🚗 Used Car Price Prediction in Germany with Machine Learning

Development of a Machine Learning-Based Prediction Model For Used Car Listing Prices in Germany

This repository contains the implementation and results of a machine learning project aimed at predicting the fair listing price of used cars in the German market. Our study focuses on resolving the problem of information asymmetry in online used car transactions by providing data-driven price estimation tools for both sellers and buyers.

---

Table of Contents

1. Project Overview
2. Motivation
3. Methodology
4. Results
5. Contributions
6. Paper
7. Technologies

---

Project Overview

• Project Type: Machine Learning Regression
• Target Market: German used car listings
• Goal: Predict realistic listing prices of used cars using multiple ML models and categorical encoding techniques
• Models Used: Multiple Linear Regression (MLR), Random Forest Regression (RFR)
• Dataset Source: Kaggle - German Used Cars Dataset

Motivation

The used car market has been rapidly expanding as consumers seek alternatives to expensive new vehicles. However, pricing uncertainty and market opacity often lead to unfair deals. Our project aims to bridge this gap by:

• Accurately predicting expected listing prices
• Helping sellers avoid undervaluation
• Empowering buyers with pricing insights

Methodology

1. Dataset & Preprocessing

• Source: German online car marketplaces (2016)

• Initial Features: 21 variables (brand, model, mileage, fuel type, etc.)

• Final Feature Selection: Focused on the top 5 most frequent car brands and frequently listed models

• Preprocessing steps:

  • Removed duplicate listings (based on vehicle ID and timestamp)
  • Created Car_Age as a derived feature
  • Removed outliers from Horsepower and Listing_Price
  • Filtered invalid or unspecified category entries

2. Feature Encoding

We applied two encoding techniques for categorical variables:

• One-hot Encoding
• Target Encoding (Mean encoding based on Listing_Price)

These were evaluated in combination with different regression models to understand how encoding influences performance.

3. Models

1. Multiple Linear Regression (MLR)
2. Random Forest Regression (RFR)

Both models were trained using 5-fold cross-validation. Hyperparameters for RFR were tuned based on grid search and performance optimization.

Results

Model	Encoding	R² Score	RMSE (€)
MLR	One-hot	0.741	4,214.52
MLR	Target	0.699	4,546.95
RFR	Raw	0.867	3,042.17
RFR	Target	0.892	2,734.41

🔍 Key Observations

• MLR works better with one-hot encoding, as it handles linear relationships more effectively.
• RFR performs significantly better with target encoding, capturing nonlinear patterns in the data.
• RFR with target encoding emerged as the best-performing model, outperforming all other configurations.

Contributions

• Investigated encoding strategies (one-hot vs. target) on categorical variables
• Conducted comparative analysis between linear and nonlinear regressors
• Validated results using standard evaluation metrics (R², RMSE)
• Highlighted implications of encoding choices on model performance

Paper

This project was conducted as part of a university research initiative and has been summarized in the following paper:

Development of a Machine Learning-Based Prediction Model for Used Car Listing Prices in Germany Kim et al., Sookmyung Women’s University, 2025

📄 Read Full Paper(PDF)

Technologies

• Python 3.9, KNIME Analytics 5.2.5
• pandas, numpy, scikit-learn
• Matplotlib, Seaborn (for visualization)
• Jupyter Notebook

Future Work

• Extend price prediction to include classification of pricing tiers (low, medium, high)
• Incorporate external economic indicators for more robust modeling
• Deploy an interactive web-based prediction tool

Contact

For any inquiries or collaboration requests, feel free to contact:

• Jisoo Kim – jisugim168[a.t.]gmail.com