Application Research Based On Dependency Nearest Neighbor Weighting Algorithm In WSVR And ?-TSVR

In the era of big data,the high-frequency data of stock market contains almost all the information in the market.If the analysis and prediction is based on high-frequency data,it will be more valuable.At present,most of the stock market researches are based on the daily data,but there is a lack of research on high frequency data.In order to deepen the research of high-frequency data and the mining of the local information of data,a Dependency Nearest Neighbor(DNN)weighting algorithm including the estimation of the model parameters are put forward in this dissertation.Then the DNN weighting algorithm is applied to improve the Epsilon Twin Support Vector Regression(?-TSVR).We also provide the derivation process to the dual problem of the improved model.The empirical results of daily data and minute data of A-share in Shanghai Stock Exchange(SSE)show that the revised model is very effective in extracting the local information of the samples.The main work of this dissertation is as follows:Firstly,we proposed DNN weighting method.DNN algorithm calculates the nearest neighbor points according to the position of each sample in the sample space.This dissertation proposed a DNN weighting algorithm based on DNN algorithm.The weighting algorithm measures the distance between samples and the dependency of sample points,which can better mine and retain the local features of sample points.Then the grid search method is proposed to determine the parameters of the weighting algorithm.And the DNN weighting method is applied to Weighted Support Vector Regression(WSVR).The experimental results in the UCI public dataset and stock data show that the proposed method has high fitting accuracy and verifies the feasibility and effectiveness of the DNN weighting algorithm.Secondly,we improved the ?-TSVR model basing on DNN weighting algorithm.The improved model is named DNNW-?TSVR for short.In order to process high frequency data,a DNN weighting algorithm is proposed in the ?-TSVR model.And grid search method is used for determining optimal parameters of optimal Dependent Region(DR)domain.The derivation process of the dual problem is given.Also,the dual problem of the improved model is solved by Succesive Over Relaxation(SOR)algorithm.Thirdly,the improved DNNW-?TSVR model is used for forecasting stocks' close price in the Shanghai Stock Exchange.We collect the daily data,60-minute data and 5-minute price data of 25 stocks of China's SSE A shares.Then we calculate some technical indicators for forecasting the closing price.The empirical results show that the improved model has good prediction and generalization performance,especially for high-frequency data.It shows that the improved model can extract and retain the local information of the samples.The improved model also can predict very well.Based on the improved D-nearest weighting algorithm and the construction of DNNW-?TSVR model,this dissertation uses a variety of data samples to nonlinearly analyze and predict the volatility of financial markets,and it has good effects.