Predicting business failure using artificial intelligence system

This thesis was submitted for the award of Doctor of Philosophy and was awarded by Brunel University LondonPredicting business insolvency is considered one of the main supportive sources of information for decision making for financial institutions, investors, creditors, and other participants in the business market. Financial reporting systems provide relevant information that can be used to assess the financial position of firms. It is crucial to have classification and prediction models that can analyse this financial information and provide accurate assurance for users about business health. Recent studies have explored the use of machine learning tools as substitute for traditional statistical methods to develop classification models to classify firm insolvency according to financial statement information. However, these models have no ideal classifier, since each provides a certain percentage of wrong outputs, which is a crucial consideration; every percentage of wrong response can mean massive financial losses for stakeholders. Therefore, this study proposes new insolvency classification and perdition models based on machine learning modelling techniques to develop an improved classifier. Individual modelling techniques using statistical methods and machine learning were used to develop the classification model of business insolvency. The results showed that machine learning method outperformed statistical methods. Deep Learning (DPL) achieved the highest performance based on all performance measurements used in the study, and it was the best individual classifier, with average accuracy of 97.2% using all-years dataset. Ensemble- Boosted Decision Tree classifier ranked second, followed by Decision Tree classifier. Thus, it has been proven that DPL modelling approach is useful for business insolvency classification. A key contribution in enhancing individual classifier outputs is the use of traditional combining methods with two new aggregation methods in business insolvency (Fuzzy Logic and Consensus Approach). The Consensus Approach showed the best improvement in the results of all individual classifiers with average accuracy of 97.7%, and it is considered the best classification method not only in comparison with individual classifiers, but also with traditional combiners. This study pioneers the development of a time series business insolvency prediction model with Big Data for UK businesses. The aim of the model is to provide early prediction about a business health. Three prediction models were developed based on Nonlinear Autoregressive with Exogenous Input models (NARX), Nonlinear Autoregressive Neural Network (NAR), and Deep Learning Time-series model (DPL-SA) and achieved average accuracy rates of 83.6%, 89.5%, and 91.35%, respectively. The results show relatively high performance in comparison with the best individual classifier (deep learning)