Essays on Bankruptcy Prediction and Systemic Risk in Financial Institutions

Abstract

Financial institutions play an important role in the economic environment through eco- nomic stability, economic growth, management of risk and general development of the economy. These activities of financial institutions often expose them to risks and sometimes lead to their collapse. Financial bankruptcy causes major disruptions which usually result in a rippling effect on the economy. Accurate bankruptcy prediction is important to financial institutions, regulators, investors and central banks as it guides investment decisions and help mitigate financial losses. Prediction of bankruptcy is significant for the whole economy to help prevent financial crisis. In an interconnected global financial system, the need to understand and manage systemic risk is crucial to prevent financial crisis and maintain the economic stability economies across the world. Following the recent financial crisis of 2007-2008, systemic risk has gained prominence. The study focuses on systemic risk within the South African financial system using econometric models to identify and measure the risks within the financial system. The research provides a framework for detecting early warning signals of systemic risk. The findings from this study offer insights for policymakers, financial institutions, in- vestors and other stakeholders to develop strategies which mitigate risks and enhance the resilience in the financial system. The first empirical study investigates the application of state of the art ML models to predict bankruptcy in the US financial system leveraging on their ability to handle large dataset, deal with non-linear relationships and have higher predictive accuracy. We investigate the effects of different feature selection and different parameter opti- mization techniques on the performance of the ML models. We also implemented the SMOTE technique to mitigate the adverse effect of class imbalance on the predictive ability of the models. The thesis employs multiple models including support vector machine (SVM), online sequential extreme learning machine (OSELM), extreme gra- dient boosting machine (XGBoost) and light gradient boosting machine (LightGBM) using the CAMELS features. We performed experiments using financial ratios of both failed and solvent US banks to demonstrate the effectiveness of our proposed mod- els for bankruptcy prediction. The experimental results indicate that LightGBM out- perform the rest of the state-of-the-art machine learning approaches for bankruptcy prediction although the other ML algorithms peformed well on both the training and testing data. The first empirical study investigates the application of state of the art ML models to predict bankruptcy in the US financial system leveraging on their ability to handle large dataset, deal with non-linear relationships and have higher predictive accuracy. We investigate the effects of different feature selection and different param- eter optimization techniques on the performance of the ML models. We also imple- mented the SMOTE technique to mitigate the adverse effect of class imbalance on the predictive ability of the models. The thesis employs multiple models including support vector machine (SVM), online sequential extreme learning machine (OSELM), extreme gradient boosting machine (XGBoost) and light gradient boosting machine (LightGBM) using the CAMELS features. We performed experiments using financial ratios of both failed and solvent US banks to demonstrate the effectiveness of our proposed mod- els for bankruptcy prediction. The experimental results indicate that LightGBM out- perform the rest of the state-of-the-art machine learning approaches for bankruptcy prediction although the other ML algorithms peformed well on both the training and testing data. The second empirical chapter examines the systemic risk contribution of financial insti- tutions in South Africa to the overall risk in the financial system. The chapter utilised the VaR and conditional value at risk method based on GARCH models. This study investigates the VaR and CoVaR using SGARCH, EGARCH and GJRGARCH and com- pares the forecasting results of each volatility model. The distributions used in the analysis are the Student-t, Skewed Student-t and generalised error distribution. We use the forecast volatilities corresponding to each model to predict the VaR and Co- VaR of each of the indices. We test the predictive performance of the estimated models using backtesting procedures. For each estimated model, we computed the VaR, Co- VaR and the ∆CoVaR. Using a sample of daily equity prices spanning the period from February 19 2002 to May 31 2023, our results reveal that in South Africa, Nedbank is the most systemically important financial institution followed by Standard bank while Absa contributes little to the systemic risk in the South African financial system. There is the need for financial institutions, regulators, central banks and other stakeholders to implement regulatory measures such as higher capital adequacy requirements for systemically important institutions. The findings show that Nedbank is the systemi- cally important financial institution. In addition, the CoVaR indicator provides a more comprehensive view of risk and provides valuable information for risk management and investment decisions. The third empirical chapter evaluates and compares the performance of univariate GAS framework model combined with heavy-tailed asymmetrical distribution in estimating iii time-varying joint VaR and ES to estimate one-day-ahead tail risk forecast of daily eq- uity returns of financial institutions. We used the MAE and RMSE to rank the predictive abilities of the models. Also, we employed backtesting procedures and loss functions to evaluate the predictive accuracy of the different models. The results indicate that although all the models were adequate for tail risk forecasting, their predictive abilities were different. The findings suggest that GAS model with heavy-tailed distributions are appropriate for modelling financial returns. The best model across all financial in- stitutions is the model with the Ald distribution at 1% and 2.5% risk levels. Results suggest that Absa was the least risky financial institution and required the least capital to absorb its losses. The fourth and fifth chapters use a variety of GARCH and GAS models with different distributions to predict risk and uses the Model Confidence Set approach to evaluate the forecasting performance. We compare the Value-at-Risk (VaR) forecasts delivered by alternative model specifications using the Model Confidence Set (MCS). We anal- yse the predictive and the forecasting ability of various GAS and GARCH frameworks for the daily returns of South African financial institutions for the period May 2003 to May 2023. We further examine the impact of different model re-estimation strategies, different forecasting periods and different quantile levels on out-of-sample VaR risk prediction performance of the selected GARCH models using 12 GARCH and 7 GAS models. The findings suggest that the quantile levels, different forecasting periods and different model re-estimation strategies had little impact on the forecasting accuracy of the GARCH models. Also, the findings from the joint VaR and ES models indicate that the risk models developed using the GAS framework exhibited homogeneity, are time-invariant and percentile-independent.