The generation of synthetic data for risk modelling

Abstract

Both recent advances in technology and changes in regulatory requirements have led to the increased popularity of advanced analytical techniques in risk management. These techniques are data intensive and therefore require a minimum amount of historical data. Financial institutions often have limited historical data available, due to the cost or to incomplete data collection in the past. This paper proposes an approach for generating synthetic data that allows risk parameters, such as probabilities of default (PDs), to be quantified when data are limited. The approach consists of imputing synthetic model drivers within an existing data framework by leveraging partial historic data and/or information derived from expert opinions or external sources. Synthetic proxies are produced for drivers with no data, limited data or data of poor quality. The synthetic drivers are generated consistently to adhere to existing or expert-driven correlations among all variables. In a logistic regression setup, the authors illustrate the approach using stylised data from real estate transactions and show how model performance metrics of PD estimates can be improved. The authors conclude that a financial institution with limited or no historic data on important model drivers can use expert views or publicly available data to improve estimates of risk parameters.