Anomaly detection using machine learning techniques: a comparative study on first payment default prediction in retail loans

Abstract

The banking sector is one of the sectors that adapt quickly and effectively to the modern methods provided by technological developments and is extremely flexible. As the digitization of banking products and applications becomes more widespread, data sources related to user behavior and products are also increasing. As a result, the datasets generated by these contemporary data sources have high potential as they allow banks to make their business more efficient by implementing better customer segmentation and targeting, and by improving risk management-related practices such as fraud detection in various ways. Various machine learning algorithms are widely applied and used in many different areas of banking. Loan default prediction is a crucial problem for banks to acquire less risky customers and avoid customers who are likely to default and go bankrupt. Detection of first payment default is a matter of great importance in order to prevent possible fraudulent activities and attacks against banks and to prevent possible financial losses that they may cause. In this study a dataset with 75000 observations has been used to benchmark and find an optimal model for the first payment default prediction problem. 11 different algorithms from supervised, unsupervised, and reinforcement learning based approaches are applied to the same dataset and compared to each other using the metrics such as F1 and lift score. In the model building process, the best-performing models in each of the supervision types have been selected. The isolation forest algorithm has been selected as the best performing unsupervised model, LightGBM has been select as the best supervised model while the DQN with 0.1 λ value has been selected as the best reinforcement learning model after the performance comparison. As the final step, the ensembling by weighted average techniques has been applied to the best 3 models. However, it is seen that the LightGBM model performs better than the ensemble models in terms of both F1 and lift scores. As a result, the LightGBM model is selected as the proposed model of the study, and the model has been interpreted and discussed in terms of the business perspective of the first payment default problem.