hcan-m/usedcarpriceprediction

Price prediction on Large Dataset

Used Car Price Prediction

Used Car Price Prediction Analysis
Huseyin Can Minareci

Data available at:

https://www.kaggle.com/orgesleka/used-cars-database\

Used Car Price Prediction Analysis.pdf --- pdf print of Jupyter Notebook
Used Car Price Prediction Analysis.html --- html print of Jupyter Notebook
Notebook.ipynb --- all codes in Jupyter Notebook
presentation.pptx --- presentation
Warning: Running all codes in notebook takes around 2-3 hours.

Summary

In this project we aimed to find the best regression model for used cars dataset to be able to predict used cars price.

Dataset was scraped with Scrapy from German eBay and there were some mistakes, duplications and outliers. As first step we removed them from the dataset and stayed with 79.9% of the original data.

There were 20 variables in the original dataset and we had to remove "dateCrawled", "postalCode", "abtest", "offerType", "nrOfPictures", and "seller" because they were not usefull for our purpose. We created new features:

"daysBeforeSold" by using "dateCreated" and "lastSeen" "namelen" by using "name" "age" by using "yearOfRegistration" After that we worked on missing values columnwise and impute what we could impute meaningfully and drop the rest. At the end we had 75.8% of the original data ready to use for price prediction.

We used log-transformation in price and powerPs and one hot encoding for categorical variables to be able to perform machine learning algorithms.

We used Linear Model, K Nearest Neighbor and Random Forest. The most successul one was Random Forest with 88.01% r2_score. Followed by KNN with 83.70% and the worst one was as expected Linear Regression with 78.67%.