A Research On Cigarette Sales Forecast Based On MOESOA-Stacking Algorithm

Cigarette sales forecasting is the basis for tobacco commercial enterprises to promote market-oriented reform,and the accuracy of sales forecasting is related to the accuracy of cigarette launch.Too much or too little cigarette launch will cause imbalance between supply and demand in the market,resulting in various industry chaos such as inverse market price or sky-high price of cigarettes.However,since cigarette sales forecasting is a process of balancing the plan and the market,it is necessary to consider the multiplicity of objectives in forecasting.The current cigarette sales forecasting is not comprehensive enough in terms of external influencing factors,model accuracy is not high enough,and it is difficult to find the optimal parameters and does not solve the problem of multi-objective optimization of cigarette sales,which leads to the weak guidance of cigarette launch.Based on the above problems,the thesis investigates the cigarette sales forecasting problem based on historical cigarette sales data,using integrated learning and multi-objective optimization as the main means,and combining cigarette market state evaluation theory to construct and optimize the cigarette sales forecasting model,so that the forecasting results can meet the model accuracy and also be close to the market target state value.First,to solve the problem of incomplete setting of external variables in cigarette sales forecasting,data preparation and feature engineering are carried out.The external factors affecting cigarette sales were analyzed and numerically processed to form a sample set together with the historical sales data exported from the provincial cigarette marketing platform,and then undergo three steps of feature generation,feature processing and feature selection secured the reliability of the data and laid the foundation for the construction of the prediction model.Second,in order to solve the problem of insufficient accuracy of the prediction model,the Prophet-LSTM integrated prediction model is proposed.The Stacking integration method was used,Prophet and LSTM are used as the base model and linear regression is used as the meta-model,which make up the lack of interpretability in deep learning model and the lack of learning ability of traditional time-series model.After the experiments,we found that the integrated model outperformed the single base models,the MAE,MSE,RMSE,MAPE of the combined model improved by0.241-2.637,0.616-4.985,0.497-13.997,and 1.482%-14.838%,respectively,compared with the comparison models.Finally,to solve the problem of multi-objective optimization of cigarette sales and the difficulty of parameter search in the optimization process,a model tuning method based on multi-objective egret swarm optimization algorithm(MOESOA)is proposed,and the convergence effect of this optimization algorithm is verified by test functions.The optimization of the Stacking model parameters using MOESOA minimizes the prediction error while narrowing the gap between the predicted value and the market target value.This study has implications for tobacco commerce in formulating cigarette launch policies and optimizing the market state.