Time Series and Machine Learning methods for Demand Forecasting: a case study of machine seller in prefabricated concrete industry

Abstract

In this research, Time series and Machine learning methods are conducted to find a suitable monthly demand forecasting models and parameters for a medium-scaled machine seller company in Thailand, who is currently using Simple Moving Average as the current method. Historical demand data are collected for a half decade and used as a training set and cross validation set for different forecasting models (Exponential Smoothing, Autoregressive Integrated Moving Average and Long-Short term memory). The history of demand indicates that the patterns are composed of many types of time series components, including stationary, trend and seasonality. The performance of models is measured by Relative Total Absolute Error (RTAE) because some months contain zero value of demand. The results show that Long-Short term memory is the most suitable method as compared to others.