Forecasting the Prices of Used Cars : A Comparative Analysis of Supervised Learning Algorithms

We build forecasting models to predict second-hand car prices capable of predicting multiple car makes and models. Using a total of 141 car attributes we fit 15 machine learning models and rank their ability to accurately predict used automobile prices according to two performance criteria. In addition, we implement the model confidence set (MCS) procedure for model comparison. LightGBM algorithm outperforms the other 14 models acording to both RMSE and MAE criteria, and is the only model in the MCS when evaluated on the in-sample (training dataset) basis. When the comparison is done out-of-sample (on an independent test dataset) LightGBM still ranks first according to MAE but is ranked third according to RMSE. Nevertheless, it is included in the 90\% MCS and thus its performance cannot be distinguished from the performance of the top ranking model on the basis of statistical evidence. Feature importance analysis shows that the age of the car, horsepower, location, and height are some of the key attributes in generating accurate second-hand car predictions