Title: Opening the black box: Deep quantile neural networks for loss given default prediction
Authors: Ralf Kellner - Saarland University (Germany)
Maximilian Nagl - University of Regensburg (Germany)
Daniel Roesch - University of Regensburg (Germany)
Maximilian Nagl - University of Regensburg (Germany) [presenting]
Abstract: A flexible combination of quantile regression and neural networks for Loss Given Default prediction is proposed. The results show a superior performance compared to linear quantile regression. This may be caused by non-linear behaviour in higher quantiles, especially in Europe. By using a novel feature importance measure, we quantify the importance and direction of every input variable. This makes neural networks as interpretable as linear models. Moreover, we show that the macroeconomy is up to two times more important in USA than Europe and increasing in quantiles. The macroeconomy is most important in the US, whereas in Europe collateralization is essential.