COVID-19 Liquor Sales Forecasting Model

Abstract

The COVID-19 pandemic brought significant shifts in consumer behavior, impacting the Consumer-Packaged Goods (CPG) sector, including a 22% sales drop for Coca-Cola in concentrate sales for soda fountains [8]. This study aims to devise a demand forecasting framework to assist CPG firms in navigating similar crises, ensuring precise demand predictions, preventing stockouts, and optimizing supply chains for profit gains, using five years of Iowa Class “E” liquor sales data employs various statistical, machine learning, ensemble, and deep learning methods across different product categories and durations. Different techniques have been applied to deal with missing data and outliers, adding new features like lag and simple moving averages (SMA) to the dataset for seasonality and trend and implementing feature engineering. The statistical method is a good starting point to get some benchmark results. The support vector regressor (SVR) model yields the best result (near 99% accuracy) out of all the models, and the outcome of the SVR model was consistent across datasets and products. Ensemble methods also produce consistent performance across products (average accuracy between 93% and 95%). Long short-term memory (LSTM) network performance was below expectation (average accuracy between 79% and 87%).