Data-driven Integrated Decision On Demand Forecasting And Inventory Optimization For An E-tailer

The cost,profit and customer satisfaction of an enterprise are directly affected by its inventory management level.Companies face great challenges in making inventory decisions under demand uncertainty.At the same time,with the development of e-commerce and the acceleration of enterprise informationization,the data collected by e-tailers is increasingly rich.It is of great significance to mine and make full use of the feature data,such as commodity price lists,item's promotion information and catgory,which can affect the goods' demand,in an e-tailer's inventory decisions.Based on the above background,this thesis studies the data-driven integrated decision on demand forecasting and inventory optimization for an e-tailer,which can be summarized as follows:(1)A ML-Weighted integrated decision model is proposed to address the contradiction between the on-target inventory service level and overstocking in an e-tailer's warehouse.The featue data is employed to link the order decision to “similar”scenarios sold historically.Therefore,the historical demand with “similar” situation and its weight can be used to approximate the conditional expected surplus inventory.As for the inventory service level constraint,it is required to meet the target shortage rate under the selected confidence level.Finally,ML-Weighted model can be transformed into mixed integer programming(MIP)model.(2)Prevailing machine learning predictive methods(i.e.,KNN and CART)are refitted to the purpose of prescribing optimal decisions based directly on corresponding feature data and historical sales data.In addition,simulation data is generated to show the asymptotic optimality of the proposed algorithms.It is found that with the increase of the sample size,the inventory decisions obtained by the proposed algorithms converge to the theoretical optimal decisions.(3)The real data provided by a large-scale domestic e-tailer by the name of Jingdong is employed to verify the effectiveness and advantage of ML-Weighted model and proposed algorithms.The results are compared with those of existing sequential decision methods.It is shown that,on the average,the performances of the integrated decision methods are better than those of the sequential decision methods in both the realized inventory service level and the surplus inventory level.What's more,when the on-target service level is higher than 70%,the KNN-Weighted integrated decision algorithm proposed by this thesis can achieve higher service level while reduce more surplus inventory,which provides inventory guidance for e-tailers in the era of big data.The main distinguish between this thesis and traditional inventory decision methods(i.e.,forecasting demand or estimating demand distribution at first,and then solving the optimal inventory decision)is that with the model and algrithoms proposed in this thesis,the optimal inventory decision can be obtained directly by data.