Second-Hand Car Prices Prediction using Machine Learning Algorithms

This project was developed for my Bachelor's thesis at the Faculty of Organizational Sciences, Information Systems and Technologies. The goal was to predict prices of second-hand cars listed on the Serbian platform PolovniAutomobili using machine learning algorithms in Python.

Project Overview

The objective of this project is to accurately predict prices of second-hand cars in euros using various ML models. The dataset consisted of 23,759 cars for training and 5,940 vehicles for testing. Coefficient of determination [R2] was the main metric used for models evaluation. Hyperparameters for Random Forest and Gradient Boosting Model were optimized using cross-validation. Furthermore, optimized GBMs were used to predict 90% prediction intervals.

GBM model performed best, achieving R2 of 0.94027 and MAE (mean absolute error) of 1230.770€ on test data. The prediction intervals contained the true values only 43.38% of the time. The 0.50 quantile GBM model identified the production year as the most important attribute, followed by engine power, kilometerage, model name and others.

Data Acquisition

Vehicles were scraped from the website PolovniAutomobili using BeautifulSoup and Tor. SQLAlchemy and MySQL were used to store the data, resulting in a total of 30,788 rows and 50 columns.

Technologies Used

• Programming Language: Python 3.11.8;
• Database: MySQL, SQLAlchemy;
• Web Scraping: BeautifulSoup, Tor, Stem, Selenium, tbselenium;
• Data Analysis: Pandas, NumPy;
• Visualization: Matplotlib, Seaborn;
• Machine Learning: scikit-learn;
• Configuration: PyYAML, Hydra.

Data Preprocessing

The Initial Cleaner component processed each column and removed outliers. The resulting data was passed then into a Preprocessor Pipeline, which prepared train and test set separately for modelling. Price (output column) was log-transformed to further improve model performance.

Results

Below are the tables of performance metrics for the optimized models on both the training and test sets.

Train Set Metrics

Model	RMSE	MAE	R2
RF	1191.538	499.685	0.98717
GBM 0.50	1579.571	547.138	0.97746
GBM 0.05	2470.072	974.428	0.94488
GBM 0.95	1435.852	818.770	0.98137

Test Set Metrics

Model	RMSE	MAE	R2
RF	2726.884	1311.138	0.93259
GBM 0.50	2566.822	1230.770	0.94027
GBM 0.05	3449.857	1715.0888	0.89211
GBM 0.95	2879.736	1624.346	0.92482

Additionaly, there is a comparison of R2 between the default and optimized ML models, along with a plot depicting true versus predicted values on the test set (GBM), including prediction intervals.

Comparing Default and Optimized Models on Test Set using R2 Metric (GBM)

Bachelor’s Thesis and Presentation

Both the thesis and presentation are available as PDFs in this repository. Files:

• Bachelor's Thesis (Serbian)
• Presentation (Serbian)

Acknowledgements

I really want to thank my mentors Sandro Radovanović and Andrija Petrović for their guidance and support throughout this project.