ARCHIVES

The work lays emphasis on the management of risk of the bank loan portfolio through the analysis of the operations of the Portfolio Management team. These operations involve the formation of Collateralized Loan Obligations (CLOs), hedging through stock-specific and index Credit Default Swaps (CDS), strategic selling of loans, non-payment insurance and risk participation. Since all loan documents were mainly in PDF, the project utilized the Natural Language Processing (NLP) methods to process unstructured data and detect the missing document or loan errors / anomalies in an efficient manner. A solid data pipeline was established to fetch data on loans from the Oracle databases and then transformed via ETL. A lot of data cleaning, feature scaling, and feature engineering were done in Python(2.x/3.x) and pandas and NumPY which guaranteed high quality of data. To learn intrinsic and combined effects, Exploratory Data Analysis (EDA), univariate and bivariate analysis were performed. Principal Component Analysis (PCA) and Factor Analysis were the dimensionality reduction methods utilized to make a model more efficient. Statistical significance, cross-validation, and ROC plots were used to validate and predict based on predictive analytics and machine learning algorithms to predict important portfolio metrics. Dynamic programming methods of reinforcement learning have also been used in order to optimize decision-making strategies. The quantitative loan data indicated a 28 percent increase in the success of missing document detection and anomaly detection, a 15 percent gain in predictive value of the metrics of risk in a portfolio, and a 22 percent decrease in the processing time of the loan data extraction and validation. The conceptual dashboards allowed business stakeholders to have actionable insights, in order to make data-driven decisions and more effective risk reduction methods.