Demand forecasting in production planning for dairy products using machine learning and statistical methods

Abstract

Demand forecasting is an essential task in manufacturing of every industry. Efficient forecasting relieves the excessive stock and out-of-stock problem, reducing revenue loss. This research performs a direct multistep forecast approach of demand forecasting on 8 dairy products of 5 different dairy production plants with 5-year data. Widely used traditional statistical method and the state of the art deep learning method for sequence problems are picked. ARIMA and LSTM. The models are compared in many aspects, monthly observations against weekly observations, univariate against multivariate, and statistical against deep learning using model error and business metrics. The result shows that both statistical and deep learning method are reliable and are suitable to be used in demand forecasting. There is no single best optimization algorithm. ARIMAs predict the future in an average smoothed straight line. It shows the best result on few wavering series, whereas LSTMs predict the future value follow the seasonal of series. It beats ARIMAs on strong trend series. Training the model on monthly observations provides lower error score because of monthly series generally has lower fluctuation than weekly series which is easier to forecast.