An artificial intelligence model to predict financial distress in companies listed on the Johannesburg Stock Exchange

Abstract

As a prerequisite for an informed decision, a company’s financial results are undoubtedly one of the most important aspects to be considered in a financial distress prediction model. To rely purely on financial results for prediction is a risk. The dilemma is that financial variables are backward-looking and point-in-time measures of a company’s financial results. Ever-changing quantitative non-financial variables could enhance the decision-making process and should therefore be taken into consideration. For this research, an artificial intelligence model based on a unique combination of financial, market and quantitative non-financial variables was developed and tested against internationally and South African-developed financial distress prediction models in order to determine its prediction accuracy. Various levels of the artificial intelligence model were separately tested against the two statistical financial distress prediction models. Empirical results of the study proved that a financial distress prediction model enhanced with market and quantitative non-financial variables yielded more accurate results than a model based purely on financial variables. A two-pronged overview of the theoretical development of financial distress prediction models was given to establish a foundation for the development of a financial distress prediction model for the study. The reliability, popularity and further development of a statistically based financial distress prediction model were constrained. Constraints such as reliance on outdated financial information in a highly dynamic operating environment and the advent of computer technology and artificial intelligence contributed to a new era in financial distress prediction. Despite its purported success, neural networks were also subject to various limitations. In an effort to overcome the critical limitations and constraints experienced in the application of neural network models, researchers have developed derivative financial distress prediction models. Most of these models are still at the stage of static modelling and are built with sample data, which is collected over an extended period of time. However, variables in the economic and company environment change over time and if the financial distress prediction model is not aligned or adjusted to these changes, the financial distress prediction model could lead to financial distress concept drift. This important criticism against the financial distress prediction models formed the foundation of the study. In an attempt to deal with the constraints experienced with neural network models, the study applied support vector machines to the financial distress prediction problem. The main difference between neural networks and support vector machines is the principle of risk minimisation. While neural networks implement empirical risk minimisation to minimise the error on the training data, support vector machines implement the principle of structural risk minimisation to minimise the generalisation error by constructing an optimal separating hyperplane in the hidden feature space, using quadratic programming in order to find an optimal solution. The primary objective of the study was to develop an artificial intelligence-based financial distress prediction model, which incorporated a unique combination of financial and quantitative non-financial variables from a South African perspective. The intention with the proposed financial distress prediction model was to provide a more accurate and timeous company financial health and distress prediction on a financial distress continuum compared with a statistical financial distress prediction model. A phased approach was followed, first by identifying the variables most often applied to financial distress prediction studies. A principal component analysis was conducted in the final selection of financial and market variables and the model development. The leading, coincident and lagging business cycle indicators as published by the South African Reserve Bank were selected as proxy for quantitative non-financial variables. A financial distress prediction model was developed based on machine learning principles, enhanced with market and quantitative non-financial variables and compared with existing financial prediction models. The empirical results demonstrated that different combinations of financial, market and quantitative non-financial variables enhanced the accuracy of financial distress prediction models.