Pre-Owned Car Price Prediction Using Machine Learning
Techniques
Andy Zhu
Computer Science, Rensselaer Polytechnic Institute, Troy, New York, U.S.A.
Keywords: Pre-Owned Car Price Prediction, Machine Learning, Linear Regression, Decision Trees, Random Forest.
Abstract: As the car industry continues to evolve, car price prediction models have been a highly focused topic for
research. In the current pre-owned car marketplace, buyers are impelled to decide whether a vehicle is within
a reasonable price range. From visiting car dealers to websites, the task of finding the true worth of a car is
laborious, especially for those who have limited knowledge about cars. The rise of online car listing platforms
calls for buyers and sellers to be informed about the values of vehicles in the market. A tool that could
determine the honest value of a vehicle would be crucial to the automotive marketplace. This paper proposes
several machine learning algorithms that aim to predict the prices of used vehicles based on the features of a
car. By using linear regression, decision trees, and neural network, this study explores models that best capture
the price of used cars from pre-existing car sales data. The models’ performances are assessed to determine
the most suitable model for future price prediction applications.
1 INTRODUCTION
In today’s growing automotive market, pre-owned
vehicles make up the majority of auto sales in the
United States. In 2021, pre-owned auto sales consisted
of nearly 78% of all vehicle sales in the U.S. (Carlier
and Mathilde 2023. Through the advent of online car
listing platforms, people looking to purchase second-
hand cars can conveniently search car listings around
their areas on their personal computers. While these
platforms facilitate a search experience to help
consumers find cars within affordability, there is no
unified algorithm shared by these platforms for
determining the price of a car. Competition exist
between selling platforms, and each platform has their
own measurement of how much a car is worth.
The demand for transparency in the U.S.’ ever-
rising pre-owned car industry is an issue worth
emphasizing. Consumers deserve to know the
worthiness of a car without the convolution of online
car selling platforms advertisements when, at the end
of day, the goal of the platform is to make the
consumer make a purchase. In this paper, different
machine learning techniques are applied to discover
trends and patterns on pre-existing car sales data,
exploring the models which predict the true cost of a
vehicle. Several machine learning models are
implemented to best determine the retail price of a
used car based on features such as make, model, year,
and mileage.
2 RELATED WORKS
Price prediction models for used vehicles have been
at the core of other studies that utilize various
techniques of machine learning.
Venkatasubbu and Ganesh used Lasso regression,
multiple regression, and regression tree to develop
their models. Samruddhi proposed supervised
machine learning through K Nearest Neighbor
algorithm to determine the used car prices. Cross-
validation was ultimately used to evaluate the model.
Gegic et al made prediction models through an
artificial neural network, support vector machine, and
random forest in ensemble fashion. Heavy
classification was done to categorize prices into
broader classes (cheap, moderate, expensive) for
separate model training thereby minimize price
fluctuations. Sharma and Sharma utilized linear
regression with one-hot encoding and attained a
relatively high accuracy of 0.86. Pandit et al
compared the methods of Lasso Regression, Ridge
Regression, Bayesian Ridge Regression, and the
output can be visualized through a web application.
356
Zhu, A.
Pre-Owned Car Price Prediction Using Machine Learning Techniques.
DOI: 10.5220/0012810100003885
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 1st International Conference on Data Analysis and Machine Learning (DAML 2023), pages 356-360
ISBN: 978-989-758-705-4
Proceedings Copyright © 2024 by SCITEPRESS – Science and Technology Publications, Lda.

3 METHODS AND MATERIALS
This study’s data comes from a dataset which contains
completed auto sales in the United States in recent
years (Sharma and Sharma 2020). Figure 1 shows the
overall process of approach and Table 1 lists the
details of the cars from the data.
Figure 1: Process Diagram (Picture credit: Original).
A car has multiple features from the dataset.
