Sales Forecasting Method Research For Department-level Of The Large Chain Supermarket

Large supermarket chains always have enormous departments, taking Wal-Mart as an example, in order to facilitate the management, according to the category and attributes of goods, each store has been divided into hundreds of independent departments which are seen as independent commodity “retail stores” in the “retail market”. The sales forecasting of department level, on the one hand, can help department manager master the sales situation, so as to adjust inventory and develop appropriate promotion plan; on the other hand, the summary of the statistical projections for each department can help improve the overall sales forecasting accuracy. In this paper, over 2000 departments from 45 Wal-Mart stores in the United States are taken as examples to find the optimal model and algorithm based on department-level sales forecasting. The key points are as below:1.According to the characteristics of the data and the requirements of the forecasting model, the row data with huge missing values need to be cleaned, organized, and standardlized, then integrate multiple data sets into a complete data set. Due to the limited sample size of sales data of each department, five kind of explanatory variables, a total of17, which are possibly affect the sales forecasting. First of all, historical sales data fotward one week, two weeks, three weeks and four week. Secondly, the moving average of sales in two weeks, three weeks and four weeks. Thirdly, social economic factors, including temperature, CPI, unemployment and gas prices. Forthly, promotion events variables, such as holiday clearance, sweepstakes, and free gift. The last one is holiday factors.2.The data set after finishing embodies the characteristics of the department level in large supermarket chains: the limited historical sales data, an enormous size and many influencing factors. Thus support vector regression(SVR), multiple linear regression(MLR)provision, Random forests(Random Forest) and single hidden layer neural network(ANN),four kinds of models commonly used in regression prediction are used to model training and validation for the sales and explanatory variables of each department.3.To optimize sales forecasting model of multi-departments, a hybrid regression model, PCA-SC-SVR, is set up based on clustering algorithm and support vector regression.This hybrid model contains the principal component analysis, spectral clustering and support vector regression, a matrix with multiple explanatory variables in the data set is seen as the object of cluster. After the cluster, support vector regression model is establishedto predict sales for different clusters. In order to validate the performance of this hybrid model, PCA-SC-SVR based on principal component analysis and k-means clustering algorithm and support vector regression is compared to the Single model single SVR without clustering algorithm. Multiple sets of experimental results show that PCA-SC-SVR hybrid prediction model can effectively optimize the prediction accuracy of a single model.