A Novel Hybrid Forecasting Model Based on Time Series

Abstract

The forecast of cigarette sales is crucial for tobacco companies to formulate long term development policies and optimize the inventory control system. Considering the long-term trend and seasonal fluctuation characteristics of cigarette sales sequence, a hybrid method including wavelet decomposition, auto-regression and fusion of several intelligent algorithms is proposed. The original time series of cigarette sales are first decomposed into two components of low-frequency component and high-frequency component which simulate overall trends and seasonal fluctuation, by using wavelet decomposition. Then they are predicted by auto-regression and fusion of several intelligent algorithms, and the final result is the combination of the predictions of the two components. The experimental results show that the hybrid forecasting model performed well, and its minimum MAPE (mean absolute percentage error) is 3.58%. Compared with BP (back-propagation neural network), SVM (support vector machine), and ELM (extreme learning machine) algorithms, the hybrid model proposed in this paper decreases MAPE by 2.01%, 1.58%, and 0.93%, while the stability of the model increased by 16.92%, 22.85%, and 56.09%, respectively. The proposed hybrid model provides a valid way for improving the accuracy and stability of time series forecasting.