Using local linearization to identify the dynamics of features in second-hand automobile pricing

Abstract

In this study, nonlinear feature interactions are analyzed using application oflocal linearization on machine learning algorithms in second-hand automobile price modeling. To study the phenomenon in nonlinear models, a colossal amount of data collection is conducted for three years using advanced python scraping. The collected 313570 observations are from two main largest online listings website. The automobile prices were modeled as a function of an automobile's technical, visual, geospatial and macroeconomic features. The dataset is processed, transformed and modeled using basic linear least squares regression and advanced gradient boosting algorithm to compare linear and nonlinear modeling performance. Nonlinear modeling performed significantly better performance with 93% in R2 with 6529 TL mean absolute error compared to linear model performance at 58% in R2 with 23214 TL mean absolute error. In this research, we show that high-performance nonlinear models can be built to understand market dynamics while maintaining a similar level of interpretability as linear models. This research opens new opportunities in the application of advanced analytics in business and academic research in explaining how a specific prediction is made, understanding a very complex phenomenon and developing interpretable high-performance models in risk, insurance, and health care. Understanding why a model makes a prediction is essential for trust, actionability, accountability, debugging, and many other tasks.