Price: Price of the car
Levy: Tax or levy imposed on the car
Manufacturer: Brand or maker of the car
Model: Specific version or variant of car
Production Year: Year which car was produced
Category: Body style of car (Jeep refers to SUV)
Leather Interior: Leather seats and interior
upholstery
Fuel Type: Gasoline/Hybrid/Diesel
Engine Volume: Capacity of car’s engine
Mileage: Total number of miles driven
Cylinders: Number of combustion chambers
Gear Box Type: Transmission system type
Wheels: Drivetrain Configuration (4x4, FWD,
RWD)
Color: Airbags: Number of airbags
When assessing a car’s value, all its attributes
come into play. Variables such as price, year, mileage,
etc. are numerical features which can be readily
applied in the model building. The remaining features
fall into categorical and require conversions. One-hot
encoding is used to transform the categorical data into
a format that is machine-readable. Table 2 shows the
structure of the data after one-hot encoding is applied.
Table 3 contains the training and testing set sizes.
Table 1: Data overview of preprocessed data.
Table 2: Sample data after preprocessing.
Price
Year
Engine
Volume
Mileage
(km)
Fuel
Type
Category
Gear
Box
Type
Wheels
Color
0
13328
2010
3.5
186005
1
1
1
0
1
1
16621
2011
3
192000
0
1
1
0
0
2
3607
2011
2.5
168966
1
0
0
1
0
3
11726
2014
1.3
91901
0
1
0
1
1
4
39493
2016
2
160931
1
0
1
0
1
Pric
e
Lev
y
Manu.
Mod
el
Pro
d.
Yea
r
Catego
ry
Leath
er
Interi
or
Fuel
Type
Engi
ne
Volu
me
Milea
ge
(km)
Cylind
ers
Ge
ar
Bo
x
Ty
pe
Whe
els
Col
or
Airba
gs
133
28
139
9
LEXUS
RX
450
201
0
Jeep
Yes
Hybr
id
3.5
18600
5
6.0
Aut
o
4x4
Silv
er
12
166
21
101
8
CHEVY
Exqu
i.
201
1
Jeep
No
Petro
l
3
19200
0
6.0
Tipt
.
4x4
Blac
k
8
360
7
862
FORD
Esca
pe
201
1
Hatchb
ack
Yes
Hybr
id
2.5
16896
6
4.0
Aut
o
4x4
Whi
te
8
117
26
446
HOND
A
FIT
201
4
Hatchb
ack
Yes
Hybr
id
1.3
91901
4.0
Aut
o
Front
Silv
er
4
394
93
891
HYUN
DAI
Sant
e Fe
201
6
Jeep
Yes
Dies
el
2
16093
1
4.0
Aut
o
Front
Whi
te
4
Pre-Owned Car Price Prediction Using Machine Learning Techniques
357

Table 3: Partitioning of data into training sets.
Ratio of dataset
Number of entries
Training Set
0.8
15389
Test Set
0.2
3847
The following are the techniques used in this
study:
Linear Regression: Linear regression is a
supervised learning technique that predicts a
continuous variable based on a number of predictor
variables. It offers simplicity in implementation and
computation. Linear Regression can be applied to the
entire dataset without additional parameter tuning.
However, the simplicity means it does not capture
more complex relationships between data as effective
as other models.
Random Forest: Random Forest is an ensemble
learning algorithm that combines the decisions from
multiple decision trees to improve predictive
performance. Each tree is built on a random subset of
data, ensuring diversity to reduce variance of the
model. The model captures both linear and complex
nonlinear relationships, make it a great fit for
predicting car sales prices based on various input
features.
XGBoost: Extreme gradient boosting
(XGBoost) is a powerful algorithm that excels in
speed and predictive accuracy. Unlike Random Forest
which builds trees in parallel, XGBoost builds trees
sequentially, where new trees correct errors of
previous ones. This model handles bad and missing
data automatically, reducing the need for data
imputation. Overall, gradient boosting is a robust
method well-suited for predictive tasks.
