Predicting car prices with machine learning
How to use data science to buy a car
The process of buying a pre-owned car can take a lot of time. You can find thousands of cars online and offline and you can easily spend hours or days comparing them as the pricing of pre-owned cars is not very transparent. Therefore, I decided to search for a car in the nerdiest way possible: by scraping a car website and creating a model to predict the market value of pre-owned cars.
The data
Since I want to have a dataset to both train the machine learning model and to select potential cars to buy, I decided to scrape a Dutch car website myself instead of using an old Kaggle dataset. The scraping can be done with Selenium or BeautifulSoup and this resulted in a dataset of c. 45k pre-owned cars with prices between €5k and €25k, excluding diesel cars.
After cleansing the data (and checking the distribution), the dataset includes many relevant features that can be used to predict the market value of each car in the dataset.
Checking the distribution of cars by manufacturer shows that the number 1 manufacturer in my dataset is Volkswagen, followed by Opel and Renault and the majority of cars in my dataset are hatchbacks followed by SUVs.
plt.figure(figsize = (20,15))
cars['Brand'].value_counts().plot(kind='bar')
plt.figure(figsize = (20,15))
cars['Type'].value_counts().plot(kind='bar')
To visualize the dataset and to better understand the relationships between the variables, we can create a correlation heatmap. This heatmap shows high positive correlations between the variables year and price, and between the new price of the car and the actual price. A high negative correlation can be observed between the variables year and km (odometer). These correlations make sense, so this is a good starting point for the data modeling part.
# calculate correlation matrix
corr = cars.corr()
# plot the heatmap
plt.figure(figsize =(20,15))
ax=sns.heatmap(corr, xticklabels=corr.columns, yticklabels=corr.columns, annot=True, cmap=sns.diverging_palette(220, 20, as_cmap=True))
bottom, top = ax.get_ylim()
ax.set_ylim(bottom + 0.5, top - 0.5)
Data modeling
To predict the market values of the cars in my dataset, I use a random forest model. Before I can start modeling I first have to transform the categorical variables into dummy variables to transform every unique value of a variable into its own binary variable.
cars = pd.get_dummies(cars, drop_first=True)Secondly, I need to scale the independent variables. The scaling of the independent variables is important because the max value of the year variable is 2021, whereas the max number for km (odometer) is 385,000 km. Therefore, if you wouldn’t scale the data a small change in the km variable would have a much larger impact on the model compared to the same change in the year variable.
from sklearn.preprocessing import StandardScaler
X_head = cars.iloc[:, cars.columns != 'Price']
X = cars.loc[:, cars.columns != 'Price']
y = cars['Price']
X = StandardScaler().fit_transform(X)Now we can train and test the random forest model. The random forest basically builds multiple decision trees and merges them together (it operates as an ensemble) to get a more accurate and stable prediction. I use 80% of the dataset to train the model and 20% for testing it.
from sklearn.ensemble import RandomForestRegressor
from sklearn.model_selection import train_test_split
from sklearn.metrics import mean_absolute_error as mae
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=.20, random_state=0)
model = RandomForestRegressor(random_state=1, max_depth=50)
model.fit(X_train, y_train)
pred = model.predict(X_test)
print(mae(y_test, pred))
print(cars['Price'].mean())
model.score(X_test,y_test)The model resulted in an MAE (mean absolute error) of 896 for a mean price of €13,498 and achieved an accuracy score of 94.3%!
We can visualize one single decision tree of the random forest model with the following code:
from sklearn.tree import export_graphviz
from sklearn import tree
from sklearn.datasets import load_boston
from dtreeviz.trees import dtreeviz # will be used for tree visualization
from matplotlib import pyplot as plt
plt.rcParams.update({'figure.figsize': (12.0, 8.0)})
plt.rcParams.update({'font.size': 14})
plt.figure(figsize=(20,20))treepicture = tree.plot_tree(model.estimators_[50], feature_names=cars.columns, filled=True)
It is impossible to read the entire decision tree, so we can limit the depth of the decision tree to check a single decision tree with a limited number of variables. To classify a new point, simply move down the tree, using the features of the point to answer the questions until you arrive at a leaf node where the class is the prediction. But keep in mind that all variables are scaled, so all the feature values in the tree are also scaled.
Checking the feature importance shows that the year variable and new_price variable are the key features to calculate the market value of used cars, followed by the KM (odometer) variable.
Since the new_price of the car also indirectly accounts for other car features, I also train the model without this variable. This would still result in an accuracy score of 93.8%. The car features, such as horsepower (PK variable), have become more important in this model.
I used the trained model to create a dataset of thousands of cars including car features, the price, and the estimated market value based on the random forest model. This allowed me to filter out the best car deals and explore these cars in further detail. The model also allowed me to check the estimated market value of cars that I spotted on other websites and offline, and a couple of days ago I bought a car based on this small data science project!
Thank you for reading! If you have any questions or comments regarding this article, please feel free to comment below.
