Forecasting China’s Total Retail Sales of Consumer Goods: A Multi-Model Comparative Approach

Abstract

An accurate forecast of the TRSCG is an important requirement for macroeconomic monitoring and policymaking since it represents a valuable index related to the domestic consumption of China. In the current research, 304 monthly values of TRSCG from January 2000 to April 2025 are obtained from the National Bureau of Statistics of China for comparing the accuracy of statistical, machine learning, and deep learning techniques. Data preprocessing, trend decomposition, and stationarity analysis (Augmented Dickey–Fuller test) have been performed, after which the original non-stationary series has been made stationary using first-order differencing. Four forecasting methods, including ARIMA, SARIMA, LightGBM, and Long Short-Term Memory (LSTM), are tested using RMSE, MAE, and MAPE measures of errors. It was found that the SARIMA method is much more accurate than ARIMA thanks to its ability to capture seasonal patterns; meanwhile, LightGBM gives good forecasting results when using macroeconomic factors. Among all other methods, LSTM is the most accurate one thanks to the ability to capture nonlinear patterns and long-term temporal dependency.