Application of Long Short-Term Memory for prediction of monthly tobacco sales volume

Abstract

The prediction of tobacco sales is crucial, for it informs businesses about future demand trends, facilitating production optimization and minimizing stockouts or overstocking. The tobacco market's complexity, with numerous influencing factors and rapid fluctuations, poses significant challenges for accurate sales forecasting. Responding to this challenge, this paper introduces the application of the LSTM (Long Short-Term Memory) network approach to develop a predictive model to accurately forecast future sales volumes. This paper initiated by conducting data processing on open-source tobacco sales datasets, utilizing box plots to identify and eliminate outliers, and subsequently filling in missing values with Hermite interpolation, thereby completing the preprocessing stage. Following this, based on the distinct sales data of four different brands, this paper independently established an ARIMA (Autoregressive Integrated Moving Average) time-series prediction model and a LSTM (Long Short-Term Memory) network prediction model seperately. Subsequently, the performance of these two prediction models was examined. Upon comparison, the ARIMA model exhibited suboptimal performance in predicting nonlinear and complex time series, thereby leading to the conclusion that the LSTM model outperforms ARIMA in terms of predictive capability. Compared with the ARIMA time series prediction model, the LSTM model has better performance in terms of robustness to outliers and long-term prediction ability model fitting, which can provide analysis and support for decision-making in the tobacco market.