Demand forecasting of fast moving consumer goods based on modeling of time series and deep learning methods

Abstract

Demand forecasting has played an important role in organizations' ability to meet future customer requirements. In this way, current research is focused on the construction of models that reduce the margin of error in forecasting. However, most Colombian companies use classical models to project their demands, ignoring models based on Machine Learning and Deep Learning. This paper presents a comparison between the performance of a set of demand projection models for a group of Fast Moving Consumer Goods (FMCG) of the Cerescos S.A.S. company. Specifically, the demand data set for the company's goods and a set of macroeconomic variables for Colombia from 2015 to 2018, were used to forecast the demand for the first 21 weeks of 2019. After researching similar studies and analyzing FMCG demand behavior and time series, the forecasting models were chosen. Thus, Multiple Linear Regression, SARIMA-MLR and Recurrent Neural Networks (RNN) models are evaluated for each of the study goods based on 2019 actual data, using MSE, MAD, CFE, TS and MAPE as forecast error measures. According to the results, the RNN models with double LSTM layers present the highest forecasting performance for the 4 studied goods compared to the other models. Also, for inventory purposes, the use of the model with time series learning and multivariate learning is recommended. Finally, the performance of the models and their potential use for the forecasting of goods in Colombian companies is discussed.