E-commerce Sales Demand Forecasting And Inventory Optimization

With the popularization and rapid development of e-commerce,e-commerce enterprises' orders have been demonstrating such characteristics as small quantities,multiple batches and differentiation.Accurate prediction of customer's demand and distribute goods in advance so as to optimize inventory and reduce warehouses' costs have gained wide attention in logistics industry.The existing predicting and distributing algorithms usually generate large errors and are only for small scale problems,which makes it hard to precisely solve the demand and distribution model for warehouses.As a result,enterprises fail to fulfill customers'demand timely and lessen their costs.Therefore,scientific forecasting and reasonable distributing models and solving algorithms have been the research focus in recent yearsIn this paper,the frontier distribution centers of e-commerce enterprises under a two-level distribution network are taken as our research subject.The distributing models,including replenishment distributing of suppliers and allocation of second-level distribution center,are presented.A corresponding forecasting and the inventory model solution algorithms are proposed based on mixed strategy.The object of this model are to optimize the inventory and reduce total warehouse costs for enterprises.According to the principle of data preprocessing and combined forecasting,the combined forecasting method of weighting univariate and multivariate combinatorial is used Auto-ARIMA is deployed in the univariate approach.Traditional machine learning methods and deep learning methods are utilized for multivariable methods.The traditional machine learning methods include SVR.RF and Boosting(Boosting includes AdaBoost,GBRT,XGBoost,LightGBM and CatBoost).while the deep learning methods include Simple RNN,GRU and LSTM.The algorithm was verified with the real historical sales data of the case company for 2 years.Through data analysis,the RMSE of univariate Auto-Arima is 4.796,which has small error and better performance.In the comparison of multivariable approaches,the error mean and maximum error of the deep learning method are better than the traditional machine learning method.Among them,the error rate of LSTM results is the lowest,and there is no dilemma that the error rate of individual commodities is too high,so the multivariable method adopts LSTM method.The combined forecasting method of Auto-Arima and LSTM is adopted in this case.The experiment displays that mean RMSE of combined forecasting method is 0.102,which is better than the single methodThe inventory replenishment and allocation model based on combined sales demand forecasting with replenishment lead period are established.This model optimization obj ective is to reduce the total inventory cost including the replenishment cost of each commodity,the sales loss cost caused by each distribution center failing to meet customers demand,the additional performance cost of requesting the performance of the first-level distribution center for the second-level distribution center failing to meet customers demand,and the holding cost of each commodity.Since the model will increase the number of decision variables and excessive constraints over time,this study is divided into two stages:replenishment and allocation.In the replenishment decision,the commodities that need to be replenished are firstly identified and settled by the classical EOQ model.In the allocation decision,the commodities that need to be allocated are first analyzed.For the commodities that need to be allocated,the greedy search algorithm based on priority is deployed to work out the allocation model.Considering the transportation capacity and inventory constraints,the allocation algorithm allocates multiple bi-level distribution centers from first-level distribution center.Finally,the case data and the simulation algorithm are used to authenticate the replenishment and allocation algorithm and compared with the real scheduling algorithm.The comparison results demonstrate that the algorithm adopted in this study has a lower total inventory cost.