AI Enabled Service Market Logistics

Abstract

Uncertainty about upstream suppliers’ ability to deliver ordered quantities on time is one the reasons that manufacturers and retailers need to keep safety stock in inventory. Through accurate prediction of suppliers’ delivery performance the uncertainty can be quantified and used by material planners in their decision-making process. Representing the deliver performance of an individual supplier as a time series, the uncertainty can be predicted through probabilistic forecasting: estimation of the future probability distribution given past observations. This thesis presents two recurrent neural network models, using encoder-decoder architectures, for multi-step ahead probabilistic forecasting of the delivery performance of suppliers to Volvo Group Trucks Operations Service Market Logistics. The models are evaluated on mean quantile loss for a number of quantiles over a 14 week forecast range. One model, DeepAR, outperformed exponential smoothing models generated by the forecast package in R on four out of five quantiles.