Modelagem preditiva da relação entre preços de carros seminovos e novos no Brasil de 2013 a 2024

Abstract

This study analyzes the ratio between used and new (zero-kilometer) vehicle prices in the Brazilian automotive market, using monthly data from 2013 to 2024. The main objective was to develop a robust predictive model for this relationship, considering its strategic relevance for vehicle rental companies, financial institutions, and manufacturers in asset management and pricing. The research applied different methodological approaches, comparing three techniques: multiple linear regression, Random Forest, and XGBoost. The results demonstrate that, although tree-based models showed superior performance in the training set, they exhibited a significant tendency toward overfitting. In the evaluation with data not used in training, XGBoost outperformed the others, achieving an R² of 0.736 in the test set, compared to 0.701 for linear regression and 0.693 for Random Forest. The temporal analysis revealed a long memory in relative prices and an important inflection during the pandemic in 2020, followed by gradual normalization. The predictive model developed offers the automotive industry a valuable tool to anticipate the evolution of residual values of vehicles, contributing to more effective strategies for fleet management, asset disposal, and rental service pricing, essential elements for the competitiveness and financial sustainability of companies in the sector.