Predicting Used-Vehicle Resale Value in Developing Markets: Application of Machine Learning Models to the Kazakhstan Car Market

The challenge of identifying a pre-owned vehicle that fits within a specified budget is primarily due to the multiple factors that determine its price, including the model year, mileage, condition, and equipment. Consequently, customers often turn to experienced professionals for assistance in identifying an ideal vehicle appropriate for their budget. The significance of accurate price prediction and classification in the used vehicle market has been explored [1], underlining how accurate pricing is vital for both buyers and sellers in ensuring fair transactions and reducing information asymmetry. In tandem with this, effective classification aids buyers in making informed decisions tailored to their preferences and requirements [1].

In the study of Arora et al. [2], the efficacy of machine learning models in predicting car resale values was evaluated. A comprehensive dataset, encapsulating various car-related attributes such as make, model, year of manufacture, mileage, condition, location, among others, was collected and preprocessed. Machine learning algorithms, including linear regression, decision trees, random forests, and support vector regression, were then applied to train predictive models using this dataset [2].

In another study conducted in Morocco [3], a similar methodology was adopted, with a dataset consisting of various features related to used cars collected for the prediction of used car prices using machine learning techniques. Contrary to the traditional linear regression models, a novel S-curve model was proposed for the estimation of used car prices [4].

The complexity of rule-based algorithms often impedes a customer's ability to predict cost accurately, even when they have a clear understanding of what they intend to purchase. Inductive-based learning serves as an effective strategy in this regard, providing insights into cost derived from the dataset [5]. In a novel approach, a back propagation neural network model based on mean encoding was utilized for predicting used car prices, resulting in improved prediction accuracy [5].

Further exploration of machine learning regression models with ensemble techniques for predicting resale car prices was conducted [6]. A multi-layer perceptron (MLP), a type of feedforward neural network architecture typically used for regression tasks, was employed [7]. To minimize the difference between predicted and actual prices, the MLP was trained using backpropagation [7]. In a comparative study, the Light Gradient Boosting Machine (LightGBM) was found to outperform random forest in terms of prediction accuracy [8].

The suitability of machine learning approaches for this specific use case has been amply demonstrated. A comprehensive study on the application of machine learning techniques for predicting and classifying vehicle prices revealed the effectiveness of the proposed approach in accurately predicting prices for sedans, SUVs, and trucks [9].

Various machine learning algorithms and techniques were explored, highlighting their strengths and limitations [10]. It was also demonstrated that ensemble methods, such as random forests, often outperform individual algorithms in predicting vehicle prices [11]. After collecting and preprocessing a dataset of car features, experimenting with different regression models, applying feature selection techniques, and evaluating performance using appropriate metrics, the effectiveness of machine learning in predicting car prices was established, indicating potential benefits for the automotive industry [12].

In the present research, this problem is addressed, introducing a machine learning approach to predict the price of pre-owned cars in Kazakhstan. This paper aims to facilitate prospective buyers in making informed decisions when purchasing a used car by providing them with necessary information. A step-by-step model is offered, ensuring they secure the best vehicle possible.

The concept of regression, a cornerstone in statistical modelling, entails estimating the relationship between a dependent variable and multiple independent variables. Anil Kumar [13] delineated several prevalent techniques, including linear regression, decision trees, random forests, support vector machines, and neural networks. Each algorithm's strengths and limitations are expounded upon, accentuating their aptness for predicting car prices based on varying data types.

In the study of Amik et al. [14], a dataset encompassing various features of pre-owned cars in Bangladesh was compiled, including brand, model, year of manufacture, mileage, condition, and location. Subsequently, an array of machine learning algorithms, such as linear regression, decision trees, random forests, support vector machines (SVM), and artificial neural networks (ANN), were employed to train models for predicting car prices. Highlighting the influence of different inputs on the output, regression analysis emerges as a potent tool for forecasting and prediction, with the target variable representing the prediction goal and the independent features shaping its value.

Further analysis was conducted to investigate the various factors implicated in the pricing of used cars, including age, mileage, brand, model, condition, and market demand [15]. Correlation analysis and regression models were employed as statistical techniques to ascertain these variables. In the final step of price prediction, a machine learning model predicated on support vector regression (SVR) was developed to predict used car prices.

