Risk Management for a Commercial Lending Portfolio Using Time Series Forecasting and Small Datasets

Abstract

Most risk managers use the Expected Credit Loss model to create allowances and provisions by calculating expected credit loss for the next 12 months. The expected loss is an estimate of the next 12 months actual loss percentage. This approach, by far, is considered to be a standard in the banking domain since it uses estimates from BASEL compliant Probability of Default, Loss Given Default and Exposure at default models. But the limitation of the approach is it's over conservative nature, and mangers end up reserving much more than is optimally required. This paper argues that time series forecasting techniques help in estimating next 12 month's actual loss percentage more accurately than the expected loss approach and hence is more appropriate technique of loss forecasting. The paper addresses the challenges of developing time series models with small datasets. Aggregating quarterly reported transaction data at the portfolio level, the data points get largely condensed and it is difficult to get more than 25 quarterly reported data points. To solve this problem associated with the small number of observations; this paper suggests a simple technique to simulate monthly data points using random numbers, under specific assumptions. ARIMA models are developed on the simulated data to estimate the next 12 months loss. The time series models show a lower prediction error as compared to the Basel compliant Expected Loss approach.