Portfolio Decision And Application Through Time-varying Coefficient Regression Analysis

The Markowitz’s mean-variance model plays an important role in modern portfolio theory. The traditional portfolio selection model is however a static version, without considering the time-varying characteristics of financial assets’return and risk. In addition, the classical mean-variance model solved by quadratic programming has at least two disadvantages. On the one hand, it lacks efficiency in the solution process. On the other hand, it is limited to solve a type of portfolio for variance risk. Therefore, we proposed the idea of applying mean regression and quantile regression to solving the mean-variance model and mean-VaR model. Furthermore, we consider a dynamic portfolio model through time-varying coefficient regression analysis, which provides the basic tool for investors to obtain better investment performance (higher return or lower risk) with portfolio weights changing. The main contributions of this thesis includes two parts.(1) We propose a new method, Flexible Least Square (FLS), for dynamic portfolio selection model. The new method has at least two notable advantages. First, it enables the classical mean-variance portfolio to be solved through a classical regression approach instead of a complex optimization method. Second, it can be used to solve the dynamic portfolio model to get time-varying weights. In order to illustrate the performance of our new model, we select the Shanghai Composite Index and 16 stocks of Shanghai Stock Exchange for empirical application. The empirical results show that our dynamic portfolio model is superior to those traditional portfolio models in terms of return, risk and the Sharpe ratio.(2) We propose a new quantile regression model with time varying coefficients by considering both the structure of traditional quantile regression model and the character of time varying coefficients in flexible least square approach. The new model can be estimated by utilizing a flexible quantile regression approach, which is obtained by adding dynamic error to the loss function of traditional quantile regression model. It is very powerful to investigate the regular patter of a wider data. It has wide application future because it can not only reflect time varying feature of regression coefficients, but also reveal how predictors influence on the complete conditional distribution of responses. In this thesis, the model has been successfully applied to portfolio section analysis and to constructing a VaR risk based dynamic portfolio scheme. We conduct empirical analysis and compare the new scheme with those traditional portfolio schemes, including VaR risk based static portfolio, variance risk based dynamic portfolio and variance risk based static portfolio. The empirical results show that our new portfolio scheme is superior to the other three portfolio schemes in terms of return, variance, Sharpe ratio and VaR.