Improvements in Loss Given Default Forecasts for Bank Loans

Abstract

Standard-Nutzungsbedingungen:Die Dokumente auf EconStor dürfen zu eigenen wissenschaftlichen Zwecken und zum Privatgebrauch gespeichert und kopiert werden.Sie dürfen die Dokumente nicht für öffentliche oder kommerzielle Zwecke vervielfältigen, öffentlich ausstellen, öffentlich zugänglich machen, vertreiben oder anderweitig nutzen.Sofern die Verfasser die Dokumente unter Open-Content-Lizenzen (insbesondere CC-Lizenzen) zur Verfügung gestellt haben sollten, gelten abweichend von diesen Nutzungsbedingungen die in der dort genannten Lizenz gewährten Nutzungsrechte. Terms of use: Documents in Pitfalls in Modeling Loss Given Default of Bank Loans AbstractThe parameter loss given default (LGD) of loans plays a crucial role for risk-based decision making of banks including risk-adjusted pricing. Depending on the quality of the estimation of LGDs, banks can gain significant competitive advantage. For bank loans, the estimation is usually based on discounted recovery cash flows, leading to workout LGDs. In this paper, we reveal several problems that may occur when modeling workout LGDs, leading to LGD estimates which are biased or have low explanatory power. Based on a data set of 71,463 defaulted bank loans, we analyze these issues and derive recommendations for action in order to avoid these problems. Due to the restricted observation period of recovery cash flows the problem of length-biased sampling occurs, where long workout processes are underrepresented in the sample, leading to an underestimation of LGDs. Write-offs and recoveries are often driven by different influencing factors, which is ignored by the empirical literature on LGD modeling. We propose a two-step approach for modeling LGDs of nondefaulted loans which accounts for these differences leading to an improved explanatory power. For LGDs of defaulted loans, the type of default and the length of the default period have high explanatory power, but estimates relying on these variables can lead to a significant underestimation of LGDs. We propose a model for defaulted loans which makes use of these influence factors and leads to consistent LGD estimates.

Forecasting recoveries in debt collection: Debt collectors and information production

“…In defining what relevant changes are in this setting, we refer to results from the recovery rate literature that predicts collection rates in a methodologically similar manner. Bellotti & Crook () report adjusted values between 10.5% and 11.1% in a linear regression on a comprehensive set of debtor, contract, spatial, and macroeconomic characteristics; Gürtler & Hibbeln () report values between 4.4% and 18.9%; Loterman, Brown, Martens, Mues, and Baesens () examine six different loan data sets and find values between 1.2% and 44.12%. The value is considerably higher in studies that include information on collateral—up to 76.9% for Ingermann et al () or up to 61% for Qi & Yang ().…”

“…The data contain the monthly collection payments for each individual account, starting from the date of transfer to the collection agency. Following Dermine & de Carvalho () and Gürtler & Hibbeln (), we analyze payments over a standardized payment period. Accordingly, only accounts that have been in debt collection for this minimum number of months are considered.…”

Time Matters: How Default Resolution Times Impact Final Loss Rates

“…The latter additionally analyse the improvement in estimation when segmenting LGDs with decision trees before estimation, hereby, constructing a mixture of the LGD distribution. Gürtler and Hibbeln (2013) focus on the effects of censoring on LGD observations and the negative consequences when censoring is ignored (see Section 2 for further information on effects of censoring). They do not provide a methodical solution, but suggest to restrict the data set to avoid biased estimates.…”

From Power Curves to Discriminative Power: Measuring Model Performance of LGD Models

“…Besides OLS and logit regression new models have been developed recently: inflated beta regression ( [6]), generalized beta regression ( [7]), censored gamma regression ( [8]), zeroadjusted gamma regression ( [9]), and mixture-models ( [10] and [11]). In [12] the authors point out some problems that arise in LGD estimation and show how they may be solved. All these models have been developed to accurately take into account the special shape of the empirical LGD distribution.…”