What Drives the Price of a Used Car?

This project analyzes over 400,000 used car listings to identify what factors influence a car's price. I built this for a dealership that wants to fine-tune its inventory and pricing strategy based on real-world data. The project follows the CRISP-DM process and is structured for both technical and non-technical audiences.

🧠 Business Objective

• Understand which vehicle attributes increase or decrease resale value
• Guide dealership decisions for pricing and inventory
• Present insights clearly for business use i.e. which brands are overpriced or undervalued

📁 Dataset Overview

• Original source: Kaggle (3M+ listings)
• Filtered version used: ~426,000 listings
• After cleaning: ~93,000 dealership-relevant listings
• Includes fields like:
  • Make and model
  • Year, mileage, fuel type, condition
  • Title status, transmission, drive type
  • Region and color

🛠️ Data Cleaning and Preparation

• Removed listings with invalid VINs, duplicate entries, or missing critical data
• Filtered out junk i.e. cars listed for under $1,000 or with salvage titles
• Standardized inconsistent model names (e.g. “F 150” to “F-150”)
• Created:
  • car_age (based on listing year)
  • log_price (to normalize skewed price distribution)
  • region_grouped (to reduce high-cardinality noise)
• One-hot encoded all categorical variables for modeling

📊 Modeling Process

• Split data into training and test sets (80/20)
• Used pipelines with ColumnTransformer for scaling and encoding
• Trained 3 models:
  • Linear Regression
  • Ridge Regression (best performer)
  • Lasso Regression (feature shrinkage)
• Tuned with GridSearchCV and LassoCV
• Evaluated using:
  • RMSE
  • R²
  • 5-fold cross-validated R²

📈 Performance Summary

Model	R² Score	CV R² (5-fold)	RMSE
Ridge Regression	0.669	0.460	0.435
Linear Regression	0.669	0.460	0.435
Lasso Regression	0.669	0.458	0.435

• Ridge was selected due to better generalization and handling of one-hot encoded features

🔍 Key Findings

• Top positive price drivers:
  • Ferrari, Porsche, Tesla
  • Low mileage and newer year
  • Clean title and excellent condition
• Top negative price drivers:
  • Models like PT Cruiser, Saturn, and Cobalt
  • Poor condition or excessive mileage
  • Older age or uncommon cylinder types
• Depreciation insights:
  • Estimated ~36% price drop per year of age
  • Estimated ~17% depreciation per odometer log unit
• Used grouped bar plots to visualize:
  • Top and bottom feature impacts by category
  • Manufacturer and model trends
  • Drive type, fuel type, and transmission effects

📂 Repo Contents

• usedCarAnalysis.ipynb – Main notebook with step-by-step analysis
• model_coefficients_comparison.csv – Feature weights for each model
• model_results_summary.csv – Evaluation metrics comparison
• vehicles.csv the dataset used for training the model
• Visuals and grouped plots inline

🚙 Who This Helps

• Dealerships pricing used cars for retail
• Data-driven sales managers and inventory planners
• Analysts trying to predict market pricing behavior

🧪 Price Predictor Tool

I also included a basic price prediction program inside the notebook. After training the Ridge model, the pipeline is saved and can be reused to estimate log-price for any new listing, which can then be converted back to dollar price.

You can test custom input data i.e. a 2018 Toyota Camry with 50,000 miles and clean condition, and the model will return a price estimate based on learned patterns.

This can help dealerships:

• Run bulk price forecasts
• Evaluate trade-ins quickly
• Make smarter purchase decisions at auctions