Research On Brand Automobile Sales Forecast Based On Network Search Data

In recent years,China's automobile industry has developed rapidly,but the problem of overcapacity has become increasingly prominent.Automobile manufacturers are facing severe development difficulties.Therefore,scientific and accurate automobile sales forecasts are needed to provide necessary decision support for automobile manufacturers,but automobile production and sales are Complex periodic processes have difficulty in achieving accurate predictions over different time periods.However,the existing research problems mainly include improper selection of research objects,imperfect feature selection system,and limited performance of prediction models.Therefore,based on network search data,this paper takes multiple hot-selling brands as research objects and follows machines.Study the complete research process in the subject area,apply a variety of feature selection algorithms,empirically analyze a variety of predictive models,and conduct in-depth research based on different time granularities to achieve accurate and systematic prediction of brand car sales.The main research contents include the following aspects:(1)Model construction of relationship between brand automobile sales volume and Internet search data.Firstly,the limitations of the traditional time series model are analyzed.Then,according to the process of decision-making behavior of car purchase,it is shown that the Internet search data represents the purchase intention of consumers to some extent.Finally,the model framework of the relationship between brand automobile sales volume and Internet search data is established.(2)Feature selection of brand automobile Internet search keywords.In order to reduce the subjectivity and retain maximum effective information,control model complexity,solve the multicollinearity,and the problem of redundant features based on the characteristics of engineering theory,using filtration characteristics of preliminary screening in the first place,then the candidate feature set variable on the application of model based on Lasso selection algorithm and puts forward two kinds of heuristic based on Svm and RandomForest recursive variable selection method for feature selection,according to different brand keywords library by using three kinds of schemes to get more optimal keywords feature set,prepare for model establishment.(3)Multi-time granularity prediction research on brand car sales.Using the three linear machine learning algorithms of penalty linear regression,support vector regression and random forest,the short-term prediction model of brand automobile sales is established based on multiple sets of feature subsets.The results show that the overall performance of feature subsets obtained by RandomForest-based heuristic recursive variable selection algorithm is obtained.Optimal,random forest model index evaluation is the best;for different time granularity in-depth study,it is found that the relationship between medium and long-term network search data and corresponding sales related trends is significantly enhanced,but there are differences in the optimal feature subsets at different time granularities;In the prediction,the performance of the random forest model is still the best,and the average of the optimal MAPE for all brands is 2.74% and 2.94%,respectively.By analyzing all the experimental results,the multi-time granularity prediction scheme of brand automobile sales based on web search data proposed in this paper basically realizes the whole process architecture including data acquisition,data preprocessing,feature selection and prediction modeling,and has reached More accurate forecasting needs.