Using an Asymmetric Loss Function to Alleviate the Risk of Loan Collateral Overvaluation

Financial institutions are increasingly adopting machine learning-based valuation models to evaluate loan collaterals. However, most machine learning algorithms do not differentiate between the risks associated with the undervaluation and overvaluation of such assets. From the perspective of a lender, the risks of overvaluing loan collateral are more critical than those that arise from undervaluing them. In this study, we alleviate this risk of overvaluation by explicitly considering an asymmetric loss function. We customize a gradient boosting machine (GBM) by specifying an asymmetric loss function, and assigning a higher penalty for overvaluation. This customized GBM is then used to predict house prices in Gimhae, South Korea. The results show that the GBM effectively reduces overvaluation while maintaining prediction accuracy. Researchers and practitioners need to intentionally bias their machine learning algorithms to incorporate the asymmetric risks associated with their businesses. The approach proposed in this study can help stakeholders make informed decisions in the lending process, thereby mitigating the risk of default by borrowers, and ultimately promoting sustainable lending practices.