中国烟草学报

目前，新品卷烟投放预测存在数据区间较短、数据波动较大等问题，制约了传统基于时间序列的卷烟投放模型在新品卷烟上的应用。为解决这一问题，创新性地将合成控制法与机器学习模型结合，采用稳定的成熟卷烟产品销量趋势对新品卷烟的销量趋势进行模拟和校准。以广西玉林市18个月卷烟销售数据为研究样本，将真龙（海韵中支）作为研究对象，基于随机森林算法对数据中缺失值进行相应处理后，通过合成控制法计算成熟卷烟产品对新品卷烟的拟合权重。应用效果检验发现，基于长短期记忆模型（LSTM）对新品卷烟预测的准确率仅为33.31%，在加入模型标签体系的多种机器学习模型后，集成学习模型（Weighted Ensemble）算法准确率可达到88.56%。最终，在将合成控制法和多种机器学习模型融合后，算法准确率达到了94.17%。该算法解决了当前新品卷烟销量预测局限性问题，并填补了新品卷烟投放模型的研究空白。

At present, new product cigarette launch forecasts are hindered by short data intervals and significant fluctuations, limiting the use of traditional time-series models for new varieties. This paper innovatively combines the synthetic control method with machine learning to simulate and calibrate new cigarette varieties' sales trends using data from mature products. This paper takes 18 months of cigarette sales microdata in Yulin City, Guangxi as the research sample, and takes the new product Zhenlong(Haiyun Medium) as the research object. After processing the missing values in the data accordingly based on the random forest algorithm, the fitted weights of mature product cigarettes to new product cigarettes were calculated by the synthetic control method. The application effect test revealed that the LSTM model's prediction accuracy for new varieties was only 33.31%. In contrast, the Weighted Ensemble algorithm, enhanced with multiple machine learning models, achieved an accuracy of 88.56%. By fusing the synthetic control method with multiple machine learning models, we achieved an algorithm accuracy of 94.17%. This research overcomes limitations in predicting new variety cigarette sales and bridges the gap in new product launch models.