See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/327415208

Valuation of Used Vehicles: A Computational Intelligence Approach

Conference Paper · May 2018

DOI: 10.1109/ISMS.2018.00011

CITATIONS READS

0 3,400

2 authors, including:

Kaneeka Vidanage

28 PUBLICATIONS   6 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Kaneeka Vidanage on 10 September 2018.

The user has requested enhancement of the downloaded file.

 Valuation of Used Vehicles: A Computational Intelligence Approach
Amjadh Ifthikar
Department of Software Engineering
Informatics Institute of Technology
Colombo, Sri Lanka
amjadh.2014262@iit.ac.lk
Abstract ​— In most of the countries in the world, the
number of vehicles have increased exponentially in the past
decade. As a result, the used vehicle trade has flourished. In
order to determine the resale value of a vehicle or for
financial requirements such as vehicle leasing services, used
vehicle valuation is paramount. In order to meet the future
demand, a proper mechanism using computational
intelligence is needed for vehicle valuation in order to
overcome the inefficiencies of conventional and subjective
valuation mechanisms.This paper focuses on discussing
aspects regarding used vehicle valuation, previous work
related to the domain as well as the proposed solution. The
proposed solution intends to use real time web scraping and
machine learning in order to determine the value of a given
vehicle.
Keywords ​— ​vehicle; valuation; computational intelligence;
web scraping; machine learning
I. INTRODUCTION
Valuation, in general, is the mechanism of finding the
present market value of a future income flow. In
automobile valuation, the value presented in a valuation
report should represent the correct resale value of a
vehicle.
When people seek the assistance of a financial
institution in order to find the necessary financing for a
vehicle, they focus on obtaining the real value of the
vehicle. In order to get it, a valuation report is requested.
Hence, a proper mechanism to obtain a value for a vehicle
is paramount in such a scenario.
Due to the competition in the vehicle leasing industry,
it is vital for a leasing company to give a reasonable as
well as profitable price to remain a profitable company. A
manual mechanism is used in Germany when calculating
the second-hand vehicles and mentions that the system in
place consists of major limitations in the attributes
considered and also there is no proper mechanism to
capture the attributes correctly [1].
Valuation of vehicles is important to leasing
companies in order to avoid losses as well as for people
who have bought new vehicles and are willing to sell the
vehicle at an appropriate price [2].
Valuation is an appraiser's opinion - judgment from the
gathered information - it depends on the valuer's skills,
knowledge, and experience; furthermore, proper data
Kaneeka Vidanage
Department of Software Engineering
Informatics Institute of Technology
Colombo, Sri Lanka
kaneeka.v@iit.ac.lk
analysis methods are very significant to make a decision
in an appraisal of a property [3].
With the number of vehicles increasing exponentially,
the current process of manual valuation will become a
difficult, time consuming and ineffective process.
Therefore there is a clear research gap where a
technological solution can be provided to appraise
vehicles faster and much more accurately.
Hence, this paper proposes a computational
intelligence approach to automate the process of vehicle
valuation by dynamically scraping data from second-hand
vehicle selling websites and using machine learning to
determine the value of a vehicle. Using the system
proposed, the vehicle valuation process will be able to
become more scientific and objective.
Once the proposed system is developed, apart from
formal vehicle valuation processes, a tool of this nature
can also support various other use cases as well. They are
listed below.
● For accounting purposes in situations such as
creating final accounts at the end of the financial
year, the residual value of the assets is calculated.
● In order to decide insured amount in a vehicle
insurance so that when an accident occurs, you
get a correct value.
● To decide the amount received by the insurance
holder in case of a condemned accident.
● To assist vehicle sellers to decide an appropriate
selling price.
● To help used vehicle dealers when buying and
selling.
● To help someone who is going to buy a vehicle
to decide whether the price he/she is paying is
correct.
In this paper, an analysis of the theoretical background
and related work regarding the domain of vehicle
valuation have been conducted. Subsequently, the
proposed solution is explained in detail.
II. THEORETICAL BACKGROUND
Valuation is a process of finding the present value of a
future income stream of a return in a property based on