In a novel approach, Bayesian Additive Regression Trees (BART) were introduced to account for heteroscedasticity, a scenario wherein the variability of the response variable differs across different levels of the predictor variables [16]. A study focused on car sales prediction applied linear regression, decision tree regression, random forest regression, support vector regression (SVR), and artificial neural networks (ANN) [17].

The selection of regression methods is contingent upon the type of data and the desired outcome. The shape of the data, in tandem with the most compatible technique, influences the outcome of the model.

Machine Learning, an essential research branch, encompasses the process of enabling computers to learn and evolve through data and algorithms [18]. This technology is widely employed within the realm of Artificial Intelligence. Machine Learning algorithms facilitate the creation of models based on training data, enabling predictions and decisions without the necessity for explicit instructions.

Three main types of machine learning have been identified: supervised, unsupervised, and reinforcement learning. Supervised learning necessitates labeled data and primarily aims to link input features with output labels. Conversely, unsupervised learning leverages unlabeled data with the objective of unveiling hidden patterns within the data. Reinforcement learning, a process reliant on trial and error, interacts with the environment whereby positive feedback and rewards are granted when the agent progresses towards its goal, and negative feedback is utilized for improvement. Various use cases are associated with these learning types: Regression and Classification for Supervised learning, Clustering for Unsupervised learning, and game playing and Hidden Markov Models for Reinforcement Learning [19]. This research paper delves into the exploration of supervised machine learning technique, Regression.

Predictive modelling and machine learning are integral components of analytics, reviewing historical and current data to forecast future occurrences. Predictive models can discern trends within data, thereby facilitating the identification of potential risks or opportunities. The significance of Predictive Analytics is underscored across numerous industries, including retail, banking, healthcare, IT, entertainment, and social media, with its application expanding in sports and the Internet of Things (IoT) [9].

The application of AI technology spans various business facets such as customer targeting, sales forecasting, inventory control, market analysis, risk assessment, feedback gathering, and future planning. This contributes to enhanced decision-making with increased accuracy and efficiency. This research paper presents the development of a car price prediction model, proposed as a fair pricing system, with potential implications on the automotive retail sector. Prediction models can exert considerable influence on businesses by accurately forecasting the future and assisting in strategic planning.

The valuation of pre-owned vehicles in the market presents a challenge due to the vast fluctuation in prices, thereby complicating the process for both potential buyers and sellers in establishing a fair price [20]. Therefore, there is a pressing need for an intelligent system capable of accurately predicting prices and ensuring that the valuation of each vehicle is aligned with its classification. The effectiveness of such a car evaluation system is contingent upon the collection and availability of pertinent data. Such data includes the vehicle's year of production, motor type, condition, mileage, horsepower, number of doors, and weight, all of which are crucial to the comprehensive evaluation of a car [21].

However, the cost of a product is contingent on its offered features, which may not always be readily accessible, thereby complicating cost determination. This study, focused on the Kazakh market, sources its information from the reputable used car dealer, “kolesa.kz”. In order to enable the data mining model to process the raw data, it is required to undergo pre-processing and transformation into an acceptable format. Vital attributes contained within the raw data include the manufacturer, model, year of manufacture, kilometers traveled, and selling price. An initial statistical analysis of the data was undertaken to ascertain which parameters were most strongly correlated [22].

The extraction and selection of key features were facilitated by a correlation matrix, allowing for the identification and retention of the most relevant features, including vehicle price, kilometers driven, year of production, and the classification label for the car manufacturer.

The objective of this research is to scrutinize the efficacy of an array of regression methods, inclusive of Linear Regression, Decision Tree Regression, Random Forest Regression, Support Vector Machines, Neural Network, and Bagged Trees, in conjunction with machine learning algorithms. The aim of the study is to predict the resale price of used cars using pertinent features such as price, car, manufacturer, year, location, generation, body, engine volume, fuel type, mileage, transmission, and others.

Data was sourced from vendor websites and subjected to data mining methods to investigate the various factors contributing to the price estimation of a used car. The initial dataset comprises 178,397 data points with 16 attributes, including 101 car brands, 18 types of car body, and 304 car locations in settlements throughout Kazakhstan. This enables customers to accurately assess the value of their current car or the one they intend to purchase.

