基于混合模型的卷烟销售量预测  

Cigarette sales volumes forecast based on hybrid model

作　　者：朱俊江[1] 何湘竹[2] 王建树 李孝禄[1] 张远辉[1] 

机构地区：[1]中国计量大学,机电工程学院,浙江省杭州市310018 [2]中南民族大学,电信学院,武汉市430074 [3]湖北省烟草专卖局(公司),信息中心,湖北省武汉市430030

出　　处：《中国烟草学报》2016年第5期111-116,共6页Acta Tabacaria Sinica

基　　金：国家自然科学基金资助项目(No.61302191)

摘　　要：为科学制定市一级烟草专卖局的烟草投放策略,提出采用小波变换、回归分析和神经网络算法构成的混合模型对乡镇为单位的卷烟销售量序列进行预测。通过小波分解,将非平稳性销售量时间序列转化为低频分量、中频分量和残差分量三部分,分别用于模拟整体趋势、季节性波动和非平稳波动。然后,采用回归分析和神经网络算法对不同分量分别进行预测,最后将各部分预测结果叠加形成最终预测结果。以湖北省某市的卷烟销售数据为例,对所提方法进行了验证,结果表明:相比于自回归移动平均模型和神经网络算法,混合模型分别降低预测偏差率4.62%和2.58%。Forecast of cigarette sales in towns by method oftirne series analysis is meaningful, yet difficult. In order to improve the accuracy of forecast, a hybrid forecast model based on wavelet composition, autoregressive model, artificial neural network was presented： By using wavelet composition, non-stationary cigarette sales time series was decomposed into three parts： low-frequency part, mid-frequency part and residues, which corresponded to overall trend, seasonal change and stochastic volatilities, respectively. The low-frequency and mid-frequency parts were predicted by using linear regressive model, and the residues part was predicted by using artificial neural network. The method was tested by using cigarette sales data from X city in China＇s Hubei province. Results showed that the hybrid forecast model outperformed autoregressive moving average model and neural network model by reducing prediction error rate by 4.62% and 2.58%.

关 键 词：卷烟销售预测 混合模型 小波分解 线性回归 神经网络 

分 类 号：F721[经济管理—产业经济]