The Research On The Identification Of Financial Risk Of Listed Firms With Clustering Approaches

It has been thirty years since China started its securities markets,which have provided Chinese companies a steady financing source and supported the reform of equity of stated-own companies.The markets have lubricated Chinese economic reform by allowing the mass investors to participate in trading stocks to make the massive free cash flow into the production sectors.As the size of the markets boomed,more and more firms launched their IPOs.Meanwhile more investors joined the markets and many of them found themselves far away from acquiring any returns on investment after witnessing some financial turbulences in stock markets.Since then,it made people to realise that the returns always come with the risks.Therefore,the stocks exchanges published a rule which stated that the companies facing kicked out of the markets would be labelled as the companies of“special treatment”.Doubtlessly,the companies on the list of“special treatment”must expose to severe financial risks.“Special treatment”system attracts many scholars for it is treated as a tool used by the stock exchanges to governance the financial markets.Former literatures constructed financial models based on financial features,which are preferable for they are obvious and effective.Among these models,the logistic model is a popular binary predictor,but it comes from a supervised learning process relying on a balanced dataset.Unfortunately,the unbalance of financial failure data is inevitable because of the rare“risk warning”event identified by the stock exchanges.In recent years,more and more scholars start to use unsupervised models to analyse the issues in financial markets.Except for the principal component analysis,the clustering algorithm,which identifies the data by the inner structure rather than by binary labels,is also becoming popular.This feature gives the clustering algorithm a unique advantage.This research is based on two clustering approaches including K-means clustering and Gaussian-Mixture clustering,by which the author forms a cluster contains firms with high risk.The previous research on financial risk focus on financial ratios,which are straightforward and effective.Meanwhile,supervised learning may cause overfitting in under-sampling condition.The data can be unbalanced if too few events used in training processes and single‘Special Treatment'events are inadequate.The construction and training processes of the models are shown as below.The first part is the identification of K-means process.Firstly,based on 27 initial financial features and“special treatment”data used to mine the common features,the Euclidian distance is calculated to set as a standard in the selection of optimal combination of financial features.The specific method is using the thought of combination to make multiple feature groups,each of which contains variables less than seven,then selecting the combination with smallest Euclidian distance as the optimal.Secondly,after the features are transformed into two dimensions,the data can be used by K-mean model to form clusters and the cluster containing most special-treatment firms is labelled as high-risk cluster.Finally,it is reasonable to judge the model by recall and precision.The second part is the identification of Gaussian-Mixture Model?GMM?process.Firstly,after the process of predicting the“special treatment”events by GMM's clusters,it is obviously for us that the feature combination with highest F1-scores is the optimal combination.The specific method is to set the number of components as 2 and 3respectively before forming the feature combination which is up to²₂₇+³₂₇.Secondly,with knowing the optimal feature combination,it is the time to training the model with the data and form the risky cluster.Finally,it is the time to use the recall and precision again to judge the effectiveness of the GMM in identification of the financial risk.The main conclusions:?1?After mining the financial data of listed firms,it appears that tangible assets ratio,cash ratio,current debt ratio,non-current debt ratio represent more commonalities.?2?The risky cluster created by unsupervised learning algorithms explains the risk well.Many firms in the cluster will find themselves on the list of‘Special Treatment'in the next few years.Investors are expected to take caution in these firms.?3?The cluster created by Gaussian-Mixture Model performs significantly better in recall?58.76%?and precision?38.52%?within 1 year than created by K-Means.