Authors: Dr. Jasjit Singh Samagh, Urvita and Chandan
Abstract: Accurate valuation of used automobiles remains a critical challenge in the automotive resale market, where traditional manual estimation methods suffer from inconsistency, subjectivity, and limited scalability. This paper presents a comprehensive comparative analysis of two fundamental machine learning regression techniques—Linear Regression and Random Forest—for automated car price prediction. We developed and evaluated two complete prediction systems: a web-based application using Linear Regression integrated with Streamlit, and a desktop GUI application employing Random Forest with Tkinter interface. Both systems were trained and tested on comprehensive used car datasets comprising over 6,700 vehicle records with features including brand, manufacturing year, kilometers driven, fuel type, transmission type, ownership history, engine specifications, and market pricing. The Linear Regression model achieved an R² score of 0.87, Mean Absolute Error (MAE) of 0.34 lakhs, and Mean Squared Error (MSE) of 0.18, while the Random Forest approach demonstrated superior performance with R² score of 0.94, MAE of 0.28 lakhs, and MSE of 0.60. Our comparative analysis reveals that Random Forest's ensemble learning approach captures non-linear relationships more effectively, achieving 7% higher variance explanation than Linear Regression, though at increased computational complexity. Statistical significance testing confirms that Random Forest's performance improvement is statistically significant (p < 0.01). Both systems provide real-time predictions through user-friendly interfaces—web-based for broader accessibility and desktop-based for offline usage. This research contributes practical insights into algorithm selection for automotive price prediction, demonstrating trade-offs between model simplicity, interpretability, and accuracy while providing deployment-ready solutions for diverse stakeholder requirements.
DOI: https://doi.org/10.5281/zenodo.19830626
</p