Used Car Price Prediction

This project applies machine learning to predict the selling price of used cars based on a comprehensive set of vehicle features and historical data. The notebook demonstrates a full workflow from data acquisition and cleaning to model training, evaluation, and interpretation.

---

🚗 Project Objective

Goal:
Predict the selling price of a used car using its attributes (year, mileage, fuel type, engine specs, accident history, etc.), enabling more accurate and data-driven valuations for buyers, sellers, and automotive platforms.

---

📂 Dataset Overview

Source: Kaggle: Used Car Price Prediction Dataset
Records: 4,009 entries

Features:

• Vehicle Details: Brand, model, model year, mileage, fuel type, engine, transmission
• Color: Exterior and interior color
• Condition: Accident history, clean title status

Target:

• Price (in USD)

Data Quality:

• Some missing values in fuel_type, accident, and clean_title
• Outliers filtered (price > $150,000 or mileage > 300,000 removed)

---

🛠️ Workflow Summary

1. Data Preparation

Cleaning:

• Removed currency symbols and standardized numerical columns (price, mileage)
• Filtered out extreme outliers
• Handled missing values:
  • Filled categorical nulls with 'Unknown'
  • Dropped rows with missing price

Feature Engineering:

• Created car_age from model_year
• Extracted horsepower and engine size from engine descriptions

Encoding:

• Categorical variables (brand, model, fuel_type, etc.) encoded numerically

---

2. Exploratory Analysis

Descriptive Statistics:

• Inspected distribution of model_year, price, and mileage
• Examined feature correlations via heatmap

Insights:

• Newer cars and those with lower mileage generally command higher prices
• Accident history and clean title status impact price

---

3. Modeling

Model Selection

Multiple regression algorithms were evaluated:

• Linear Regression
• Ridge Regression
• Lasso Regression
• Random Forest Regressor
• Gradient Boosting Regressor
• XGBoost Regressor (if installed)
• LightGBM Regressor (if installed)

Preprocessing

• Feature scaling applied to numerical features
• Train-test split used for robust evaluation

Cross-Validation

• 5-fold cross-validation performed for each model to ensure fair comparison

Hyperparameter Tuning

• Grid search applied to find optimal parameters for ensemble models

---

4. Evaluation

Metrics

• R² (coefficient of determination)
• RMSE (Root Mean Squared Error)
• MAE (Mean Absolute Error)

Feature Importance

• Identified which features most strongly influence price predictions

Error Analysis

• Examined residuals and mispredicted cases to extract further insights

---

📊 Key Results

Cross-Validation Performance

Model	R² (CV)	RMSE (CV)
LightGBM	0.866	9,533.74
XGBoost	0.860	9,740.78
GradientBoosting	0.833	10,649.21
RandomForest	0.827	10,836.48
Ridge	0.640	15,653.04
Lasso	0.640	15,653.44
LinearRegression	0.640	15,653.45

Best Model: LightGBM

• Test MAE: 5,253.88
• Test RMSE: 7,892.63
• Test R²: 0.89

Top Predictive Features

• Car age
• Mileage
• Brand/model
• Engine size and horsepower
• Accident history, clean title status

Ensemble models (LightGBM, XGBoost, GradientBoosting, RandomForest) substantially outperform linear models by capturing non-linear relationships between features and price.

---

🔍 Insights

• Car age and mileage are the strongest predictors of price depreciation.
• Brand and model significantly affect baseline price.
• Accident history and clean title status impact resale value.
• Engine specs (size, horsepower) and transmission type add predictive power.

---

🚀 How to Use

Clone the Repository

git clone https://github.com/yourusername/used-car-price-prediction.git
cd used-car-price-prediction


---

## 📊 Results

Model comparison on the held-out test set; **LightGBM** performed best:

| Metric | Value |
|---|---|
| MAE | $5,254 |
| RMSE | $7,893 |
| R² | 0.89 |

Gradient-boosted trees on tabular data with mixed numeric/categorical
features — the same model family I use professionally for production
travel-time forecasting.

---

*Part of my portfolio — more at [abadeanlou.com](https://abadeanlou.com).*