For backtesting/following a PD model there are quite some know, relevant statistical tests (Gini, KS,…).But how about LGD models: of course you compare the predicted with the observed LGD’s. But which statistical indicators can you use to follow an LGD model? When do you conclude that a model no longer performs well? Are there confidence interval for the predictions, stability indices, …? Thanks

Rating Agencies first model a measure of LGD by regressing it on relevant factors such as Collateral, Firm level, Industry, Macroeconomic, Geographic info. and then perform inverse beta transformation to normalize it. Further Prediction Intervals, Prediction Error rates (MSE), Correlation with Actual Recovery Rates, and Prediction of larger than Expected Loss and Power Curves are used to validate the model.

QuoteOriginally posted by: vikashpungliaRating Agencies first model a measure of LGD by regressing it on relevant factors such as Collateral, Firm level, Industry, Macroeconomic, Geographic info. and then perform inverse beta transformation to normalize it. Further Prediction Intervals, Prediction Error rates (MSE), Correlation with Actual Recovery Rates, and Prediction of larger than Expected Loss and Power Curves are used to validate the model.vikashpunglia: thanks for your quick reply! I have a few questions concerning your reply:1) how do you calculate the prediction interval? Which distribution to use?2) when will you accept the model and when will you conclude that an LGD model no longer performs well: are there any statistical tests or do they use 'rules of thumb' (eg mean observed LGD 30% higher than predicted implies model change)?Thanks a lot!

1) How do you calculate the prediction interval? Which distribution to use?A disclaimer first. I am no expert in this field. Pls take my reply with a pinch of salt. Can't one calculate the prediction for Beta distribution? I have no inkling. If one can then the percentage of sample within 5%-95% bounds can be used.2) When will you accept the model and when will you conclude that an LGD model no longer performs well: are there any statistical tests or do they use 'rules of thumb' (eg mean observed LGD 30% higher than predicted implies model change)?What according to you are the utility of Prediction Intervals, Prediction Error rates (MSE), Correlation with Actual Recovery Rates, Prediction of larger than Expected Loss and Power Curves. One can keep on calculating these metrics to ensure that the model is working fine.

QuoteOriginally posted by: vikashpunglia1) How do you calculate the prediction interval? Which distribution to use?A disclaimer first. I am no expert in this field. Pls take my reply with a pinch of salt. Can't one calculate the prediction for Beta distribution? I have no inkling. If one can then the percentage of sample within 5%-95% bounds can be used.2) When will you accept the model and when will you conclude that an LGD model no longer performs well: are there any statistical tests or do they use 'rules of thumb' (eg mean observed LGD 30% higher than predicted implies model change)?What according to you are the utility of Prediction Intervals, Prediction Error rates (MSE), Correlation with Actual Recovery Rates, Prediction of larger than Expected Loss and Power Curves. One can keep on calculating these metrics to ensure that the model is working fine.Problem is that for retail credits the recoveries are absolutely not beta distributed. There are usually two peaks, near zero and near one. Very hard to use a regression technique for such a distribution! So difficult to calculate confidence intervals, ...

Problem is that for retail credits the recoveries are absolutely not beta distributed. There are usually two peaks, near zero and near one. Very hard to use a regression technique for such a distribution! So difficult to calculate confidence intervals, ...I would not call myself an expert in the LGD field, but I have done some research recently. The distribution you described can in my opinion be modelled via beta distribution (I am talking about the distribution for recovery rate for a single credit exposure). For example beta distribution with:a = 0,575510204b = 0,383673469andEX = 0,6std_dev(x) = 0,35gives the distribution, with 2 peaks (near zero and one). I am not talking about the case when regression technique is used for the LGD forecasting. I totally agree with you that there seems to be no established techniques for the validation of LGD models comparing to the methods used for PDs.