Stacking Model Combined with Sequential-Model Based Optimization for Used Car Valuation

Private cars are considered to provide unrivaled comfort to their owners. In developing countries, buying and selling used cars is inevitable. The valuation of a used car is determined by various factors, making it difficult to determine a fair selling price for them. To address this problem, this study uses a stacked generalization (stacking) algorithm to combine machine learning (ML) methods that have shown promising results in used car valuations researched in other studies. Both the sequential model-based optimization (SMBO) algorithm and hyperparameter optimization were adopted to obtain a better accuracy. The initial price of a used car, which is a feature that affects its price, has not been addressed much in previous studies; thus, we examined it. Through this, we could create another model to determine the residual value of a car. The results confirmed that the optimized stacking model outperformed the other algorithms in terms of predictive ability on both the used car price and residual valuation models. Feature analysis indicated the significant effect of the initial price on the used car’s price. This analysis also proved that the reduction in a used car’s valuation was not only influenced by the usage level of the previous owner, which is reflected in the age, mileage, and remaining warranty period of the car, but also by the primary physical attributes of the car such as its brand, type, engine capacity, and initial price.

Download paper here