ADCAIJ: Advances in Distributed Computing and Artificial Intelligence Journal
Regular Issue, Vol. 12 N. 1 (2023), e31490
eISSN: 2255-2863
DOI: https://doi.org/10.14201/adcaij.31490

Cryptocurrency Price Prediction Using Supervised Machine Learning Algorithms

Divya Chaudhary^a, Sushil Kumar Saroj^a

^aDepartment of computer science and engineering, MMMUT, Gorakhpur, India, 273010

divyachaudhary098@gmail.com, sushil.mnnit10@gmail.com

ABSTRACT

KEYWORDS

1. Introduction

A cryptocurrency is an encrypted digital currency that prevents forgery and double expenditure. It is not created by a centralized authority or banking system and is a distributed virtual currency that may be traded via encrypted means. Mined virtual currencies may be purchased on a crypto exchange. Furthermore, not all e-commerce sites accept the purchase of cryptocurrencies. Even well-known cryptocurrencies such as Bitcoin are seldom employed for retail transactions. In contrast, cryptocurrencies have gained popularity as trading tools due to their rising value. They are also used in a limited capacity for international transfers.

Blockchain is the technology that supports Bitcoin and the vast majority of other cryptocurrencies. It maintains a tamper-proof ledger of events and keeps a record of who possesses what. All peer-to-peer transactions are recorded on a blockchain which is a decentralized ledger. Participants may validate transactions even without a requirement for a centralized authority using this technology. A blockchain is, at its most fundamental level, a publicly verifiable catalog of transactions. Based on how they are employed, the component units of digital currencies are known as either coins or tokens. Some are intended to be used as mediums of exchange for goods or services, some as repositories of worth, and still others to participate in specific software applications, such as gaming and investment derivatives.

Regardless of the perspective, Bitcoin is a high-risk investment. Cryptocurrency does not fulfill the economic qualities of a currency, based on the viewpoint and characteristics of conventional money. The crypto market is one of the most tempting places for financial speculators, indicating growth in deceptive activities on social media. Some investors, especially those with a high-risk appetite, are interested in Bitcoin investments. Numerous people have had substantial gains from digital market speculation. Consequently, forecasts are used by market analysts and speculators. Due to variations in each coin’s prognostic capacity, machine learning, and artificial intelligence algorithms are rather fascinating today.

Cryptocurrencies with slow growth are more dependable than those with fluctuation. In contrast, rate patterns are very controversial and unpredictable. Based on historical data, machine learning model predictions allow organizations to formulate extremely accurate hypotheses about the predicted outcomes of a query. Machine learning can only be used to forecast cryptocurrency values when users analyze historical price fluctuations before purchasing and selling cryptocurrencies. The values of cryptocurrencies change from time to time, shifting unpredictably, as their tempo is unknown. As a consequence, forecasting is tough for academics and related fields.

The goal of this research paper is to predict the price of Bitcoin, Ethereum, Dogecoin, and Bitcoin Cash for five days using the current close price, using the supervised machine learning algorithms (SVM), linear regression, and decision tree, while comparing and evaluating the accuracy of all three methodologies. The rest of this paper is organized into a literature review in Section 2. The proposal is described in Section 3. Performance evaluation is presented in Section 4. Finally, the paper is concluded and future lines of research are discussed in Section 5.

2. Related Work

Cryptocurrencies are progressively changing the dynamics of the traditional money market. Many methods have been proposed for predicting the closest value for the currency. It is stated that social media sentiments also have a huge role in the changing dynamics of cryptocurrency. Therefore, many research works were performed previously that state the various factors that affect the currency market. This part includes the literature-based efforts on cryptocurrency price prediction using machine learning, deep learning, and social media sentiments. Several recent research works are investigated.

