Previsão de vendas do comércio de varejo com técnicas clássicas e de aprendizado de máquina

Abstract

Sales demand forecasting in the retail bussiness presents many challenges due to the many variables that can affect sales behavior. Furthermore, sales volume varies according to the product or product category in question. Accurate forecast plays an important part in stock replenishment operations by maintaining adequate stock levels and avoiding stockouts leading to reduced expenditures related to logistic costs. However, time series forecasting methods are not always appropriately used in the industry. The present study compares time series forecasting methods (of statistical and Machine Learning approaches) on sales data. The preditive performance of each method is evaluated by the error between predicted and observed values. A normalization is then made by using the mean absolute scaled error (MASE) between the considered model and a baseline naive method. The comparisons are made for 18 time series of aggregate sales of 18 departments. The results show that the random forest model performed best when considering aggregate time series.