An Investigation Of Cigarette Sales Forecast

With the intensification of competition in the tobacco industry,more and more related persons become to pay attention on how to forecast cigarette sales.Sales from tobacco enterprise to retailers are very different from those between manufacturers and sellers,which are often strictly planned,and those from retailers to consumers,which are very hard to track.With the gradual improvement of tobacco industry information facilities,sales information from tobacco enterprise to retailers has been saved since more than ten years ago.With these data,forecasting the sales volume from tobacco enterprise to retailers without any other hard to get data is possible.According to the above background,after analyzing relevant researches and forecast methods used by an enterprise,some problems are discovered.On this basis,an evaluation system is constructed.The scores of the existing forecast methods based on the evaluation system is possible to be further increased.So a new cigarette sales forecast method is proposed.Cigarettes sales volume in several periods can be consider as time series,which can be turned into the input and output of a supervised learning problem using moving window.After that,one neural network of MLP,RNN and LSTM should be picked and trained to forecast sales.When using this three neural networks to forecast cigarette sales,some adjust should be made,including input,output,activation function,optimization algorithm,and hyper parameters,which is illustrated in this thesis.After that,some experiments are did to adjust the hyper parameters,and also verify the effectiveness of the forecast method.Besides,with the comparison of the performance of different neural network under different situation,the picking rule of the above three networks is discovered.When the sales volume of the cigarette has an obvious periodic change,and the cigarette is valuable enough to cost extra human capital to adjust the window size,the multilayer perceptron should be chosen.Under other estimates,one of the other two networks should be chosen.The recurrent neural network performs better on forecasting weekly sales,as the long short-term memory performs better on monthly sales.Finally,to test the performance of the forecast method proposed in this thesis,it is scored with the evaluation system mentioned above.With the comparison of relevant researches and forecast methods used by an enterprise,the forecast methods proposed in this thesis get higher scores,so adapt better to the purpose of this research.