The Used Cars' Price Forecast Based On LightGBM

After more than 30 years of development,the used car market in our country is expending fast.The amount of annual transactions of used cars has grown rapidly,from 250,000 in 2000 to14.34 million in 2020,an increase of 57 times.The used car market is facilitated by national policies,which promotes the continuous development of the market.While the market is being expanded,some problems have also been exposed.For example,the market transaction process is not transparent,there is a lack of unified evaluation standards and so on.Compared with the developed countries,the domestic used car market still has a big gap currently,which indicates that Chinese market has great potential.Therefore,it is of great practical significance to establish a scientific,standard,and unified used car valuation method.The thesis sorts out the literatures on the price evaluation of used cars,and then summarizes the experience of the researchers.Based on the characteristic price theory,10 indicators such as new car price,car brand,and the value displayed by the odometer are selected to build an evaluation system,and then the valuation model of used cars is established based on the LightGBM.Then empirical analysis is analyzed,R~2and MAE are selected as evaluation indicators,and the values of LightGBM are 0.985 and 0.491.The LightGBM model is used to analyze the importance of each indicator to price prediction.The results show that the price of the new car has the greatest impact on the price of used cars,followed by the odometer value,brand,registration time,displacement,number of transfers,emission standards,intake form,speed change type.Finally,the prediction results of LightGBM,XGBoost and Random Forest Model are analyzed and compared.The R~2of XGBoost is 0.983 and MAE is 0.831,and these two values for Random Forest Model are 0.980 and 0.719,respectively.The values of R~2for the three models are similar,but the LightGBM's MAE is the smallest.That's to say,the smaller the gap between the predicted value and the actual value,the smaller is for the error.In terms of model efficiency,to train the same amount of models,LightGBM takes a total of 99seconds(1.65 minutes),XGBoost takes 2,718 seconds(approximately 45 minutes),and the time for the Random Forest Model is 7,757 seconds(around 129 minutes).LightGBM model greatly improves the training speed of the model while improving the accuracy.People have a higher pursuit of efficiency in big data era.High-efficiency models have competitive advantage.Therefore,the proposed used car evaluation model based on LightGBM has good practical application value.