Applications of Early Warning Systems for Customer Segmentation of Wholesale Banking Clients
Details
- Authors: Alessandra Amato
- Title: Applications of Early Warning Systems for Customer Segmentation of Wholesale Banking Clients
- Supervisior: Prof. Dr. Marcos Machado, Prof. Dr. Jörg Osterrieder, Prof. Dr. João Luiz Rebelo Moreira
- Degree: Master of Science
- University: University of Twente
- Year: 2023
- Status: Working Paper
Summary
Development of a Customer Segmentation model integrating Early Warning Systems for Wholesale Banking clients
Abstract
In the rapidly evolving landscape of Wholesale Banking (WB), Early Warning Systems (EWS) has increasingly become a vital resource for financial institutions aiming at monitoring their credit portfolio and preemptively detecting financial distress scenarios. For instance, ING Bank has tried to leverage the overwhelming wave of data imposed by the phenomenon of Big Data by implementing the Advanced Risk Integrated Application (ARIA), the company’s EWS tool devel- oped to surveil their commercial clients and to generate a number of warning in the presence of a potential risk incurring. However, since the current active triggers are only capable of de- tecting ongoing negative changes, ING has tried to explore innovative ways to expand the value delivered by the tool and introduce new solutions for the identification of potential up-selling opportunities. Among all the possible data-driven techniques that nowadays companies have started to rely on in order to maximise revenues and enhance their profitability, automated Customer Segmentation (CS) represents one the most successful and effective techniques de- veloped. Therefore, the goal of this study focused on the investigation and implementation of a novel CS model, integrating early warning triggers, by answering the following main research question:
"How to design and integrate early warning signals into a new CS model in order to identify potential business opportunities within banks’ WB credit portfolio entities?"
In order to align the outcomes of the model developed to the initial business objectives, the research defined a number of requirements that the artifact should have presented related to its segments’ orientation, identifiability and actionability. On the basis of the aforementioned characteristics, the research designed and introduced several different variables that aimed at providing a comprehensive and complete overview of the risk scenario associated with each client. The attributes in question, which can be obtained from the preprocessing and feature engineering of historical records of clients’ internal data, internal triggers and external triggers, defined the client’s risk profile from several perspectives: the progress and growth the entities have faced through the months in terms of EAD, RWA, allocated limit, outstanding amount and expected loss, the evolution of the client’s credit quality rating, the average number of monthly early warning raised by each borrower and, finally, the client’s current activity status, credit limit and outstanding balance recorded in the last month of the study.
On the basis of the insights derived from a systematic literature review on the application of EWS and CS in the field of finance, two popular clustering algorithms have been deployed, namely K-Means and DBSCAN, along with dimensionality reduction techniques such as cor- relation analysis and Principal Component Analysis (PCA). Moreover, the Elbow Method and Silhouette Score were also used to validate the models deployed.
The assessment and interpretation of the clusters generated was performed through the imple- mentation of a number of analyses that explored the different segments from multiple aspects, such as the tightness and separation of the subgroups formed or the density and descriptive statistics of the customers’ distribution. From these studies it was discovered that the use of PCA slightly improved the compactness and distinction among the clusters, compared to the dataset derived from the correlation analysis. In addition, it was also observed that DBSCAN clustering algorithm proved to be unsuitable and inefficient for the type of data under exam- ination, as no real conclusion and meaningful insight could be derived from the exploration of its clusters. Finally, a risk-reward analysis and risk exposure analysis related to, respectively, the comparison between the average number of negative and positive monthly triggers and the juxtaposition of the growths detected for the outstanding amount and the respective EAD value, were included as well.
In conclusion, the research contributed to obtaining a deeper understanding of the financial health of ING’s WB clients, enabling decision-makers to re-adapt strategies and deliver more custom and targeted services based on each segment emerging needs. In addition, the study was able to bridge the gap between EWS and CS by introducing a novel perspective on strategic risk monitoring.