Comparing Predictive Models using k-Fold Cross-Validation for Auto Auction Prediction

摘要

According to the National Auto Auction Association (NAAA), the wholesale auto auction industry moves large numbers of vehicles at an estimate of $100 billion on a yearly basis. The objective of wholesale auctions is to sell cars efficiently, quickly, and accurately. NAAA statistics state that more than nine million cars are sold in auto auctions in the United States out of the 17.7 million cars that reached the auction in 2016. These statistics indicate that a considerable number of cars remain unsold. Exploiting an original, rich dataset consisting of car records collected from an auto auction in the United States, this study aims to develop and compare the performance of three predictive models namely: artificial neural network (ANN), decision tree (DT), and logistic regression (LR) for predicting the selling probability of cars to increase auction profits. The models were developed using sensitivity analysis of K-fold cross-validation to measure the impartial classification of the predictive models and compare their performance. Three performance measures were used for evaluation: accuracy, sensitivity, and specificity. With the objective being achieving the highest increment of profit, the best classifier in this study was found to be ANN with 72% accuracy, 75% sensitivity, and 62% specificity.