Analysis Of CPI Forecasting In Jiangsu Province Based On SARIMA And BP Neural Network Model

With the start of the 14 th Five-Year Plan,the construction of a new Jiangsu province that is "strong,rich,beautiful and high" has made significant achievements,and a decisive achievement has been made in building a moderately prosperous society at a high level.Meanwhile,a new journey towards socialist modernization is about to begin.However,the international situation is serious and the impact of the Corona Virus Disease 2019 is far-reaching,so government workers need to work hard to ensure the stability of the market and the people’s sense of well-being and security.The Consumer Price Index(CPI),as an important indicator to measure the level of consumption as well as inflation and deflation,has far-reaching significance in analyzing and forecasting CPI.Based on previous studies,this paper explores the establishment of a suitable analytical forecasting model based on the monthly CPI data of Jiangsu Province.By introducing some theoretical knowledge of traditional time series model,BP neural network model and combinatorial model,the foundation for empirical analysis is laid.The article selects the monthly CPI data from January 2006 to December 2020 as the training sample and the data from January to December 2021 as the test sample.The model evaluation indexes are mainly selected as root mean square error(RMSE)and mean absolute percentage error(MAPE).First,several suitable models are initially identified through the process of smoothing,model identification and parameter estimation for the training samples,and then the optimal model is screened using the AIC criterion,and then the corresponding fitted predictions are performed.Then,using the same training samples,a BP neural network model was established by adding momentum terms to the traditional neural network structure and learning the training several times.Next,a combinatorial model was established.The CPI data were decomposed into linear and nonlinear parts to be processed separately.A time series model is built for the linear part,and a BP neural network model is built for the nonlinear part,which is the residual series,to obtain the combined SARIMA-BP model,and the predicted values of the two models are added together to be the predicted values of the combined model.Finally,on the basis of the two single models,a combined model based on the weight coefficient method is built,and the weights are calculated by four methods including: simple weighting method,error square and inverse method,dominance matrix method,MAE and least squares weight coefficient method.The combined models were constructed and predicted in turn,and the RMSE and MAPE values of each model were calculated.Comparing the seven models obtained,we find that the RMSE value of BP neural network decreased from 1.437 to 0.380 and the MAPE value decreased from 0.0036 to0.0033 compared to the SARIMA model.Therefore,the BP neural network model is better than the SARIMA model,but its short-term forecasting ability is not as good as that of the SARIMA model.The RMSE and MAPE values of each combined model are lower than those of the two single models,indicating that the combined model can improve the accuracy of the single model to a certain extent and can more fully extract the information in the original data.The prediction accuracy of the combined models constructed by different weight assignment methods also varies,among which the best result is the combined weight coefficient model constructed based on MAE and least squares method,with RMSE value of 0.199 and The MAPE value is 0.0017,both of which are the smallest values.In general,for time series data with both linear and nonlinear characteristics such as CPI,the combined model analysis is better than the single model analysis,and the conclusions drawn are more informative.Finally,the optimal model is used to forecast the CPI data of Jiangsu Province from January to June 2022.