Yu-Hua Mo¹, Chao Deng¹, Fei-Jie Huang¹, Qian Tan¹, Yuan-Kun Li^{2 1}China Tobacco Guangxi Industrial Co., Ltd., China, ²PBC School of Tsinghua University, China

In order to accurately predict the ef ect of new product cigarette marketing strategy.We take 18 months of cigarette sales data in city B of province A as the research sample, take new cigarette C as the researchobject, and use the random forest method to fix the errors and missing data. Then, we first use the mature cigarette brand's short-term historical sales and multiple labeling systems including the mature cigarette brand's historical sales data, retailer sales data, merchant circle crowd portrait data. Based on various machine learning method, we calculate the fitting weights of mature cigarettes to new cigarettes and thensimulate and predict the sales trend of new cigarettes. The application ef ect test found the accuracy of new cigarette sales prediction based on the traditional LSTM model was only 33.31%. In comparison, the prediction accuracy of the new model we constructed can reach 94.17%. We address the limitations encountered in new cigarette sales prediction, and fill the research gap in new cigarette launch models.

Cigarette sales forecast; Synthetic control method; Machine Learning; Multiple labeling systems.