Used Car Price Prediction System Based on Gradient Boosting Algorithm: From Data Exploration to Interactive Web Application

This article introduces an end-to-end used car price prediction project, covering data exploration, feature engineering, model comparison, and Streamlit web application deployment. The core uses a gradient boosting regression model, which achieves a prediction effect with an R² of approximately 0.95 on real datasets, providing data-driven price references for both buyers and sellers of used cars.

Traditional pricing in used car transactions relies on experience or simple depreciation, which is difficult to reflect market dynamics. This project predicts prices through machine learning models to meet the reasonable pricing needs of both buyers and sellers. At the same time, it demonstrates a complete machine learning engineering process and provides an interactive web application to lower the threshold for ordinary users to use.

The Kaggle used car dataset (301 records, 9 features) is used, covering dimensions such as basic vehicle information, usage status, and technical configuration. Feature engineering includes calculating Car_Age (2024-Year), removing model names to avoid overfitting, and performing label encoding for categorical variables (such as fuel type). The target variable is the actual transaction price (unit: 100,000 rupees).

Four regression algorithms are compared: Linear Regression (R²=0.74), Ridge Regression (R²=0.74), Random Forest (R²=0.95), and Gradient Boosting (R²=0.95, RMSE=1.19, cross-validation R²=0.94). Gradient Boosting has the best balance between prediction accuracy and generalization ability, making it the final choice.

1. The new car guide price is the primary predictive factor; 2. Depreciation accelerates after 5 years of vehicle age; 3. Diesel cars retain value better than gasoline cars; 4. Automatic transmissions have a premium; 5. The lower the mileage, the higher the retention rate; 6. Dealer channel prices are higher than personal transactions.

A web application developed based on Streamlit. Users input information such as new car guide price, mileage, fuel type, etc., and can obtain the predicted price (including ±8% interval), depreciation percentage, and decomposition of factor impacts. Transparent explanations enhance users' trust in the results.

• Buyers: Prioritize diesel automatic transmission models with 5-8 years of age, focus on low mileage; although dealer channels are slightly more expensive, the risk is controllable.
• Sellers: Try to sell the vehicle within 5 years, choose dealer channels, and keep complete maintenance records.
• Car dealers: Integrate the model into the inventory management system to achieve automated pricing, and use the model to identify undervalued car sources to optimize procurement.

This project fully demonstrates machine learning engineering practice from data exploration to deployment. The gradient boosting model performs excellently, and the web application realizes business implementation. In the future, we can explore introducing more features (such as brand, region), trying deep learning models, building real-time price monitoring systems, or developing mobile applications.