Study On Sales Prediction Based On A Kind Of New Processing Method Of Data Imbalance

After more than ten years of vigorous development,Chinese e-commerce has become a red sea market.All medium and small sellers of e-commerce platform feel the deterioration of competitive environment.Simply relying on increasing advertising and marketing costs to buy traffic in order to increase orders is no longer effective,has been unable to remain invincible in the competition.How to make fine operation,more accurate and reasonable formulation of procurement and production plans,to reduce inventory as far as possible,speed up the turnover of goods,is an effective way for small and medium-sized sellers to save costs and improve profits.Compared with using traditional statistical methods to predict sales,it is a better solution to achieve more accurate prediction of sales through machine learning.This paper studies the sales forecast of a snack e-commerce enterprise.It can be divided into two parts: The first part is the weekly summary sales forecast of whole store.The second part is to predict the sales volume of each specific product in the online single store.For the summary sales of the whole store,this paper adopts two models based on different principles: ARIMA model based on inertia and LSTM model based on memory learning,as well as the combination of the two models,to predict the sales of the online whole store.Sales forecasts for specific goods,because more detailed attributes are available.In this paper,the integration model GBDT based on tree model is used for modeling.In addition to making full use of various attribute information,numerous weak learners in the model also make the model generalization performance stronger and the prediction results more accurate.In the process of predicting the sales of specific goods,it is very difficult to accurately forecast the activity sales data due to the large volatility compared with the daily sales data.In order to obtain more accurate prediction results,first of all,the data is marked with activity label or normal label by constructing relevant indicators,so as to divide the data into active and inactive parts.Then modeling the two parts separately can better improve the prediction effect.In addition,in order to solve the problem of data imbalance,this paper proposes a data imbalance processing method which can further significantly improve the model performance and improve the prediction accuracy.This method is advanced and universal.