Neural Network: Neural networks function like
human brains. They consist of interconnected nodes
(neurons) that process data in layers. They employ
multiple hidden layers between input and output
layers, allowing the model to learn precise patters
from data. A neural network is also more acquainted
to unstructured data, leveraging heavy computational
power for its performance. While neural networks can
capture complex relationships, they also require more
time to tune and refine. This study’s implementation
is inspired by Tamoghno’s neural network models
(Tamoghno 2020).
4 MODEL IMPLEMENTATION
The models are based on supervised learning and
make predictions about car prices using information
from the training sets and subsequently are applied to
the test dataset for evaluation. Figure 2 shows the
results from the linear regression model.
5 LINEAR REGRESSION
The linear regression model is based on Animesh’s
implementation and ideas are derived from Ali’s work.
Figure 2: Predicted(y) vs Actual Value(x) of selling price
(Linear Regression) (Picture credit: Original).
6 RANDOM FOREST
REGRESSOR
Hyperparameter tuning is done using GridSearchCV
to optimize the performance of the model. The
hyperparameters used are in Table 4, and Figure 3
shows the results of the Random Forest Regression.
Table 4: Optimal hyperparameters for random forest.
n_estimators (number of trees in the forest)
30
max_depth (maximum depth of the trees)
2
min_samples_split (minimum number of
samples required to split an internal node)
2
min_samples_leaf (minimum number of samples
required at a leaf node)
100
Figure 3: Predicted(y) vs Actual Value(x) of selling price
(Random Forest) (Picture credit: Original).
DAML 2023 - International Conference on Data Analysis and Machine Learning
358

7 XGBOOST
Gradient boosting, inspired by Mello’s “XGBoost:
Theory and practice”, is applied to the dataset. Table
V outlines the hyperparameters used for tuning the
XGBoost model, and the results are shown in Figure
4.
Table 5: Parameters for XGBoost.
learning_rate (step size for each
iteration in gradient boosting)
0.01, 0.05, 0.1
max_depth (maximum depth of
individual decision tree)
4, 6, 8
colsample_bytree (fraction of
features sampled for tree growth)
0.6, 0.7, 0.8
alpha (L1 regularization term for
controlling complexity)
0, 0.5, 1
n_estimators (number of
decision trees to include in the
ensemble)
500
objective (function specifying
the task)
reg:squarederror
Figure 4: Predicted(y) vs Actual Value(x) of selling price
(XGBoost) (Picture credit: Original).
8 NEURAL NETWORK
Table 6 shows the performance of the neural network
model across various structures having different
number of layers.
Table 6: Neural network evaluation on layer.
Summary
Layers
2
3
4
R
2
-Value
0.66
0.64
0.65
MAPE
31.54%
27.80%
29.98%
Hidden layers are intermediate processing stages
for the neural network to recognize and learn patterns,
features, and relationships from the data. The results
reveal that the neural network with three hidden layers
has the lowest MAPE score, indicating the highest
level of accuracy in predictions. The network with
four hidden layers has a slightly larger MAPE score
but performs well in terms of accuracy. The network
with two layers has the largest MAPE score,
indicating the lowest accuracy.
9 MODEL PERFORMANCE
EVALUATION
Table 7: Overall Model Evaluation.
Model
R
2
-Value
MAPE Score
Linear Regression
0.576
38.95%
Random Forest
0.765
25.34%
XGBoost
0.789
24.48%
Neural Network
0.65
28.98%
Table 7 lists the models’ performances. Linear
regression provided a baseline for this study’s
prediction assessment. The R
2
suggested the model
provided a moderate level of predictive capability and
in explaining the variance of the dataset. However, the
MAPE score remained relatively high, indicating a
significant deviation from the actual car prices in the
training set. The Random Forest model showed a
considerable improvement over linear regression and
captured larger portions of variance. The decreased
MAPE score implies better accuracy in predictions.
The gradient boosting algorithm further improved the
model’s predictions, achieving the highest R
2
among
the models implemented. The enhanced accuracy can
be attributed to XGBoost’s ability to capture complex
non-linear relationships in the dataset. The neural
network model did not perform as well as Random
Forest or XGBoost, which suggests the model needs
further optimization.
