Abstract
The risk of going bankrupt is of significant interest to shareholders, creditors, employees of a firm and lenders. Following the financial crisis and the spate of corporate bankruptcies in recent times, bankruptcy prediction models are being used more than ever and for a variety of purposes, including monitoring the solvency of financial and non-financial firms by regulators, assessment of loan security, going-concern evaluations by auditors, the measurement of portfolio credit risk, and assessing securities exposed to credit risk. High episodes of corporate bankruptcies in the U.K. and elsewhere have precipitated a re-evaluation of the methods that underpin corporate bankruptcy prediction.A plethora of literature is devoted to assessing the risk that corporate firms will go bankrupt. Two dominant tools used in developing bankruptcy prediction models are statistical and machine learning algorithms. The scope of this thesis is limited to reviewing some of the most frequently used bankruptcy prediction tools in the literature.
This thesis reviews the predictive abilities of two statistical bankruptcy predictive tools, namely Multi-Discriminant Analysis (MDA) and Logistic regression. Also, the following artificial intelligence tools were examined - Artificial Neural Network (ANN), Support Vector Machines (SVM), Rough Sets (RS), Case-Based Reasoning (CBR), Decision Tree (DT) and Genetic Algorithms (GA). We included the Hazard Model, a popular tool used in predicting a firm's time to default.
We provide a systematic review and quantitative meta-analysis of corporate bankruptcy prediction models. The systematic literature review (SLR) allows us to provide syntheses of the state of knowledge in the bankruptcy prediction field to identify future research priorities. Our systematic literature review shows the Artificial Neural Network (ANN) to be the most popular machine learning algorithm used in bankruptcy prediction studies. Still, the results from the ANN studies differ in accuracy level and other performance evaluation metrics.
The use of systematic reviews to summarise and appraise the literature on bankruptcy predictive models is gathering pace. However, none of the previous systematic review studies in this knowledge domain has used a meta-regression to create a model describing the linear relationship between study-level covariates and the effect size. Also, the criteria used to assess the methodological qualities of the primary studies included in the reviews are unclear in previous studies. Moreover, most of these studies did not consider clustering in the analysis (e.g., intra-cluster correlation coefficient).
The present study differs from previous studies in two ways- first, we use meta-analysis to combine quantitatively the evidence from eligible studies identified from the systematic review of the literature and explore sources of between-study variations. Second, we use meta-regression to empirically examine the effects of the study characteristics such as sample size, percentage of failed firms in the sample, model validation methods, and type of input datasets on predicting bankruptcy event rate. The results from the quantitative meta-analysis reveal evidence of between-study heterogeneity. This result is not surprising because several factors such as sample size, type of input data sets, percentage of bankrupt firms in the sample and the type of validation methods used in the primary studies can influence the magnitude and direction of the effect size. Following the meta-analysis results, we empirically explored the potential sources of these variations using meta-regression.
| Date of Award | 2022 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Chioma Nwafor (Supervisor) & Sanjukta Brahma (Supervisor) |