Multi-view locally weighted regression for loss given default forecasting

Abstract

Accurately forecasting loss given default (LGD) poses challenges, due to its highly skewed distributions and complex nonlinear dependencies with predictors. To this end, we propose a multi-view locally weighted regression (MVLWR) method for LGD forecasting. To address the complexity of LGD distributions, we build a specific ensemble LGD forecasting model tailored for each new sample, providing flexibility and relaxing reliance on distribution assumptions. To address complex relationships, we combine multi-view learning and ensemble learning for LGD modeling. Specifically, we divide original features into multiple complementary groups, build a view-specific locally weighted model for each group, and aggregate the outputs from all view-specific models. An empirical evaluation using a real-world dataset shows that the proposed method outperforms all the benchmarked methods in terms of both out-of-sample and out-of-time performance in LGD forecasting. We also provide valuable insights and practical implications for stakeholders, particularly financial institutions, to enhance their LGD forecasting capabilities.