Forecasting And Assessing Of The Risk For Chinese Stock Market

Ever since the entrance of WTO in 2001, China has been faced with the ever-changing international tendency in its financial market, which makes the financial market risk management very important, thus the measurement tool of financial market risk—value-at-risk (VaR) has become the standard for many countries to estimate its financial market risk.Therefore, the counting methods of VaR have been the hot topic for many researchers.According to the fundamental theory and main methods of VaR counting, this article briefly introduces the common models of international VaR counting:simple moving average models, exponentially weighted moving average models, various ARCH models, historical simulation and its improved models. we will apply those VaR models in logarithmic return series of Shanghai stock comprehensive index by comparison, reference, quotation, etc, and will have some empirical analysis below:Firstly, the Statistical Characteristics of sequence are discussed according to the theories and methods of analysis of Financial time series, and it turns out that the mean value is zero and it doesn't follow normal distribution but with the characteristics of leptokurtosis and heavy tail; sequence is stable and not auto correlative, and it is not totally independent but with higher order autocorrelation. ARCH effect exists in sequence.Secondly, VaR models are tested by three likelihood ratio tests of Christoffersen (1998). The test result shows that the models considered have good probability of the unconditional or conditional coverage failure in the stock market in China. Thus more VaR models can be offered.Then, VaR models are compared. The first step is to choose three loss functions as the standard to evaluate and to choose the most classical Riskmetrics model as the benchmark model, then uses the "reality check" of White (2000) and the stationary bootstrap of Politis & Romano to compare between the models. The result shows that the direct comparison between two models can produce models with better predictive ability than the benchmark model. Given the problems of data-snooping, more models are compared with benchmark model contemporarily, and the result is that only fewer models have better predictive ability than the benchmark model.All considered, none of the VaR models can get the best predictive ability under different standards and different Confidence levels.