Intellectual Technology of Analysis and Price Forecasting of Used Cars

Abstract

For the profitable sale of a used car, people should not only be guided by their own or third-party experts' evaluation, but also use all other suitable resources. Such resources can serve as price prediction systems that, using the common features of a car (such as a car manufacturer, car model, mileage, fuel type, body type, etc.), are able to predict the possible price of a car. Such systems can help in decision-making not only to ordinary car dealers, but also to agencies involved in the ordering and bulk transportation of used cars from abroad. To select the key features and identify the optimal structure and parameters of the models, relevant datasets should be selected, the intelligence analysis and selection of features will be conducted, after which building of a number of machine learning models has begun, from which the optimal model was chosen by certain criteria. In order to build an information system and test the functionality of the proposed intellectual technology, two comparable datasets for used cars of the USA and Ukraine were selected. Python methods and libraries have been systematized for intelligence analysis and general recommendations for their application for the task have been formulated. The general principles of intellectual technology, which is tested on the selected datasets, are offered. In particular, a exploratory data analysis of US data was conducted and a rule for filtering anomalous, and possibly erroneous, data was substantiated. Many possible models were selected, their training was carried out and the optimal one was selected according to the R-squared criterion. The cost of the car has been predicted to an accuracy of 86.1%. A similar problem is solved for data on Ukraine. An accuracy of 85.6% was achieved. This has proven the workability of the proposed technology and has yielded useful results in practice.