Evaluation and Comparison of Linear Regression Models for Predicting the Asking Price of Used Cars: A Case Study of Car Listings on Divar
Evaluation and Comparison of Linear Regression Models for Predicting the Asking Price of Used Cars: A Case Study of Car Listings on Divar
Fereshteh Dehkhoda1
1) Master’s Degree in Computer Software, Damavand Science and Research University
IOI Code
:
ACEF-ACAGBD
Publication :
6th. International Congress On Engineering, Technology and Applied Sciences(etas2027.com)
Abstract :
Estimating the price of used vehicles is a multivariate problem, given the variations in technical specifications, physical condition, and mileage. This study aims to compare the performance of regression models and examine the impact of response variable transformation, feature selection, and regularization on predicting the asking prices of used Peugeot vehicles. The initial dataset consisted of 9,512 listings from the Divar platform across the cities of Tehran, Mashhad, Isfahan, and Shiraz; after removing seven observations with suspicious price values, 9,505 observations remained for analysis. Preprocessing steps included standardizing the production year, calculating vehicle age, aggregating low-frequency categories, and encoding categorical variables.
The data were divided into a training set with 6,653 observations and a test set with 2,852 observations. Five approaches were compared: basic linear regression, linear regression with a Box–Cox transformation, a Box–Cox model combined with feature selection, and Ridge and Lasso regressions incorporating this transformation. Feature selection and the tuning of regularization parameters were performed using ten-fold cross-validation.
The test set results indicated that, among the models evaluated, linear regression with a Box-Cox transformation and all features achieved the lowest Mean Absolute Percentage Error (MAPE) at 9.33%; for the baseline model, this metric was 11.96%. The Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) for the selected model were calculated to be approximately 31.56 and 17.20 million Tomans, respectively. Furthermore, 87.38% of the model s predictions fell within a range of ±15% of the price listed in the advertisement, exceeding the target threshold of 80%. In this experiment, feature selection and regularization did not yield further improvements over the Box-Cox model that utilized all features.
The findings indicate that transforming the response variable can improve the performance of linear models in estimating the asking price of vehicles. However, generalizing these results to other time periods, different brands, or final transaction prices requires independent evaluation.