Better Efficiency on Non-performing Loans Debt Recovery and Portfolio Valuation Using Machine Learning Techniques

The following research is based on a portfolio of non-performing loans (NPLs), which was previously acquired and managed by a collection agency, the company under study is one of the owners of the portfolio. The study compares the efficiency and performance of several machine learning algorithms to develop and implement a forecasting tool to estimate the recovery rate of NPL portfolios. These models help to enhance and support the debt collection operation, allowing to forecast the number of debtors that will be recovered in the lifetime of the portfolio, as well as to efficiently manage resources (recovery task force) by reducing costs and expenses. The application aims to support the valuation process at the time of portfolio purchase. The study shows that the application using a binary ranking approach based on the XGBoost model outperforms other techniques, offering good results. It is also evident that product type was one of the most influential variables among the different models. The model using this algorithm could serve as a decision support tool, precisely in the operation of purchasing a portfolio of unprofitable debts, as it allows the quantification of the client’s debt to be recovered by identifying the group of potential debtors with the highest probability of compliance, which would result in a faster and more efficient debt collection process.