Research On Demand Forecasting Model For Product Of S Enterprise

With the rapid development,food consumption has changed from subsistence consumption to healthy and enjoyment consumption.People pay more attention on food's basic nutrition and health rather than satiety,goodness and safety in the past.In order to meet people's personalized needs,make-to-order production type of multi-variety,small batch production has gradually replaced mass production and became the mainstream of the market.However,for the traditional small and medium-sized food manufacturers,due to the limited hardware equipment in the enterprise site,the relative lack of management personnel,the relatively imperfect management and insufficient control of the market terminal,the enterprise demand management has brought great challenges.Bullwhip effect increases the instability of marketing,production,supply and inventory management.In order to ensure market supply and delivery rate,enterprises often reserve a large number of finished products in advance as safety inventory or carry out intermittent emergency procurement.As a result,on the one hand,the inventory of finished products and raw materials occupies a large amount of working capital,increasing the inventory cost;On the other hand,too long storage of finished products may cause safety problems,resulting in overstock and loss of sales.Company's profit margins are being squeezed.Therefore,the accuracy of product demand forecasting becomes more and more important.In view of this,the short-term demand forecasting model of products is constructed based on SARIMA(Seasonal Autoregressive Integrated Moving Average),SVR(Support Vector Machine Regression),BP(Back-Propagation)neural network and Stacking algorithm,taking the demand of S enterprise as an example,finally selecting the prediction model with smaller error through experimental comparison and analysis.Firstly,constructing the influencing factors of product requirements.Based on the existing data of the enterprise,through preliminary exploration and visual analysis,the preliminary sales volume,number of customers and promotion sales are selected as the main influencing factors.Secondly,choosing prediction method.According to the historical literature,the traditional time series model and the machine learning algorithm are both selected.Thirdly,building prediction models.In the beginning,this article constructs three single models.After considering the merits and demerits and applied situation of the SVM and BP neural network,such as SVM's great advantage in the small sample prediction,and BP neural network has good ability of learning and memory but easy to fall into local minimum,Stacking algorithm is introduced to integrate SVM and BPNN,further increasing the robustness of the model.Finally,RMSE,MAE and MAPE are used to evaluate the experimental results.This paper studied the current sales situation of S enterprise,and the monthly sales volume from April 2014 to December 2018 was analyzed.The he first 54 data were taken as the training set,last 3 data were taken as the testing set.The results showed that:On the single model prediction,SARIMA model's error is biggest because it does not take other factors into account.SYR has better prediction accuracy,BP neural network is slightly better than SVR;Further,Stacking method is better than single model in predicting effect,because Stacking algorithm reduces the risk of single item model's error,of great significance in the practical application of enterprise,and can provide reference for enterprise management decision.