Not logged in.

Contribution Details

Type Master's Thesis
Scope Discipline-based scholarship
Title Developing an Event Based Monitoring Model using information regarding account movements
Organization Unit
Authors
  • Shiyu Qiu
Supervisors
  • Erich Walter Farkas
  • Markus Heusler
Language
  • English
Institution University of Zurich
Faculty Faculty of Business, Economics and Informatics
Number of Pages 53
Date 2020
Abstract Text The internal rating models currently used by banks to assess the solvability of small and medium-sized enterprises (SME) heavily rely on their yearly financial statements. In order to detect counterparties with looming credit deterioration early, we follow a different approach, which is based on continuous monitoring of the client's transaction data. We do so by creating transactional features that summarize the operational condition, financial behavior, and liquidity of the counterparty across different time windows. By combing the transactional features with business sector information, we develop both a traditional logistic regression model as well as a machine learning XGBoost model. It turns out that the XGBoost model has a better performance on capturing relations between transaction data and yearly rating information, reflected by an area under the ROC curve (AUC) of 0.857. In addition, we build specific models for clients belonging to different industries, finding that the models for the financial sector and the real estate sector perform the weakest, being due to insufficient data quantity and weak features, respectively. Therefore, we finally build three XGBoost models; one common model for the raw materials production, industry, construction, commerce, service and restaurant sectors, and two separate models for the particular sectors mentioned above. The model for the six sectors exhibits a good and robust performance with an average AUC of 0.874, which illustrates the effectiveness of transaction data when it comes to early tell apart clients with good and bad ratings. We also discuss the choice of the threshold used to define “good” and “bad” rating classes and test the robustness of the models for further real application.
Export BibTeX