Products Demand Forecasting Based On Correlation-driven Clustering And Attributed-driven Classification

In recent years,product demand forecasting has been valued by more and more companies,and accurate demand forecasting can greatly help the company's operations.On the one hand,with accurate new product demand forecasts,companies can make correct operational decisions and help new products introduction;on the other hand,with accurate old product demand forecasts,companies can better manage inventory and reduce costs,thereby maximizing company profits.In the traditional analogue-based clustering forecasting method,there is a mismatch between clustering and prediction procedure.To solve this problem,this paper designs a product forecast method based on correlation-driven clustering and product attribute-attribute classification,which captures the correlation between the sales and its predicted features(Feature-response Pattern)of each old product in the training set.Similar correlations of the old products will be clustered to obtain multiple product clusters.We use machine learning algorithms to train the forecast model for each cluster.These forecast models contain all the information of all products with similar correlations in the clusters.For new products demand forecasting,it lacks historical data and needs to use historical old products to provide the information needed for forecasting.For old products demand forecasting,the prediction accuracy can also be improved by the historical data of similar old products;therefore.We adjusted the existing distance calculation method,and calculated the product attribute discrete coefficient of each category and the attribute distance between the target products and each category by using the product attributes shared by the target products and the historical old products,thereby obtaining the effective similarities between target products and different categories,target products will be assigned to the category which has highest degree of similarity.This article uses the sales data of an online store on an 020 platform in China(the platform shared online and offline inventory,online orders are satisfied through offline inventory,thus the out-of-stock rate is low,so the actual sales can be used to approximate the product demand)to designed a number of experiments,the experimental results to some extent indicate that the new method designed in this paper is feasible in the new product and old product demand forecasting problem.The forecast performance under the Random Forest algorithm are better than the traditional clustering forecasting method and non-clustering method.The forecast performance under the XGBoost and LASSO algorithms of some evaluation indicators are better than the traditional clustering forecasting method and non-clustering method.