Overview of Loss Given Default Models

Loss given default (LGD) is the proportion of a credit that is lost in the event of default. LGD is one of the main parameters for credit risk analysis. Although there are different approaches to estimate credit loss reserves and credit capital, common methodologies require the estimation of probabilities of default (PD), loss given default (LGD), and exposure at default (EAD). The reserves and capital requirements are computed using formulas or simulations that use these parameters. For example, the loss reserves are usually estimated as the expected loss (EL), given by the following formula:

EL = PD * LGD * EAD

With increased availability of data, there are several different types of LGD models. Risk Management Toolbox™ supports:

• Regression models — These are linear regression models where the response is a transformation of the LGD data. For more information on the supported transformations, see Regression.

• Tobit models — These are censored regression models with explicit limits on the response values to capture the fact that LGD can take values only between 0 and 1. Censoring on the left, right or both sides are supported. For more information, see Tobit.

The Model Loss Given Default example shows these two types of models, as well as other models, are fitted using Statistics and Machine Learning Toolbox™. Specifically, besides the regression and Tobit models, this example also includes a non-parametric, look-up table type of model; a Beta regression model; and a “two-stage” model where a classification model (cure-no cure) and a regression model (predicted LGD conditional on no cure) work together to make LGD predictions.

In addition, you can use the Regression and Tobit models to develop LGD models that include macroeconomic predictors for stress testing or to support regulatory requirements such as IFRS 9 and CECL. For more information, see Overview of Lifetime Probability of Default Models.

Model Development and Validation

Risk Management Toolbox supports the modeling and validation of LGD models through a family of classes supporting:

• Model fitting with the fitLGDModel

• Prediction of LGD with the predict function

• Model discrimination metrics with the modelDiscrimination function and visualization with the modelDiscriminationPlot function

• Model accuracy metrics with the modelAccuracy function and visualization with the modelAccuracyPlot function

The supported model types are Regression and Tobit models.

A typical modeling workflow for LGD analysis includes:

1. Data preparation

   Data preparation for LGD modeling requires a significant amount of work in practice. Data preparation requires consolidation of account information, pulling data from multiple data sources, accounting for recoveries, direct and indirect costs, determination of discount rates to determine the observed LGD values. There is also work regarding predictor transformations and screening. There is a wide range of tools available to treat missing data (using fillmissing), handle outliers (using filloutliers), and perform other data preparation tasks. The output of the data preparation is a training dataset with predictor columns and a response column containing the LGD values.

2. Model fitting

   Use the fitLGDModel function to fit an LGD model. You must use the previously prepared data and select a model type. Optional inputs allow you to indicate which variables correspond to predictor variables, or which transformation to use for a regression model, or the censoring side for a Tobit model. You can specify a model description and also specify a model ID or tag for reporting purposes during model validation.

3. Model validation

   There are multiple tasks involved in model validation, including

   • Inspect the underlying statistical model, which is stored in the 'UnderlyingModel' property of the Regression or Tobit object. For more information, see Basic Loss Given Default Model Validation.

   • Measure the model discrimination on either training or test data with the modelDiscrimination function. Visualizations are generated using the modelDiscriminationPlot function. Data can be segmented to measure discrimination over different segments.

   • Measure the model accuracy on either training or test data with the modelAccuracy function. Visualizations are generated using the modelAccuracyPlot function. Also, you can visualize the residuals.

   • Validate the model against a benchmark (for example, a champion model). For more information, see Compare Tobit LGD Model to Benchmark Model.

   • Perform a cross-validation analysis to compare alternative models. For more information, see Compare Loss Given Default Models Using Cross-Validation.

   • Perform a qualitative assessment of conditional PD predictions by using the predict function directly with edge cases. Visualize residuals using the modelAccuracyPlot function. There are examples of additional visualizations using histograms and box plots in the Model Loss Given Default example.

References

[1] Baesens, Bart, Daniel Roesch, and Harald Scheule. Credit Risk Analytics: Measurement Techniques, Applications, and Examples in SAS. Wiley, 2016.

[2] Bellini, Tiziano. IFRS 9 and CECL Credit Risk Modelling and Validation: A Practical Guide with Examples Worked in R and SAS. San Diego, CA: Elsevier, 2019.

[3] Gupton, G., and R Stein. "Losscalc v2: Dynamic Prediction of LGD Modeling Methodology". Moody’s KMV Investor Services, 2005.