Research On Time Series Trend Prediction Method And Its Application To Stock Market

Stock is an important way for people to invest and do financial management,and has getting more and more attention from people.The accurate stock forecasting cannot help people get more and more earnings,but also help government develop the national economy to some extent.Stock prediction,with the aim of predicting the future price trend of stocks,is the key to realize profit maximization of stock investment.However,the stock price is affected by many elements,such as stock company earnings,gross national product,investor sentiment,industry prosperity and so on.Therefore,stock prediction is still a challenging task.For the prediction of stock,one of the most commonly used methods is the model based on historical stock price data,which consists of two important parts:(1)constructing a reasonable similarity measurement to describe the degree of similarity between different(price)time series curves;(2)building a prediction model of time series to accurately predict the future trends of time series.In view of the above two parts,this paper mainly does the following two aspects:(1)For the aspect of the similarity measurements,a novel similarity measurement is proposed,which is called the dynamic multi-perspective personalized similarity measurement(DMPSM),for accurately describing the similarity between time series.Specifically,firstly weighing the time series for reflecting the difference in the effect of time.Secondly,embedding Canberra distance into the DTW to measure the similarity between any time series.The DMPSM cannot only reflect the personalization of stock time series,but also eliminate the impact of singularities and cope with time shifts and warpings.(2)For the aspect of prediction models,a new model named maximum correntropy criterion-autoregressive model(MCC-AR)is constructed.In this model,the actual output and the ideal output are regarded as two random variables,and the correlation entropy is used to measure the similarity between them.And furtherly,based on maximum correntropy criterion(MCC),a new optimization function of regression model is constructed to guide the determination of regression coefficient.MCC-AR can consider the influence of abnormal data and make full use of the feature that correlation entropy can control the adjustable window by adjusting its own kernel width,so as to effectively reduce the negative influence of singularities on time series.Furtherly,based on the above two aspects of work,a personalized maximum correntropy criterion-autoregressive model is proposed in this paper.Firstly,DMPSM is used to find out the most similar sequence and then,construct MCC-AR to predict each similar sequence and obtain a predicted value.Finally,these predicted values are averaged to represent the sample series.Experiments based on the measured data of 285 stocks in Shanghai Stock Exchange show that the proposed method is insensitive to singularities,has high prediction accuracy and strong robustness.