Research On The Sales Predict Method Based On Deep Learning

Sales forecasting of commodities has always been of great significance to enterprises,especially for short life cycle products such as fashion,inaccurate sales forecasting often leads to inventory backlog,cash flow occupancy and other issues.In order to improve the accuracy of prediction,scholars have proposed many methods,including two kinds of models: time series model and machine learning based prediction model.Time series model is based on historical sales data series to predict,can find the seasonality and trend of sales,and achieve better prediction accuracy in durable consumer goods.But for the short life cycle products represented by fast moving consumer goods,the prediction accuracy of time series is poor because of the unstable product sales sequence.In order to improve the prediction accuracy,scholars include variables reflecting users' purchasing behavior,and use machine learning method to predict.Most of the behavior variables used for prediction are structured data,which also solves relatively simple linear or periodic prediction tasks.In e-commerce environment,because pictures convey a lot of information about product style,color and material,but because of the technical limitations of traditional machine learning methods,unstructured information including commodity pictures has not been included in the prediction variables of sales forecasting model.Faced with the problems encountered in the research of sales forecasting,this paper combines the ability of deep learning algorithm in dealing with complex tasks and unstructured data,takes the most representative fast consumer goods in short-cycle products as the research object,follows the research paradigm of sales forecasting based on machine learning method mentioned in previous studies,and firstly analyzes consumer behavior.Selecting the factors that influence the sales volume including pictures,price and discount,historical sales as input variables of the model,three different neural network models,i.e.full-connected neural network,convolution neural network and cyclic neural network,are used to process structured data,pictures data and sales sequence data respectively,and a depth neural network is constructed to represent the characteristics,and then according to the above three kinds of depth gods.After the output of the network,the sales forecasting model is trained with the full-connected neural network,and tested with the actual data of a garment e-commerce company in Nanjing.Compared with the exponential regression method,the prediction of the deep learning model has no obvious lag.Compared with the shallow neural network,the prediction accuracy of the deep learning model is higher in the initial stage of commodity sales.Finally,it is found that the proposed sales forecasting method has higher accuracy than exponential regression and shallow neural network.The in-depth learning forecasting method based on unstructured data,such as pictures,not only provides a more accurate sales forecasting method for short and medium life cycle products of e-commerce,but also provides an effective in-depth learning method for management practice.