Empirical Research On CPI Portfolio Forecasting

Consumer price index(CPI)is often used to measure the level of consumption and cost of living of residents,and is closely related to everyone.This index can not only show the price change of daily life consumption of residents,penetrate into all aspects of residents' life,but also show the rate of inflation to a certain extent.Because of this,it provides a reference for the country in formulating economic policies and macro-control,which means that for the national economy,consumer price index is very important.Therefore,building a stable and accurate CPI prediction model is of great significance to the public and researchers.In order to achieve the purpose of CPI data prediction in China,an effective prediction model is indispensable.As time series data,many researchers apply different time series prediction models to CPI prediction.In addition to classic time series prediction,machine learning algorithms are gradually emerging in the field of macroeconomic prediction,such as neural networks,support vector machines,etc.However,due to the application scenarios and constraints,sometimes the prediction results are not stable,which leads to not fully mining the information contained in the data.This kind of problem can be solved by the combination prediction method.In this paper,the monthly CPI data of China is selected as the modeling object to establish the combined forecasting model.The data set contains the monthly CPI data of China,and the time range is from January 2000 to December 2019.First of all,ARIMA(sum autoregressive moving average),BP neural network and particle swarm optimization support vector machine model are established to predict the data,so as to build the basis of the combined model.Then,the combination model is built on the basis of three single prediction models.Different from many literatures,this paper proposes three different weight setting methods,including equal weight method,prediction error square sum reciprocal method and weight optimization method through genetic algorithm.As the name implies,the equal weight method divides the weight of multiple independent models into equal parts,and the sum of the weight of each model is 1;the reciprocal method of the square sum of prediction error is a way of using the information of prediction error to distribute the weight,and the weight of each single prediction model is the reciprocal of the sum of the squares of prediction error;unlike the rules of the two models,the core of genetic algorithm is to minimize the average absolute The first mock exam is used to assign the optimal weighting coefficient to a single model.In order to evaluate the performance of the three combined forecasting models reasonably and effectively,this paper compares the performance of each combined model with three single forecasting models,and also compares the effect of each combined model.Finally,the empirical analysis shows that the performance of the combined forecasting model is more stable than that of the single forecasting model.At the same time,in the horizontal comparison of three kinds of combined forecasting models with different weight settings,the method of optimizing combined weight by genetic algorithm has better effect in both prediction error and prediction accuracy.