Research On The Falsified Financial Statesment Detection And Financial Distress Prediction Based On Robust Mahalanobis Taguchi System

In a well developed stock market, financial reports of listed companies are the most important tools to evaluate companies' financial states, to forecast companies' foreground and investment returns used by variety of institutions and individual investors.Therefore, listed companies' financial statements are the focus of the interest related correlations: the investors need financial statesments to understand operating states of public companies, the listed companies often whitewash an odious situation of their reports by all means in order to attact investors. And more, because of the unusual market developing backgroud and local government's indulgency, falsified financial statesments are widely in existence in china stock markets and have been the most notorious enemy of ordinary investors.MTS is one kind of pattern recognition technology developed by Dr. Genichi Taguchi based on his famous quality engineering and robust design in 90s of 20th. MTS uses orthogonal array and design of experiment methods to resolve feature selsection problems in pattern recognition area firstly. It sets the signal-noise ratios of Mahalanobis Distance as the optimization targets, and selects sufficiency feature from two levels othogonal array. MTS has been successful applied in many engineering applications to improve the performance of the product/process both in Japanese and USA.This dissertation focuses on the theory development of MTS in robust statistics and the application of RMTS in finacial analysis. At first, it study the linerary statistical model , formation and property and dimension decrease of two level orthogonal arrays, proving that the two level OA is more practical than ordinary feature selection methods. It also discusses the ANOVA of SN ratios when the discriminant creterions are the function of category distance. Due to threshholds determinated by minimized risk Bayes decision and original MTS training samples not reflecting the practical quality loss of misclassified, we established the theory of threshhold determinents based on Taguchi quality loss function.The research pays great attention to robust character of MTS. In original MTS, the discriminant criterion of category is Mahalanobis Distance in multidimensional systems. Mahalanobis Distance can be computed by using the classical statistics X_n and S_n. From EDA and robust statistics, we know that the classical average vectors X_n and correlation matrix R_n are highly not resistent so that the MTS has lower breakdown point espacially when there are many outliers. We developed a arithmetic of MCD and substitute classical ststistics X_n and S_n by rubust location and distribution estimation (T_R (X), C_R (X)). In this way, we gain RMTS system which has a. higher breakdown point.