a smoother distribution and support sensitivity analysis, though data length and dependence issues remain.

Key ideas

  • Vintage-level default and recovery estimates can be aggregated into empirical portfolio loss observations.
  • A tail quantile of the resulting distribution can serve as a preliminary Credit VaR estimate.
  • Overlapping horizons create autocorrelation in vintage loss observations.
  • Short histories may miss severe losses and understate uncertainty in extreme quantiles.
  • Economic cycles cluster defaults, so independence assumptions may be unrealistic.
  • A calibrated Vasicek model can provide a smoother distribution and enable sensitivity analysis.

Tags

Fixed incomeRisk managementStatistics
Full text
# A quick and dirty loss distribution and Credit VaR


# A quick and dirty loss distribution and Credit VaR












I need to create a loss distribution for a credit portfolio as the first steps to estimate the portfolio Credit VaR.

I have historical monthly account snapshots (payment history) of all accounts going back 5 years. I need a simple intuitive approach.

I will assume that defaults are uncorrelated and also the recovery rate for all defaulted loans is same (x%)

I am thinking of below steps:

-Pick a vintage- Say Jan 2015.i.e the population to study is all loans opened in Jan 2015. -For each loan in above, see how many defaulted within 12 months of origination. Say p% -Assume a standard recovery rate (x%) for defaulted loans. -Loss of this particular vintage within 1 year = Sum(pxindividual loan balance at default)

Now I repeat the calculation selecting a different vintage..i.e loans originated in Feb 2015. I have 5 years(60 months), so I will have 60 different values of expected loss. Using this, if I draw a histogram , where the x axis represents different values of the EL and y axis the relative frequency of that particular loss, I have a probability density function. Can this be used as a loss distribution?

If yes, I can proceed to calculate the x value corresponding to the 5% significance level and that will be my VaR, correct?

## Answer by Magic is in the chain (score 1)

https://quant.stackexchange.com/a/45255

It would indeed provide a QD empirical loss density, though a few problems. The default rates would have autocorrelation due to the use of the monthly snapshots (overlapping outcomes). Last 60 months data probably won’t capture extreme loss events so extreme quantiles will be inaccurate. Another subtle point is around the economic cycle, as you would have noticed, good and bad years are clustered due to the very nature of economic cycles, which will bring in additional correlation.

You can also try Vasicek loss distribution. Calibration of whose parameters such as the correlation and PD can be done using historical data and you will then get a much smoother loss distribution. Again the length of data and the autocorrelation etc will have an impact but at least you can then do some additional sensitivity analysis.

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.