Used Car Price Prediction and Feature Importance Analysis Using XGBoost and SHAP

Abstract

The used car market has rapidly expanded; therefore, the importance of the accurate and transparent pricing of a used car has increased. Traditional approaches to pricing usually rely on linear regression to infer a clear-cut mathematical relationship between vehicle attributes and pricing. This paper presents an optimized machine learning approach to predict the used car price using an Extreme Gradient Boosting (XGBoost) model. Above all, it always beats the baseline regression models. Using a dataset of 7253 records from India, it compares Linear regression, Ridge, Lasso, Random Forest, and XGBoost models. The findings from experiments reveal that XGBoost offers the highest predictive performance with a score of 0.87, and also possesses the lowest root mean squared error, compared to 0.80 for Linear Regression. To make our model more interpretable, it uses Shapley Additive Explanations (SHAP) to analyze how power, year, km driven, transmission, and engine influence price. The understanding and improvement of online pricing transparency will pave the way for consumers, dealerships, and online players to take effective action.