Social network analysis for Credit Risk Modeling

Abstract

The modern society quest for credit has led to an increase in consumer credit uptake with proportionate increase in default rates. Some of the current credit risk models deteriorate over time and need frequent validation. This is due to use of historical data that is time dependent. The models lack the flexibility to take into account changes in economic and other extreme events. This research considers the use of social network data to offer an alternative approach in consumer credit scoring. Social data is widely available with vast amount of information due to increase in social network sites and technologies. This offers time dependent data that can be harnessed to develop time dependent models that do not need validation and will not deteriorate with time. Further, the poor and young consumers lack historical data and thus social data will cover that gap. The set of agents who are part of a network are also obligors in a loan portfolio with a financial institution. The key aim is to estimate credit risk in a loan portfolio based on the agents’ behavior at the network level. The agents interactions and cyclical inter dependencies in the social economic network are estimated to derive the social and economic factors. These factors are derived from the network matrices using singular value decomposition technique and scaled into (0; 1]. The scaled data forms the credit risk analysis factors that are used to learn and train the hidden Markov model. The model emits the credit quality levels, the dynamic threshold and the credit quality scores. These outputs are in turn used to estimate the model false rates and the obligors’ delinquent cases, default rate, stopping time and survival rates. The dynamic threshold is estimated at each time period to capture the dynamics of the credit quality of the obligors in the loan portfolio and emit the default and non-default rates. Obligors are classified into four credit quality levels; poor, average, good and excellent (PAGE). Obligors with average and good credit quality levels ranges between 61:5% and 89:3%

while the excellent credit quality level was between 8% and 20:4%. The obligors classified in the false rate category ranges between 25% and 50:1%. The model performance is between 53% and 73% which is an accuracy rating of between medium and good accuracy. Sensitivity analysis and false rates in the model have a coefficient of determination of between 0:647 and 0:983. The social network model offers an alternative approach to consumer credit scoring with time dependent data. Agents’ interactions and cyclical interdependencies is an ideal approach to incorporate in consumer underwriting and capture the poor and the unbanked. The model has opened new frontiers in consumer credit scoring. Thus, the study contributes in opening up new frontiers and innovations in consumer credit scoring with social and economic data.