Abstract: The Manufacturing Purchasing Managers Index (PMI) is an important indicator in the manufacturing industry reflecting the performance of a country’s economy. However, traditional forecasting models have low accuracy for predicting such time series data. Focusing on the characteristics of non-linearity, volatility and limited data volume of PMI index in the manufacturing industry, a combined model based on one-dimensional discrete wavelet transform for data preprocessing is proposed. After the wavelet transform of time-series data, steady-state low-frequency data are processed by an auto regressive moving average – generalized autoregressive conditional heteroscedasticity model (ARMA-GARCH), while the gated recurrent unit (GRU) handles high-frequency data with strong volatility. The prediction results of each frequency band are combined to get final prediction result. In order to verify the effectiveness of the model, a certain amount of data with PMI indices is selected for experiments. Results show that, compared with other common models, the combined model established in this paper has better prediction accuracy and performance, where the mean absolute error (MAE), root mean square error (RMSE), mean absolute percentage error (MAPE) reach 0.00 329, 0.004 162, 0.65%.

Key words: purchasing managers index (PMI), wavelet decomposition, wauto regressive moving average model (ARIMA), generalized autoregressive conditional heteroscedasticity model (GARCH), gated recurrent unit (GRU)