past and current information. Valuation, in its simplest
form, is the determination of amount for which the
property will transact on a particular date.
 Further as per the Royal Institute of Chartered
Surveyors, valuation is a hypothetical transaction. The
very act of asking for a valuation is equivalent to putting
the property on the market. In asking for a valuation, the
owner has therefore made the decision to "sell" he is
willing to sell even though he knows the market is weak /
bad. He cannot become unwilling just because he is not
happy with the price (valuation figure). All valuations
should be approached only through the eyes of purchases
[3].
Further definitions of R.T.C.S. Manual, it can be
concluded that a valuation is a prediction of the most
likely selling price. A calculation of worth is an estimate
of what an investment is worth to a particular buyer or
seller, and an appraisal is a combination of the above two.
"The act or process of estimating value corresponds to the
term "valuation appraisal" defined as "an unbiased
analysis, opinion, or conclusion that estimates the value of
an identified parcel of real estate or real property at a
particular point in time."[4]
As it has been indicated in the above definitions,
valuation is a process of determining the current worth of
an assets or a property. Unbiased valuation for the
property is very important.
The quantity and usage of vehicles have increased and
the complexities of it have increased simultaneously. The
necessity of getting the realistic values for these vehicles
increased day by day for the various purposes such as
insurance, selling purposes, purchasing, accounting
purposes, etc.
III. RELATED WORK
The automobile industry is globally expanding
exponentially. As a result, the used vehicle market has
also faced an unexpected growth. in order to cope with the
future expansion of vehicle trade, research in
computational approaches to used vehicle valuation have
been scarce. Given below is an analysis of previous
approaches to used vehicle valuation and value prediction.
Listiani, in her MSc thesis [1], has compared
approaches of multiple linear regression and Support
Vector Machines (SVM) in estimating the residual price
of leased cars. In her thesis, she proves that a high
accuracy can be gained by using SVM than simple
multiple regression or multivariate regression. SVM is
predicting value of used vehicles by utilizing machine
learning systems which are more ready to manage high
dimensional information (number of features used to
anticipate the cost) and can keep away from both
overfitting and underfitting. In order to find the optimal
parameters for the SVM, a genetic algorithm has been
used. This reduces the time taken to generate the SVM.
The main downside of this approach is that the superiority
of SVM rather than simple multiple linear regression was
not depicted in basic measures like mean variance or
mean deviation.
Richardson, in his thesis [5], suggests a theory that
vehicle manufacturers are all the more eager to make
vehicles which don't depreciate quickly. By the use of a
multiple linear regression analysis he proves that hybrid
vehicles, which are vehicles with the ability of using both
a combustion engine and an electric motor, have a higher
ability to retain its market value in a used vehicle market
than vehicles with only a combustion engine. In
determining these results, he has used features like
milage, make, age and fuel efficiency of the vehicles. The
reason pointed out was environmental as well as
economical reasons because hybrid vehicles emits a very
low amount of carbon dioxide and also consume less fuel.
The data have been collected from various websites.
Wu et al. [6] used an Adaptive Neuro Fuzzy Inference
System (ANFIS) and also a parallel Artificial Neural
Network (ANN) to predict the value of used vehicles and
analyzed the results of each approach. The data for the
system have been collected from a used car website in
Taiwan. In order to predict the price, the brand of the
vehicle, the year of manufacture of the vehicle and the
engine capacity of the vehicle have been used. In order to
increase the accuracy, the availability of Anti-lock
Braking System (ABS), Traction Control System (TCR)
and Supplemental Restraint System (SRS) have been
used. The study concludes that the ANN with
Backpropagation (BP) gives a higher accuracy(lower
absolute percentage error) than ANFIS and that ANFIS
provides a better forecasting performance than ANN. The
reason provided for this conclusion is that ANN with BP
uses many neurons, which takes time to train but can
easily learn the relationship between inputs and outputs.
In USA, a large amount of vehicles are sold through
leasing [7]. Because most vehicles are been returned at
the end of the leasing period, institutions needs to appraise
the vehicles accurately in order to resell the vehicles
