University of Twente Student Theses


An improved LGD model for the hire purchase and financial lease portfolio of Volkswagen Bank GmbH Branch NL

Heeres, P.E. (2009) An improved LGD model for the hire purchase and financial lease portfolio of Volkswagen Bank GmbH Branch NL.

[img] PDF
Abstract:Problem background For its financial lease and hire purchase portfolio Volkswagen Bank aspires to use the Internal Ratings Based approach for the determination of the credit risk capital. Under this approach the bank needs to find its own estimate for loss given default using a LGD model. The risk management department of Volkswagen Bank had identified some problems in the current LGD model. Therefore a student Industrial Engineering and Management from the University of Twente was given the task to back test the current model and to find improvements to increase the predictive power of the LGD model and to make it more intuitively explainable within the applicable regulatory constraints. Recommendations and motivation It is recommended to implement the following improved LGD model. Like the current model this model consists of two sub models. This is based on the idea that a contract can go in default and then cure to become a normal working contract again without a loss being incurred. The probability of this happening is modeled by the cure model. The loss given non-cure model predicts the loss in case a contract does not cure. The model prediction for LGD can be found by multiplying the outcomes of both sub models. Cure model <<Figure confidential>> Figure 1 - Improved cure model The probability of cure model uses contract vintage to distinguish between different groups of contracts. The newest contracts cure most frequently as was shown by univariate analysis and confirmed by the debt control department. Contracts that are past their original end date cannot cure by definition and this is captured too in the improved model. For hire purchase contracts client age is important because this variable is used to model the mortality risk. In case of death of a client the contract cannot cure so as client age increases, the probability of cure decreases. For financial lease contract the whether the car is new or used is the variable that distinguishes best between groups of contracts in terms of their probability of cure. Loss given non-cure model <<Figure confidential>> Figure 2 - Improved loss given non-cure model The loss given non-cure model adds the hire purchase contracts of self employed clients to the group of financial lease contracts, because both client groups have many similarities in terms of income source and as the client age is lower for self employed clients mortality risk plays only a minor role. A distinction is made between commercial vehicles and regular cars as in case of a commercial vehicle the contract is treated more conservatively at acceptation. Loan to value is an Loss Given Default= (1 -Probability cure)xLoss Given Non-cure | An improved LGD model for Volkswagen Bank GmbH Branch NL 3 important variable to predict loss which is shown by univariate analysis and literature, therefore an equation with loan to value is used. In the group of non-self employed clients mortality risk plays an important role which is modeled using the variable client age. For contracts that have a start date before July 1st 2008 a large loss is encountered in case of death of the client, because in that case **.*** euro of the exposure does not need to be repaid. This can be modeled with an equation that uses client age. Starting July 1st 2008 the full exposure of the contract needs to be repaid even in case of death of the client. It is expected that loss can be predicted using loan to value in a similar way as for the financial lease contracts and hire purchase contracts for self employed clients. Consequences The performance of the recommended new LGD model has been tested and it shows improvements in calibration and precision. At the level of the sub models the performance is also better in comparison to the current LGD model. There is a small difference in performance of the model on a dataset on which it has been developed and data for which that was not the case, but still the model is expected to work well on new data. However for individual contracts the difference between predicted LGD and the realized LGD can be large. The new LGD model is easier to explain intuitively because the hard to explain cut-offs between categories of contracts in the current model have been removed. On average the LGD predicted by the new model will be lower compared to the current model. This is because the dataset contained an error which was resolved before the development of the new LGD model.
Item Type:Essay (Master)
Volkswagen PON Financial Services
Faculty:BMS: Behavioural, Management and Social Sciences
Subject:85 business administration, organizational science
Programme:Industrial Engineering and Management MSc (60029)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page