A Financial Distress Dynamic Early Warning Model Based On KALMAN Filtering And BP Neural Network

The global financial crisis triggered by the United States sub-prime crisis has been swep the world since the end of 2007.Countries were seeking remedy to cure this crisis and various rescue measures have been proposed. While the financial distress (FD) early-warning is from the company's financial empirical research, the FD early-warning is a complex, integrated dynamic management process from the current research and applications development, and its theory and practice related to early-warning theory, progress control, dynamic science tools, simulation technology and many other academic knowledge.The companies'financial state is of an ongoing character and has the cumulative effect. It is a gradual process of evolution that company has a financial distress. Also a temporary deviation from the normal should not be classified as a crisis company. The traditional static financial distress early-warning model is based on one-period cross-sectional data and predicts the financial distress, which ignored the historical time series data's the impact on the results, as well as the inherent flaws to the modeling approach resulted that the methodology is hard to promote the use of financial management. Therefore, we need research on the financial distress early-warning mechanism deeply, and develop an early-warning model attempt to the actual financial state which can track, control and early warning, also predict financial distress signals, and prevent or reduce the damage to the business.In this paper, we conclude that a listed company occurs a financial distress is the outcome of financial state, management quality and governance performance. Hence the comprehensive early warning indicator system includes financial indicators, management quality indicators and governance indicators. And we detect that the dynamic process of corporate performance, and observe the company's changes from good to bad, from light to heavy two-point threshold. Thus we establish a financial distress dynamic early warning model based on Kalman filtering. At the same time, we established financial distress static early warning model based on BP neural network since corporate governance indicators has static characteristics. Hence we improved the financial distress early warning research from two dimensions including time-series and cross-sectional. It highlights the superiority of early-warning model including corporate governance indicators. In this empirical research we use large sample, multi-variate, time series data, and both early-warning models have shown good stability and predictability.First of all, the research on the theory of the listed companies' financial distress early-warning indicators. Through deep analysis of the characteristics and causes of financial distress, we conclude that the deteriorating financial situation led to the financial distress is its immediate causes; its poor management is the internal basis triggered by financial distress, also the weakened corporate governance is the endogenous motivation which makes financial distress occurred. We selected 30 fiancial ratios and 11 corporate governance ratios to build a comprehensive indictor system, which can solve the theoretical basis and integrity issues of the financial distress early-warning indicator system.Second, we analyse and test the listed companies'financial distress early-waring data. We use CCER database and select 264 China listed companies as study samples. We chose 30 financial ratios from 1994 to 2008, including 2727 year-observation data to empirical study. Though descriptive statistics, noise pre-processing, non-parametric tests, correlation analysis, the standard normality conversion and principal component analysis, it can overcome the logic problems of.the existing domestic and international cross-section analysis.Then, we build a financial distress dynamic early-warning model based on Kalman filtering. Through its applicability for financial distress early-warning, stochastic filter model is established, which is real-time recursive algorithm by computer. The companies'observation data is as filter input and the companies'real state and parameters is as filter output. It links the time series update and observation update algorithm, and correct the model parameters, finanly formate the optimal filtering equation. In this study, MATLAB M program is used to achieve financial distress dynamic early warning model for Kalman filtering calculation and visualization. Maximum likelihood estimation is used for parameter identification. Meanwhile, according to the dynamic process of companies'performance, the financial state of a company is divided into three phases, we extract the company's transformation from good to bad, from light to heavy two-point thresholds. Comparing the real state and prediction, we identify the model'ability though prediction accuracy, sensitivity and specificity. The empirical results showed that the dynamics early-warning model based on Kalman filtering has long-term early warning capacities and dynamic correction function, which is superior to the existing financial distress early-warning models.Finally, we research on financial distress static early warning model based on BP neural network. Time-series and cross-sectional studies are as the two objective dimensions on research financial distress early warning, which can make up for each other's deficiencies. As the dynamic early-warning model require large time series data, and the trend of corporate governance data is not obvious, but as a deep reason trigger the financial distress, thus we establish a BP neural network model to predict financial distress early-warning, which consists of two sets of variables, whether include corporate governance measurement system. By building the 22 input nodes, 3 output nodes of the 13-layer neural network, its training and simulation, it concluded that the model'correct rate of 99.18% including corporate governance variables, which highlights the superiority of having corporate governance variables and model itself on the ability to identify the exact static performance.