Improving the predictive performance of longitudinal risk models for UK SMEs

Abstract

In 2008, the whole world was a picture of economic depression. During the credit crisis, the viability of Small and Medium-sized Enterprises (SMEs) has been profoundly jeopardised because of their vulnerability. After the credit crisis, raising the awareness of risk management in the banking sector has been needed. The launch of the Basel III regulations proposes a more stringent requirement on capital and liquidity to promote stability in the financial system. Regarding credit risk, the probability of defaults (PDs) models essentially remains unchanged. On the other hand, the expected credit loss (ECL) model under International Financial Reporting Standard (IFRS) 9 proposed by International Accounting Standard Board (IASB) is expected to bring significant influence on SMEs and Banking sectors since there is an increase of loan loss provision undoubtedly. This thesis aims to explore the performance of SMEs due to the fundamental role played in a country’s economic development. A large dataset used in this thesis includes 79 characteristics of UK SMEs from 2007 to 2010. The SMEs were pigeonholed into start-ups (growing businesses) and non-start-ups (developed businesses) considering their different behaviour during the credit crisis. However, the dataset contains a substantial number of incomplete observations, and the analysis of such dataset is a handicap. In light of this, Multiple Imputation by Chain Equations (MICE), a state-of-the-art and flexible technique, has been employed to deal with missing data. Although this technique is widely used in the medical field but not in credit risk modelling, it takes into account the uncertainty within the process of combing multiple imputed dataset to produce estimated coefficient, and each type of variables has its specific model for imputation. Once getting over the missing data problem, cross-section analysis is followed to build up the credit risk model. Logistic regression and shrinkage regression are used to analyse the relationship among the selected variables and Generalised Additive Models (GAM) is performed to capture their non-linear relationship, derive a direct marginal trend and plot how explanatory variables influence SMEs performance. Subsequently, time effects are accounted for by employing a panel model controlling the time effect using year dummy variables or macroeconomic variables. It can be found that the panel data models with firm-specific and macroeconomic variables are preferred as the AUROC is at least as other models, especially during the credit crisis. Again, the ex-post regulations, the 12-month ECL may not capture a significant increase in credit risk if the economic downturn is expected to occur at a later stage. The lifetime ECL captures this downturn and will, therefore, identify a significant increase in credit risk sooner. The panel data models are believed to capture the change in the macroeconomics during the credit risk and are appropriate to apply to meet Basel III and IFRS 9 requirements as these regulations require to consider the economic cycle.