Car Price Prediction

A production-style machine learning pipeline for predicting used car prices from Craigslist listings. Built with a clean module structure, reproducible preprocessing, cross-validated model training, and a proper test suite.

---

Problem

Used car pricing is notoriously opaque. Listings on platforms like Craigslist vary wildly for similar vehicles. This project builds a regression model that estimates a fair market price given vehicle characteristics — useful for buyers, sellers, and dealerships alike.

Dataset: Craigslist Cars & Trucks — Kaggle (~420k listings, US-wide)

---

Results

Model	MAE	RMSE	R²	MAPE
Random Forest	~$2,400	~$4,100	~0.87	~18%
XGBoost	~$2,100	~$3,700	~0.89	~16%

XGBoost outperforms on all metrics and is selected as the final model.

---

Project Structure

car-price-prediction/
├── data/
│   ├── raw/           # Original dataset (not committed)
│   └── processed/     # Cleaned data (not committed)
├── src/
│   ├── data/
│   │   ├── load.py        # Data loading
│   │   └── preprocess.py  # Cleaning + feature engineering
│   ├── features/
│   │   └── encode.py      # Encoding + train/test split
│   ├── models/
│   │   └── train.py       # RF + XGBoost training with CV
│   ├── evaluation/
│   │   └── metrics.py     # Metrics + feature importance
│   └── pipeline.py        # End-to-end orchestration
├── notebooks/
│   └── 01_eda.ipynb       # Exploratory data analysis
├── tests/
│   ├── test_preprocess.py
│   └── test_metrics.py
├── outputs/
│   └── models/            # Saved model artifacts
├── requirements.txt
└── setup.py

---

Setup

# Clone and install
git clone https://github.com/armanesh/car-price-prediction.git
cd car-price-prediction
pip install -r requirements.txt
pip install -e .

Get the data:

1. Download vehicles.csv from Kaggle
2. Place it in data/raw/vehicles.csv

---

Run the Pipeline

python -m src.pipeline

This will:

1. Load and clean the raw data (filters extreme prices, old vehicles, high mileage)
2. Engineer features (vehicle_age, age × odometer)
3. Encode categoricals and split train/test (80/20)
4. Train Random Forest and XGBoost with 5-fold cross-validation
5. Print evaluation metrics and top feature importances
6. Save models to outputs/models/

---

Run Tests

pytest tests/ -v

---

Key Design Choices

Feature engineering: vehicle_age (derived from year) outperforms raw year as a feature. The interaction term age × odometer captures that high mileage matters more for older cars.

No target encoding: Used OrdinalEncoder with unknown_value=-1 to handle unseen categories safely at inference time, keeping the pipeline simple and reproducible without leakage risk.

Model selection: XGBoost with histogram-based tree building (tree_method="hist") for efficiency on ~400k rows. Cross-validation run on training set only, test set held out until final evaluation.

---

Top Predictive Features (XGBoost)

Based on feature importance scores:

1. odometer — mileage is the strongest signal after price range
2. vehicle_age — age degrades value non-linearly
3. age_x_odometer — interaction captures wear compounding
4. manufacturer — brand premium / discount
5. type — trucks/SUVs vs sedans have distinct price distributions

---

Author

Ali Rahbarimanesh — Data Scientist & AI Engineer
LinkedIn · GitHub