Online Forecasting Model Of Supply Chain Demand Based On Incomplete Sales Information

With the rapid development of e-commerce,demand uncertainty is increasing.Accurate demand forecasting is the key for companies to decrease demand uncertainty,gain advantages in fierce market competition.Since the real needs of consumers are often unavailable,in existing researches and practices,demand is approximated by actual sales.However,in daily research,almost all sales data that can be obtained through public channels are highly incomplete,which brings great inconvenience to the demand forecast,business analysis and other activities for researchers and enterprises.Based on the previous studies of scholars,this paper organically combines the matrix factorization model,the characteristics of the product's history sales record,and the ideas of online learning,we further research on the fitting of latent factor matrix for the incomplete sales information and construction of online demand forecast model.Specifically,considering the case that only having incomplete historical sales records,this article based on the time series characteristics of sales data use the matrix decomposition model for time series data to fitting the latent factor matrix that describes the characteristics of product sales information.Further,we consider how to establish a demand forecasting model based on the latent factor matrix and the idea of online learning,which hoped that it can provide reference historical sales data for the research and operation analysis of enterprises and researchers and expand existing demand forecasting models.First,this paper optimizes the solution process of the matrix factorization model for time series data,and proposes the model MAFTIS^ALS using alternating least squares solution for incomplete historical sales records that can be used for latent factor matrix fitting and missing value estimation.Secondly,under the framework of MAFTIS^ALS,the historical sales records are fitted into a commodity attribute factor matrix and a time attribute factor matrix.Based on the representation of the historical sales records by the latent factor matrix,we establish a model named MODF-MF based on latent factor matrix for multi-products which can be used for long-term prediction or short-term online demand prediction.Finally,in order to verify the effectiveness and usability of the model proposed in this article,in the latent factor matrix fitting problem of incomplete historical sales data,we use the accuracy of missing value estimation as a measurement.Using existing applications such as mean filling method,nearest neighbor value filling method,linear interpolation method and PMF as the benchmark algorithm,MAFTIS^ALS can achieves better results than the benchmark algorithm in three real world data sets.Under the framework of time series matrix factorization,the MODF-MF model uses online algorithms like online support vector regression and online sequential extreme learning machine as the benchmark algorithm,and uses three real data sets to conduct experiments.The results show that the MODF-MF model can more effectively use the information of historical sales records and obtain better prediction results.