Commodity Demand Forecast Based On Sales Records In The Supply Chain Context

Demand is the driving force of the enterprise supply system.Forecasting commodity demand at upstream and downstream nodes is the basis for all parties in the supply chain to make ordering strategies,inventory management and distribution planning.In addition,with the rapid development of the Internet economy and further integration and development of industry and the Internet,a series of new business formats have been derived,such as C2M(Customer to Manufacturer)mode,F2C(Factory to Customer)mode and new retail mode.As a result,the uncertainty of demand is increasing,which greatly increases the difficulty of predicting customers’ demand.At the same time,it also puts forward higher requirements for the sensitivity of the supply chain.Moreover,how to achieve more accurate forecast of customer demand within the lead time is important for rational formulation of the operations and management strategies of the supply chain.Therefore,it has critical realistic significance.Based on the study of scholars,this thesis effectively combines the non-parametric theory,the randomness fluctuation and the connection in time series of customer demand to build the demand prediction model in the supply chain context.Specifically,this thesis uses a specific company in the supply chain to forecast customers’ demand for a certain commodity as a modeling background.Considering the sales records of the commodities based on a data-driven perspective.And using theoretical knowledge such as Bayesian networks as the modeling basis.Furthermore,considering how to build a demand forecasting model under the situation of none side information and with side information,respectively,in order to provide a more complete decision-making reference for realistic business operations and expand the existing demand forecasting theory.First,a non-parametric commodity demand forecasting model based on multi-layer Bayesian network Hiyes is proposed under the condition of none side information,that is,only considering the historical demand of certain commodities.Secondly,under the condition of side information taking into account,a multi-layer Bayesian demand prediction model DFSI is proposed.The above two models are based on Bayesian inference and the maximization of the posterior probability as the optimization angle to derive the optimization goals of the models.In the model solution,the corresponding algorithms are designed based on the gradient descent algorithm.Finally,this thesis designs comparative experiments to verify the validity and scientificity of the two demand forecasting models Hiyes and DFSI.Among them,the Hiyes model under the condition of none side information is compared with the existing demand forecasting models such as Croston,ARIMA,H-NBSS.Moreover,the experimental results on 6 real customer demand data sets show that Hiyes can obtain more accurate prediction results than the benchmark algorithms.Then,the DFSI model with side information uses Croston,ARIMA,SSpace,H-NBSS and Gaussian Processes to verify the prediction accuracy.And the DFSI model is tested by the real sales data of Jingdong Mall and A manufacturer.The results show that the DFSI model can obtain better prediction results than the commonly used demand prediction models.