Forecasting Tobacco Sales Trends Based on Combined Xgboost-Lightgbm Modeling

Abstract

In the contemporary market landscape, consumers increasingly prioritize cigarette products' safety and pricing transparency. This study develops multiple time-series forecasting models to accurately predict five cigarette brands' sales volume and revenue. By integrating advanced data preprocessing techniques and feature engineering, the proposed framework addresses the challenges posed by extreme sales values and macroeconomic fluctuations. Considering that the sales in some months are "extremely large" or "extremely small", this paper introduces the Savitzky-Golay filtering algorithm for data smoothing. One of the major innovations of this paper is that in addition to product sales and average prices, other factors such as the rate of price change and household consumption index are also considered. In order to analyze the importance of the impact of different factors on sales, this paper introduces the SHAP algorithm to rank the degree of characteristics and finally selects the top three factors as the indicators of sales forecast. In this paper, two efficient gradient-boosting decision tree algorithms, XGBoost and LightGBM, are used for combined prediction, and the average of the two predictions is used as the final prediction result. The final MSE of the two cigarette brand predictions are 0.0042 and 0.0036, and the goodness of fit is 0.9056 and 0.9028, respectively, and this combination model has high accuracy.