Prediction based on traditional network prediction model and LSTM deep neural network model

Abstract

The total retail sales of consumer goods is an important index to measure a country's economic development. For time series data sets, most traditional prediction models often do not make full use of the potential spatial correlation between variable pairs, which is time-consuming and has complex limitations. As a successful variant of recursive neural networks, long short-term memory networks (LSTM) have stronger nonlinear dynamic storage capability of continuous data than traditional time series models. The purpose of this study is to compare the traditional network forecasting model with the new neural network forecasting model, LSTM. Then find out the model with the least error, provide the government with an accurate forecast of the total retail sales of consumer goods in China. By comparing the two traditional prediction models with the neural network model LSTM, a more accurate long-term prediction model is obtained. LSTM model is an ideal method to process nonlinear data. The error index obtained by the deep neural network prediction model was the lowest among the tested models, and the mean square error (MSE) and mean absolute percentage error (MAPE) were the smallest, which were 1.4839 and 0.0037, respectively. This forecast can be used by governments to guide their decisions to promote faster economic growth in the future.