dc.description.abstract | Due to the global financial crisis in 2007 and 2008, cause by August 9, 2017 Subprime mortgage crisis, investors began to lose confidence in the value of mortgage-backed securities, causing a liquidity risk. The financial crisis started out of control and leads to a number of large financial institutions fail or the governments have to take over and lead to a recession. Financial institutions operating lever repeatedly Foundation, the financial statements of assets and liabilities are difficult to interpret, the traditional way of viewing difficult warning bankruptcy. Especially big as Lehman Brothers Holdings Inc., once no warning bankruptcies happen will tiger overall financial system risk, every financial institution needs to think about new tools to review investment targets, which are not in potential bankrupt risk.
The main research objective of this study is using machine learning techniques to construct an optimal bankruptcy prediction model. The research target dataset is based on 6819 companies of Taiwan, which contain 220 non-bankruptcy and bankruptcy companies, respectively. In addition, there are 95 different financial indicators, which are divided into eight categories or combinations including Solvency Solvency / Capital Structure Capital Structure ratios / Other Others / Profitability Profitability / Turnover Turnover ratios / cash flow ratio Cash flow ratios / ability to grow Growth / solvency + other. By constructing different single classifiers and classifier ensembles, the study is expected to find out the best combination of financial indicators and classifier. Moreover, the performance impact when using feature Selection for dimensionality reduction is further examined.
According the experimental results, we found based on the eight kinds of financial type datasets using the CART and MLP classifier has similar AUC. For the SVM classifier, it is not applicable because it AUC is near 0.5 only. On the other hand, classifier ensembles by the Bagging and Adaboost techniques slightly perform better than single classifiers. moreover, feature selection can enhance AUC and reduce the modeling time. | en_US |