The intention of this investigation is to provide a robust analysis of the current market for used vehicles in Kazakhstan, with the objective of developing a predictive model that can serve as a reliable tool for customers and dealers alike. The model is anticipated to help in the accurate estimation of used car prices and thus aid in making informed decisions. The expectation is that this work will contribute to the literature by providing a comprehensive and detailed study of various factors affecting used car prices in Kazakhstan.

1.png

Figure 1. Pair plot between numerical features and the price

2.png

Figure 2. Features vs importance

Feature Importance is a key part of any machine learning process, as it has the ability to directly influence model performance. By removing attributes that are not relevant or only partially relevant, you can avoid overfitting of your algorithm and reduce the amount of time needed for training. Furthermore, Feature Importance can also help improve the overall performance of your model. Feature Importance is an effective technique that assigns a numerical score to each of the fields in your dataset. It helps you identify which variables are most influential and important when it comes to predicting the target outcome. The higher the score, the more relevant or essential that feature becomes. In this document, we will employ the feature importance attribute of a model and its corresponding correlation matrix to identify the most pertinent features of the dataset as seen in Figures 2 and 3. Estimating the price of a second-hand car is quite challenging. Not all features are equally important; some may make a huge difference, while others may not be relevant at all. Determining the right ones can prove to be crucial for acquiring accurate qualifications. It is essential to focus on the significant features for effective results. To see how those remaining features are connected, creating a heatmap (illustrated in Figure 3) will be beneficial. We can observe that the fields like transmission type of car, fuel type of car, brand name of car, kilometers driven value and age of car are having impact on predicting the price of used car. We can observe that the fields like age of car, engine volume, transmission type of car, body, drive unit are having strong positive impact on predicting the price of used car. The manufacturer's brand has a weak positive correlation with price. The region where the car is located and the mileage of the car has a weak negative correlation with the price.

Figures 4 and 5 show box plots of the vehicle prices and kilometers traveled based on the manufacturers. The list of manufacturers is determined by sorting in an ascending order, beginning with the highest value:

By price: BMW, Lamborghini, Land Rover, Mercedes-Benz, Porsche, Rolls-Royce, Toyota.

By kilometers traveled: AvtoVAZ, Mercedes-Benz, Opel, Toyota.

Analyzing the amount of cars and the cost, manufactures anticipate data based on this information. Avtovaz and Opel are particularly popular among motorists due to their competitive prices and long service life – even outperforming vehicles from higher price classes such as Mercedes-Benz and Toyota.

Figure 6 shows the plots of the features called by the plotmatrix function. Analyzing the graphs, we can say that the most expensive car brands are Rolls-Royce, Lamborghini, Mercedes-Benz, Bentley, Land Rover, Porsche, McLaren. Toyota, Ferrari, BMW, Ford, Lexus, Audi, Aston, Cadillac, Dodge. The newer the car, the more expensive it is. The most expensive cities to buy a car are Almaty, Astana, Atyrau, Kostanay, Kyzylorda, Taraz, Ust-Kamenogorsk, Shymkent. The most profitable for buying are Aktau, Zhanaozen, Zhezkazgan, Kokshetau, Taldykorgan, Turkestan, Ekibastuz. The most expensive are cars with a body like a sedan, crossover, SUV. The cheapest are cars with a body like microvan, targa, fastback, hardtop. The larger the engine volume, the more expensive the car. Gasoline, hybrid and diesel vehicles are more expensive than LPG or LPG vehicles. The lower the mileage, the more expensive the cars are. Cars with automatic transmission and robot are much more expensive. Drive units with rear-wheel drive and all-wheel drive are more expensive. Left hand drive cars are more expensive. By color, the most expensive shades are black, black metallic. Slightly more expensive are cars with white metallic, green metallic, silver, silver metallic, gray, gray metallic. For repurchase, you can buy cars with the color pink, purple, lilac, blue, yellow, cherry, turquoise. Cars with these colors are much cheaper. For further resale, you can repaint the car in black or black metallic. State registration in Kazakhstan does not greatly affect the cost of the car.

3.png

Figure 3. Correlation matrix

4.png

Figure 4. Box plot of the vehicle price according to the manufacturer

5.png

Figure 5. Box plot of kilometers traveled by manufacturer

6.png

Figure 6. Plotmatrix