which are returned. In order to address this situation, the
ODAV (Optimal Distribution of Auction Vehicles)
system was developed by Du et al. [7]. Using a k-nearest
neighbor regression model, the system have the ability of
estimating the price of vehicles entered. This solution was
created in 2003. Since the United States is a huge country,
the state in which the vehicle is present also plays a
paramount role in the vehicle price.
An approach using Artificial Neural Networks (ANN)
for value prediction of used vehicles was proposed by
Shen Gongqi [8]. The features mileage, estimated useful
life, and manufacturer was considered. This model was
created in order to handle nonlinear relationships as well,
which is ignored in simple linear regression models. In the
end it was concluded that a good accuracy was gained
using this approach to predict the value of used vehicles.
Generally considering previous research in this area,
one of the main issues is the unavailability of up-to-date
data in order to predict the price of a given vehicle. In a
market of this nature, the volatility of the prices are very
high. The price of a vehicle today can drastically change
tomorrow.
The main difference between previous work and the
proposed solution is that the proposed solution will collect
data in real time from second-hand vehicle selling Toyot 121
2003 128000 Auto Petrol 3590000
websites and as a result help to take rapid market a G-grade
Toyot 121
fluctuations into consideration. Another difference is that a G-grade
2000 140000 Auto Petrol 2950000
the proposed solution gather and use data from vehicles Toyot 121
2004 106000 Auto Petrol 3750000
similar to the vehicle to be appraised. This significantly a G-grade
improves the accuracy of the valuation. Toyot 121
2001 228000 Manual Petrol 2900000
a G-grade
IV. PROPOSED SOLUTION
The proposed solution, which will be created to When analyzing the data collected it is clearly evident
become a tool for automobile appraisal agents will consist that the system have collected information on vehicles
of a web front end for the user to input characteristics of with similar brand and model of that of the sample
the vehicle to be valued. The data will be cross validated vehicle.
using an API(Application Programming Interface) that After the data collection part is completed, the data
contains details of vehicle makes and models existing in collected from several websites needs to be integrated. In
the world. order to do this a rule based data integration mechanism is
The details of the vehicle will then be sent to a web to be introduced which will eliminate duplicate vehicle
crawler which will be programmed in a way such that it entries and also make sure all data will be in the same
will collect adequate data of vehicles similar to the vehicle format.
to be valued. The data that will be collected are the brand, The integrated data will be pre-processed later. The
model, engine capacity, fuel type, transmission, mileage main processing that will be done is the filtration of data.
and year of manufacture. These factors were determined This is conducted in order to determine whether all the
by studying previous research conducted in the area and data available is similar to the vehicle that is need to be
are suggested to be the factors which govern the value of a valued and eventually eliminate data that is not similar.
used vehicle.. Once the data pre-processed, a machine learning
For this, the web crawler will scrape data from several model will be trained dynamically. In order to achieve this
websites which let the general public advertise an analysis will be conducted to determine which machine
second-hand vehicles by payment of a fee. Tough there learning algorithm/ combination of algorithms will ensure
are websites which let people advertise free of charge, the highest accuracy. The machine learning algorithms
they were not considered because people can quote used for this purpose are, Gradient Boosting Regression
unrealistic prices for their vehicles. (GBR), SVM and Naive Bayes Regression (NBR).
In order to get similar vehicles to the vehicle that In order to evaluate the accuracy of each algorithm/
needs to be valued, initially the web crawler will be the combination of algorithms, the data which is collected
using the results filtering mechanism of the websites that dynamically will be divided into two where 70% of the
will be crawled. data will be training data and 30% of it will be testing
Depicted in TABLE II is a part of the data set created data. Once the model is trained using the training data, it
by scraping data of vehicles similar to a sample vehicle will predict the values of the vehicles in the testing data.
which is shown in TABLE I. Subsequently, the Mean Absolute Error (MAE) will be
calculated as follows.
TABLE I. SAMPLE VEHICLE

