Comparison of individual,ensemble and integrated ensemble machine learning methods to predict China’s SME credit risk in supply chain finance

Supply chain finance (SCF) becomes more important for small- and medium-sized enterprises (SMEs) due to global credit crunch, supply chain financing woes and tightening credit criteria for corporate lending. Currently, predicting SME credit risk is significant for guaranteeing SCF in smooth operation. In this paper, we apply six methods, i.e., one individual machine learning (IML, i.e., decision tree) method, three ensemble machine learning methods [EML, i.e., bagging, boosting, and random subspace (RS)], and two integrated ensemble machine learning methods (IEML, i.e., RS–boosting and multi-boosting), to predict SMEs credit risk in SCF and compare the effectiveness and feasibility of six methods. In the experiment, we choose the quarterly financial and non-financial data of 48 listed SMEs from Small and Medium Enterprise Board of Shenzhen Stock Exchange, six listed core enterprises (CEs) from Shanghai Stock Exchange and three listed CEs from Shenzhen Stock Exchange during the period of 2012–2013 as the empirical samples. Experimental results reveal that the IEML methods acquire better performance than IML and EML method. In particular, RS–boosting is the best method to predict SMEs credit risk among six methods.

     

Correlation Structure and Evolution of World Stock Markets: Evidence from Pearson and Partial Correlation-Based Networks

We construct a Pearson correlation-based network and a partial correlation-based network, i.e., two minimum spanning trees (MST-Pearson and MST-Partial), to analyze the correlation structure and evolution of world stock markets. We propose a new method for constructing the MST-Partial. We use daily price indices of 57 stock markets from 2005 to 2014 and find (i) that the distributions of the Pearson correlation coefficient and the partial correlation coefficient differ completely, which implies that the correlation between pairs of stock markets is greatly affected by other markets, and (ii) that both MSTs are scale-free networks and that the MST-Pearson network is more compact than the MST-Partial. Depending on the geographical locations of the stock markets, two large clusters (i.e., European and Asia-Pacific) are formed in the MST-Pearson, but in the MST-Partial the European cluster splits into two subgroups bridged by the American cluster with the USA at its center. We also find (iii) that the centrality structure indicates that outcomes obtained from the MST-Partial are more reasonable and useful than those from the MST-Pearson, e.g., in the MST-Partial, markets of the USA, Germany, and Japan clearly serve as hubs or connectors in world stock markets, (iv) that during the 2008 financial crisis the time-varying topological measures of the two MSTs formed a valley, implying that during a crisis stock markets are tightly correlated and information (e.g., about price fluctuations) is transmitted quickly, and (v) that the presence of multi-step survival ratios indicates that network stability decreases as step length increases. From these findings we conclude that the MST-Partial is an effective new tool for use by international investors and hedge-fund operators.