Jang et al. (2017) argued that bitcoin and national currencies exhibit propagation of volatility shocks, while policy uncertainty has no influence. They deployed Bayesian neural network models coupled with characteristics obtained from Bitcoin and blockchain technology’s principles. The authors validated the Currency’s high volatility and developed an effective price time-series data predictor. Madan et al. (2015) investigate the connection between big data and cryptocurrency. Classification-wise, they addressed the topic of anticipating Bitcoin’s price changes. 98.7 percent accuracy was reached for daily forecasts, and 50-55 percent utilizing 10-minute intervals.

Chen et al. (2013) established investing strategies by analyzing and categorizing Twitter feeds. Go et al. (2009) evaluate the classification performance of classifiers trained using a dataset simplified by remote monitoring. Recent research has focused on high-frequency trading and the use of deep-learning methods such as RNN for the prediction of time series data using dense, feed-forward neural networks as function models. Andrianto et al. (2017) show that including cryptocurrency in a portfolio improves its performance in two distinct ways. The primary goal is to reduce standard deviation, while the secondary purpose is to provide investors with other allocation options. The appropriate cryptocurrency allocation ranges between 5 to 20 percent, depending on the risk tolerance of the investors.

Derbentsev et al. (2021) use two machine learning approaches, random forests (RF) and stochastic gradient enhancing machine, to forecast time series data (SGBM). The results suggest that the ML ensemble method may be used to forecast the price of Bitcoin. Particularly, there is a lack in research using ML techniques to predict Bitcoin’s price. Shah et al. (2014) predicted Bitcoin’s price using a latent source model, claiming an 89 percent return in 50 days and a 4.1 Sharpe ratio. Attempts have also been made to predict the value of Bitcoin using text data from social media networks and other sources. Georgoula et al. (2014) explored sentiment analysis by combining support vector machines with the frequency of Wikipedia page views and the network hash rate.

Matta et al. (2015) analyzed the association between the price of Bitcoin, tweets, and Google Trends page views. By analyzing Google Trends, the authors forecasted Bitcoin’s trading activity rather than its price. Inadequate sample sizes and the potential for disinformation to spread across multiple (social) media channels such as Twitter or on discussion boards such as Reddit, which dramatically raise prices, are the disadvantages of such research. Hota et al. (2017), provided a model for the prediction of time series data based on the sliding window idea and the Radial Basis Functions Network (RBFN) Artificial Neural Network (ANN) approach. It illustrates some constraints, such as the introduction of hybrid or ensemble approaches with novel characteristics. Velankar et al. (2018) sought to discover and comprehend daily Bitcoin market patterns by collecting ideal characteristics around Bitcoin prices and producing a normalized graph.

Zhao et al.’s (2021) hybrid model was used for both linear and non-linear period elements of the stock dataset for future predictions. CNN and Seq2Seq LSTMs were connected well for dynamic modeling in the non-linear time series forecast including both short- and long-term dependent patterns. Yuneline et al. (2019) examined the perspectives and features of cryptocurrencies in terms of monetary characteristics, legal aspects, economic factors, and Sharia issues. Based on the perspective and characteristics of conventional money, cryptocurrencies lack the economic characteristics of a currency. Using Bayesian-optimized RNN and LSTM, McNally et al. developed two different types of Bitcoin price forecasting models. The analysis revealed that LSTM had 52% accuracy and 8% RMSE.

3. Proposed Approach

To build a prediction system using machine learning, there is a fixed set of steps that involves data collection, training the historical data, and then predicting the result by using the appropriate machine learning algorithms. Figure 1 depicts the steps involved in the prediction of the cryptocurrency price. The machine learning algorithms used in the paper are: support vector machine, linear regression, and decision tree. For this investigation, we chose to focus on four distinct cryptocurrencies due to their widespread adoption because there is a wide availability of historical price data, and overall market capitalization, namely, Bitcoin, Ethereum, Bitcoin Cash, and Dogecoin.

3.1. Data Acquisition

Coin Metrics and Coin Codex were used to collect the data for this research paper. The datasets for all four are in different CSV format files. Each dataset contains the same number of columns each defining the attribute of the currency. The dataset contains a total of six features. The details for them are as follows: