Research On PLS-based Financial Crisis Forewarning Models

The Chinese economy has entered a new phase in recent years,with phenomena including linkage deviation among major economic indicators,continuing downward economic growth and low-level function of CPI,declining profit margin,and rising consumption against decreasing investment.New situation has increased the difficulty of corporate management,proposing higher standards to forecast financial crisis.Facing more complexity in their operation when joining fierce market competition,now companies have more hidden incentives which might lead to a financial crisis.One of the important means of eraly warning is to use the models with higer prediction accuracy.So far,studies on financial crisis forewarning models has gone through two stages,traditional econometric models,such as univariate model,MDA,survival analysis,and artificial intelligence(AI)models,including genetic algorithm,expert system etc.Researches and application on present models have made fruitful results.However,some shortcomings also emerge,mainly in three aspects.Firstly,theory basis for financial crisis forewarning and its variables determination are still insufficient.Deficiencies of micro and macro theories still exist,causing many researches to be baseless.Groundless variables might affect the prediction accuracy of financial crisis forewarning model,making it useless.Secondly,variables screening methods are still imperfect.Variables screening directly affects the validity of the financial crisis forewarning models.Present mainstream approaches,like principal component analysis,typical correlation analysis etc.,face problems like multicollinearity of data in the screening process.Few breakthroughs have been made in recent years.Fortunately,some scholars have already attempted to apply device operation control and waring system,which is widely used in AI industrial control,to financial crisis forecast.Once the fluctuation of variables exceeds a certain range,a warning will be triggered automatically.Rapid development of data mining technology,together with the maturing of variables screening-related theories and application,has made it more practical to find out more objective variables with more advanced methods and instruments.Thirdly,the accuracy of crisis forecast is still not enough.Different phases of economy,different economic structures of society,different industry cycles,and different positions of firms,all raise higher standards for model construction and forecast accuracy.Hence,judging from the micro and macro environments Chinese firms faces,it is especially important to make the model more practical by improving forecast accuracy.Thus,this dissertation focuses on studying the insufficiencies mentioned above.Based on researches of the theories and application on financial crisis forewarning both at home and abroad,the author constructs the theory basis and framework of financial crisis forewarning variables to further explore problems in variant screening.Current theories and methods are deeply analyzed and Partial Least Squares(PLS)is innovatively proposed.Its advantage to variables screening has been verified through empirical analysis.Furthermore,212 samples from Shanghai and Shenzhen stock markets,between 2011 to 2013,are picked out for case study.Components of partial least squares are extracted to construct Logistic Regression Financial Crisis Forewarning Model and BP Neural Network Forewarning Model.Accuracy of both models are verified through empirical study and effectiveness test containing companies listed above.The technical routs through this dissertation consists mainly three layers.The topic is raised in the first layer.Under the backgrounds of theory researches and application,in chapter one,the author illustrates the motivation to choose this topic and the research value.The second layer is topic research,including theoretical and practical studies.Theoretical study involves review and filing of some important research literature.Current mainstream variant screening methods are analyzed through scientific argument.Considering economic structure change and the expansion demand of companies,the author deduced the advantage of PLS in variant screening,which supports subsequent empirical test from Chapters Two to Chapter Five.In the section of conclusion and testing,mainly in Chapters Six and Chapter Seven,the author applies statistical methods of empirical studies to the extraction of forewarning variables from eligible samples dating from 2011 to 2013.Afterwards,PLS is adopted for variant screening and then both PLS-based Logistic Regressions Financial Crisis Forewarning Model and PLS-based BP Neural Network Forewarning Model are constructed and tested for validity.The third layer,Chapter Eight,is for conclusion of research,innovation,insufficiencies and guidelines for future research.Innovative ideas of this thesis include the following two points:1)Innovatively taking the screening of financial crisis forewarning variables as an independent module.Adopting current mainstream screening methods of financial crisis forewarning such as principal component analysis,typical correlation analysis etc.,the author conducts in-depth research and points out problems inside these methods,that is,multicollinearity of variables.The author also introduces theories and principles to apply PLS on variables screening.Furthermore,an empirical study has been made to compare current methods with PLS.Combining theory and practice together,the study proves PLS to be more reasonable than traditional methods,proving its key role in improving the accuracy of financial crisis forecast.2)Constructing the financial crisis forewarning models with higher prediction accuracy.The author used both theoretical and practice method,constructed variables of financial crisis forewarning,which including financial indicators and non-financial informations.Extracting high-quality ingredients through PLS,removing overlapped indicators,then applying PLS to build Logistic Regressions Financial Crisis Forewarning Model and BP Neural Network Forewarning Model on the basis of traditional measurement and artificial intelligence,the dissertation tests and verifies the accuracy of financial crisis forewarnings model by means of empirical test and validity checking,which improves prediction accuracy of both models.Test result shows PLS-based Logistic Regression Model improves the accuracy by 1%-4% in 2014,3% to 4% in 2015,comparing to Logistic Model based on principal component analysis,while PLS-based BP neural network model raises the accuracy by 4%-5% from BP model based on principal component analysis in both 2014 and 2015.