Feature Value MAE = n​−1​∑​x∈X​ |y​x −

​ z​x​| (1)
Brand Toyota
Where X is the testing dataset, y is the predicted value,
Model 121 G-grade z is the real value and n is the number of data used for
Year 2001 testing. MAE will be used as the measuring factor of the
Mileage 100,000km model accuracy.
Initially, a dynamic model was trained using the GBR
Transmission Automatic
algorithm. The configurations used to train the machine
Fuel type Petrol learning model that resulted in the least MAE are given in
TABLE II. SAMPLE SCRAPED DATA TABLE III.
The mean absolute error of the trained model for the
Yea Mileag Trans Fuel Price vehicle data collected for the vehicle described in TABLE
Brand Model
r e (km) mission type (LKR)
Toyot 121
I was LKR 150,000. Considering the mean vehicle value
2002 148000 Auto Petrol 3375000 of the vehicles, the mean absolute error is less than 5%.
a G-grade
Toyot 121 To evaluate the entire system as a whole, a set of
2003 146000 Auto Petrol 3600000
a G-grade vehicles which are in the market will be valued using the
Toyot 121 system. The error percentage will be calculated
2001 150000 Manual Petrol 2850000
a G-grade
eventually.
View publication stats

A rich picture depicting a high level flow of the As future enhancements to the system, since it is
proposed solution is depicted in figure 1. developed as a helping tool for vehicle valuators, an
mobile application can be developed to be used by
TABLE III. TRAINING CONFIGURATIONS vehicle valuators. This application can be given the
Parameter Value ability to connect to a OBDII device plugged to the
OBD(On Board Diagnosis) port of the vehicle in order to
Learning rate 0.05
get vital information on the condition of the engine
Max depth 6 which cannot be seen by external inspections.
Minimum samples leaf 5 The system can be further enhanced to facilitate
Maximum features 0.1 predictions of vehicle values. It can be achieved by
Number of estimators 1000 training a neural network to learn from data of vehicles
from different years of manufacture. Vehicle value
prediction is mainly important for vehicle leasing
companies to determine what the value of the vehicle
will be after a number of years. It is also an important
factor when making a vehicle purchasing decision.
REFERENCES
[1] M. Listiani, “Support Vector Regression Analysis for Price
Prediction in a Car Leasing Application,” 2009.
[2] S. Pudaruth, “Predicting the Price of Used Cars using Machine
Learning Techniques,” ​Int. J. Inf. Comput. Technol.,​ vol. 4, no. 7,
pp. 753–764, 2014.
[3] P. P. D. . Shayamali, “A Study on the Current System of Vehicle
Valuation in Sri Lanka,” 2013.
[4] Utah State Tax Commission, ​Real Property Valuation Standards
of Practice​. 2002.
[5] M. S. Richardson, “DETERMINANTS OF USED CAR RESALE
VALUE,” 2009.
[6] J. Da Wu, C. C. Hsu, and H. C. Chen, “An expert system of price
forecasting for used cars using adaptive neuro-fuzzy inference,”
Expert Syst. Appl.​, vol. 36, no. 4, pp. 7809–7817, 2009.
[7] J. Du, L. Xie, and S. Schroeder, “PIN optimal distribution of
auction vehicles system: Applying price forecasting, elasticity
estimation, and genetic algorithms to used-vehicle distribution,”
Market. Sci., vol. 28, no. 4, pp. 637–644, 2009.
[8] S. Gongqi, Y. Wang, and Q. Zhu, “A new model for residual
value prediction of the used car based on BP neural network and
nonlinear curve fit,” ​Proc. - 3rd Int. Conf. Meas. Technol.
Mechatronics Autom. ICMTMA 2011​, vol. 2, pp. 682–685, 2011

Figure 1. Rich picture

V. CONCLUSION
In this study, the current system of vehicle valuation,
the theoretical background and previous work carried out
on similar scenarios have been analyzed. The study also
proposes a solution to overcome the irregularities of the
current system.
Based on the initial results of the study as mentioned,
it is evident that the proposed solution is capable of
determining the market value of a used vehicle with a
high accuracy.
VI. FUTURE WORK
The proposed solution, which is at a concept stage
needs to be tested with different algorithms/combination
of algorithms to achieve the least error.