Abstract
Two major challenges in demand forecasting are product cannibalization andlong term forecasting. Product cannibalization is a phenomenon in which highdemand of some products leads to reduction in sales of other products. Longterm forecasting involves forecasting the sales over longer time frame that iscritical for strategic business purposes. Also, conventional methods, forinstance, recurrent neural networks may be ineffective where train data size issmall as in the case in this study. This work presents XGBoost-basedthree-stage framework that addresses product cannibalization and associatedlong term error propagation problems. The performance of the proposedthree-stage XGBoost-based framework is compared to and is found superior thanthat of regular XGBoost algorithm.