Handling of multicollinearity problem in modelling non-performing loans in Africa's portfolio data

Abstract

Non-performing loans (NPLs) are detrimental to profits in the banking sector. Predicting the level of NPLs using macroeconomic variables is vital in order to build mitigating actions for such scenarios to safeguard the profitability of the institution. Macroeconomic variables are susceptible to high correlations amongst each other, bringing about the problem of multicollinearity. Predicting in the presence of multicollinearity brings about unreliable and inefficient results. This study aims to find an optimal and efficient way of forecasting NPLs using Ordinary Least Squares (OLS), Ridge Regression (RR) and Principal Component Analysis (PCA) while correcting for multicollinearity. To do this, NPL data from bank X was attained, along with multiple macroeconomic variables, specifically for Kenya and Nigeria. It is critical to assess the determinants of NPLs so that effective and efficient policies can be deployed to prevent the rising trajectory of NPLs. To minimize the risks of using expert judgement, it is necessary to consider effective statistical methods for predicting NPLs. The benefits accrued from such methods include (1) minimum collection costs incurred when a loan defaults, such as less phone calls urging the customers to pay, less litigation costs when trying to recover the assets, less shortfalls incurred when disposing off the assets that have been repossessed and less auction sales if the assets have to be auctioned, to mention a few; (2) correct pricing for the risk; (3) be able to differentiate between high-risk and low-risk accounts based on the macroeconomic factors; and (4) be more prudent in granting credit to minimize losses and maximise profits. This study considers the OLS, RR and PCA in modeling the NPLs data from bank X. The results showed that multicollinearity exists for most variables. Some of the variables did not conform to the assumptions of the OLS. The models for OLS for both countries were significant, while some of the variables displayed illogical outcomes, possibly due to multicollinearity among the predictor variables. RR method solved for multicollinearity and had a relatively predictive power for Nigeria data and not Kenya, whereas PCA solved for multicollinearity and introduced a positive factor in data reduction and had a relatively better predictive power. The mean square errors (MSEs) for PCA and RR were lower than that of OLS. A key limitation was inadequate data from the banking sector due to sensitivity issue. We conclude that the data can be expanded, and the number of variables reduced so that prediction can be more precise. Further work using other methods such as GARCH can be explored to improve the prediction of the NPLs in the midst of multicollinearity.