Machine Learning for Corporate Bankruptcy Prediction

Corporate bankruptcy prediction has long been an important and widely studied topic, which is of a great concern to investors or creditors, borrowing firms or governments. Especially due to the recent change in the world economy and as more firms, large and small, seem to fail now more than ever. The prediction of the bankruptcy, is then of increasing importance. There has been considerable interest in using financial ratios for predicting financial distress in companies since the seminal works of Beaver using an univariate analysis and Altman approach with multiple discriminant analysis. The big amount of financial ratios makes bankruptcy prediction a difficult high-dimensional classification problem. So this dissertation presents a way for ratio selection which determines the parsimony and economy of the models and thus the accuracy of prediction. With the selected financial ratios, this dissertation explores several Machine Learning methods, aiming at bankruptcy prediction, which is addressed as a binary classification problem (bankrupt or non-bankrupt companies). They are OP-KNN (Publication I), Delta test-ELM (DT- ELM) (Publication VII) and Leave-One-Out-Incremental Extreme Learning Machine (LOO-IELM) (Publication VI). Furthermore, soft classification techniques (classifier ensembles and the usage of financial expertise) are used in this dissertation. For example, Ensemble K-nearest neighbors (EKNN) in Publication V, Ensembles of Local Linear Models in Publication IV, and Combo and Ensemble model in Publication VI. The results reveal the great potential of soft classification techniques, which appear to be the direction for future research as core techniques that are used in the development of prediction models.In addition to selecting ratios and models, the other foremost issue in experiments is the selection of datasets. Different studies have used different datasets, some of which are publicly downloadable, some are collected from confidential resources. In this dissertation, thanks to Prof. Philippe Du Jardin, we use a real dataset built for French retails companies. Moreover, a practical problem, missing data, is also considered and solved in this dissertation, like the methods shown in Publication II and Publication VIII.