While linear regression provided a reasonable
initial insight, it did not capture the intricate
relationships influencing car prices. XGBoost
balances accuracy and computational efficiency.
From the training results, XGBoost appears as the
preferred choice in car price prediction tasks.
The models evaluated in the study represented a
progression from basic linear regression models to
advanced learning techniques such as a neural
network. The decision for the best model depends on
the trade-offs between a model’s accuracy,
complexity, and efficiency.
Pre-Owned Car Price Prediction Using Machine Learning Techniques
359

10 CONCLUSION
In today’s thriving automotive market where pre-
owned car sales account for the majority of vehicles
sold every year in the U.S, it is imperative that
consumers are fed transparent, data-driven
information about car prices. Price control and profit
maximization are often at the best interests of car
sellers. They list prices of cars according to their
preferences, often overlooking the financial situations
of buyers. While this practice has been a tradition
among car retailers, it is important to ensure that
buyers are well-informed about the true costs of a
vehicle. Machine learning algorithms are able to
achieve this task and predict the costs of used cars
based on historical data. This allows buyers to make
informed decisions on purchasing, and from a moral
standpoint, promotes fairness in the car market
industry. A transparent and universal car price
evaluation algorithm would foster healthy competition
among car retailers and listing platforms. This study
tackles this issue by testing machine learning
algorithms that predict the true cost of used vehicles.
This analysis begins with exploring the linear
regression model and delves into more sophisticated
techniques involving random forest decision trees.
Advanced techniques such as gradient boosting and
neural network are implemented to learn intricate
patterns more precisely from feature variables in the
dataset. This study identifies the best machine
learning models to use for evaluating used car prices.
The objective of this research is to provide a method
for price prediction that ensures fairness between
consumers and sellers.
This study delves into machine learning models
for predicting used car prices. The models proposed in
this study can be adopted in similar price prediction
studies such as determining the price of real estate
properties, vintage collectibles, or electronic gadgets
in the pre-owned market. These models can undergo
heavy tuning on a larger dataset to provide more
accurate predictions. Clustering algorithms like K-
Means can also be considered, involving attribute-
based clusters. This opens the option for price range
analysis that assesses groups of cars across different
price segments. For future applications, this study will
explore larger datasets containing a broader range of
sales data across any industry.
REFERENCES
Carlier, Mathilde, “U.S. New and Used Car Sales”, (Statista
29 Aug. 2023)
P. Venkatasubbu and M. Ganesh, “Used Cars Price
Prediction using Supervised Learning Techniques”,
IJEAT, 2019, pp. 216-223.
K. Samruddhi and Dr. R. Ashok Kumar, “Used Car Price
Prediction using K-nearest Neighbor Based Model”,
IJIRASE, 2020, pp. 686-689.
Enis Gegic, Becir Isakovic, Dino Keco, Zerina Masetic,
Jasmin Kevric; “Car Price Prediction using Machine
Learning Techniques”, TEM, 2019, pp. 113-118.
Ashutosh Datt Sharma, Vibhor Sharma, “Used Car Price
Prediction Using Linear Regression Model”, IRJET,
2020, pp. 946-953.
Eesha Pandit, Hitanshu Parekh, Pritam Pashte, Aakash
Natani, “Prediction of Used Car Prices using Machine
Learning Techniques”, IRJET, 2022, pp. 355-360.
Bhattacharya Tamoghno, “An introduction to neural
networks with implementation from scratch using
Python”, Towardsdatascience July 1, 2020).
Agarwal Animesh, “Linear regression using python”,
(Towardsdatascience November 14, 2018).
Arthur Mello, “XGBoost: Theory and practice”, (Medium
August 17, 2020).
Amir Ali, “Linear regression with practical
implementation”, (Medium November 24, 2019)
DAML 2023 - International Conference on Data Analysis and Machine Learning
360