Year: 2026
Type: article
Abstract: Background: Used car prices are influenced by model year, mileage, fuel type, transmission, engine size, engine power and market position. This study compares global and segment-aware machine learning approaches for used car price prediction using the CarDekho dataset from Kaggle. Methods: After 167 exact duplicate extras were removed, 15,244 records were evaluated. The primary analysis used ten sealed 80/20 outer splits. Segment topology was selected by three-fold cross-validation within each outer-training set, and fixed CatBoost/XGBoost expert profiles were combined through an XGBoost routing classifier. Global, hard-routing, soft-gating and oracle variants were compared. All target models used log1p prices and a three-fold out-of-fold Duan retransformation correction. Results: Across the full-market splits, global XGBoost achieved R²=0.8523±0.1120, RMSE=349,460±180,834 INR and MAE=104,138±5,274 INR. Soft gating achieved R²=0.8733±0.1007, RMSE=321,661±173,497 INR and MAE=99,196±5,479 INR. Paired Wilcoxon p values were 0.0039 for R² and RMSE and 0.0020 for MAE. In the seed-42 paired bootstrap, the R² and RMSE intervals included zero, whereas the MAE reduction remained supported. The secondary IQR-restricted analysis yielded almost identical mean R² and RMSE for global and soft-gated models, with a lower MAE for soft gating. Conclusion: Segment-aware modeling reduced average full-market error, but the size of the advantage varied across splits. The contribution is a controlled empirical comparison rather than a new learning algorithm.
Authors: Rukiye Nur Kaçmaz, Hakan Onay
Institutions: Erciyes University, Division of Undergraduate Education
Language: English
Subfield: Information Systems
Field: Computer Science
Domain: Physical Sciences
Open Access status: bronze
OpenAlex ID: W7221774055