Non-parametric Regression Research And Its Application

In reality, there often appears a type of non-deterministic nonlinear relationshipamong the intricate nexus among the economic variables, and those defects of artificialerrors existing in the parametric regression model often result in the failure of meetingthe demand for economic management. However, the non-parametric regression modelcan better describe the nonlinear characteristics reflected in the sample data and thisnew model has aroused widespread attention nowadays. The present thesis studies thenon-parametric regression theory and its applications in economic science. The maincontents are as follows:Firstly, based on the exploration of the non-parametric regression model, the paperputs forward the argument that the different selection methods can be adopted as theclassification foundation of the non-parametric regression analysis, which lead to thegeneration of the types including partial as partial regression analysis, spline regressionanalysis and the orthogonal regression analysis. In addition, the author of the presentthesis, aiming to the different types of non-parametric regression analysis and focusingon the comparative chart of the fitting and simulated values, respectively implementsthe numerical simulations and draws the conclusion that the bandwidth or thesmoothing parameter is the main factors to influence the fitting result.Secondly, the thesis gives the definite procedure of the NAR(p) model and alsodeduces the model by adopting the non-parametric cycle-forecast method. In addition,the author of the present thesis carries out the empirical study of the year-on-yearhousing-sales price index in our country from November of2006to December of2010and finds that both the fitting value and the real value are quite good in the aspect of thecomparative error of the fitting result and the real result. Thirdly, in the present thesis, the author integrates the Bayesian method and thenon-parameter partial linearity regression and discovers that the Bayesian method canbe adopted to handle the situation of the variable bandwidth in m (.) and then gives theposterior distribution of the variable bandwidth. The thesis also probes the samplingmethod of the posterior distribution of samples and the steps of the Bayesiannon-parametric local linear regression algorithm. By doing the research of theyear-on-year housing-sales price index in our country from November of2006toDecember of2010using the Bayesian non-parameter partial linearity return algorithm,the author discovers, in the aspects of the average-absolute error and mean-square error,the fitting and the inference effects obviously surpass the NAR(p) model.Finally, the author summarizes the thesis’s work and proposes the direction of thefurther research.