Predicting the price of Bitcoin by the most frequent edges of its transaction network

doi:10.1016/j.econlet.2019.108655

Economics Letters

Volume 184, November 2019, 108655

https://doi.org/10.1016/j.econlet.2019.108655 Get rights and content

Highlights

•
The transaction network’s influence on the future price of Bitcoin is investigated.
•
The most frequent edges have significant predictive power for Bitcoin price movements.
•
An efficient approach is provided for applying a network dataset in day trading.

Abstract

Research on the Bitcoin transaction network has increased rapidly in recent years, but still, little is known about the network’s influence on Bitcoin prices. The goals of this paper are twofold: to determine the predictive power of the transaction network’s most frequent edges on the future price of Bitcoin and to provide an efficient technique for applying this untapped dataset in day trading. To accomplish these goals, a complex method consisting of single-hidden layer feedforward neural networks (SLFNs) is used. Based on the results, the presented method achieved an accuracy of approximately 60.05% during daily price movement classifications, despite only considering a small subset of edges.

Introduction

Bitcoin is not only the world’s leading cryptocurrency but also a pioneering blockchain technology.¹ In contrast to the traditional financial systems, where the records of daily transactions are considered highly sensitive and are kept private (Kondor et al., 2014), this technology ignores financial intermediaries and provides a public ledger that records all Bitcoin transactions. This ledger can be represented by a directed graph, called the transaction network, in which transactions temporarily connect the input and output addresses of Bitcoin users.

The research on this network has increased rapidly in recent years, but we still know little about the network’s influence on Bitcoin prices. To the best of our knowledge, only a few papers have addressed this issue to date. Among these, Kondor et al. (2014) investigated how the structural changes in the network accompany significant changes in the exchange price of Bitcoin. Akcora et al., 2018b, Akcora et al., 2018a found evidence that some subgraphs of the transaction network have a predictive influence on Bitcoin prices. Koutmos (2018) detected bidirectional linkages between Bitcoin returns and transaction activities. Other studies used blockchain, price and some network data for the daily price movement classifications. Among these, McNally et al. (2018) used the closing price of Bitcoin as the input variable and achieved an accuracy of $52.78 %$ by applying a long short-term memory (LSTM) recurrent neural network. Madan et al. (2015) reached $98.7 %$ accuracy using a generalized linear model and 25 independent variables.

The aims of the present study are twofold: to investigate the transaction network’s influence on the future price of Bitcoin and to provide an efficient technique for applying this untapped dataset in day trading. To satisfy both aims, this paper focuses only on the most frequent edges of the transaction network. The assumption behind this approach is that these edges may belong to the largest traders and probably have the highest predictive power on the future price of Bitcoin.² Thus, not only the number of fake signals on price movement but also the size of the network can be reduced, facilitating the processing of this dataset. This concept is illustrated in Fig. 1.

Section snippets

Bitcoin dataset

In the process of creating the transaction network, this paper follows the procedure of Kondor et al. (2014).³ For this reason, we use the authors’ dataset and source code of their modified client program, helping us to compile the graph. This dataset contains the raw and preprocessed data of

Results and discussion

Based on the empirical results, the most accurate model is obtained by the following parameters: $t_{T R} = 16, α = 8000$ and $β = 100$ . With this setup, for the daily price movement classifications between November 25, 2016 and February 5, 2018, we reached an accuracy of $60.05 %$ . The results are presented in detail in Fig. 4.

Compared with the results from McNally et al. (2018), we found that the most frequent edges of the transaction network have significantly higher predictive power for the price movement

Conclusions and future works

The research on the Bitcoin transaction network has increased rapidly in recent years, but we still know little about its influence on Bitcoin prices. In this paper, we examined the predictive power of this network’s most frequent edges on the future price of Bitcoin with a complex method based on single hidden layer feedforward neural networks (SLFNs). We not only found a significantly high accuracy ( $60.05 %$ ) for the price movement classifications but also showed that this information can be

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgments

This research was supported by the Pallas Athene Domus Scientiae Foundation, Hungary and the ÚNKP-18-3 New National Excellence Program of the Ministry of Human Capacities, Hungary . The views expressed are those of the author’s and do not necessarily reflect the official opinion of supporters.

References (19)

AkcoraC.G. et al.
Bitcoin risk modeling with blockchain graphs
Econom. Lett.
(2018)
DemirE. et al.
Does economic policy uncertainty predict the Bitcoin returns? An empirical investigation
Finance Res. Lett.
(2018)
DyhrbergA.H.
Bitcoin, gold and the dollar–a GARCH volatility analysis
Finance Res. Lett.
(2016)
KatsiampaP.
Volatility estimation for Bitcoin: A comparison of GARCH models
Econom. Lett.
(2017)
KoutmosD.
Bitcoin returns and transaction activity
Econom. Lett.
(2018)
UrquhartA.
The inefficiency of Bitcoin
Econom. Lett.
(2016)
AkcoraC.G. et al.
Forecasting bitcoin price with graph chainlets
CoindeskC.G.
Bitcoin price dataset
(2018)
HoT.K.
Random decision forests

There are more references available in the full text version of this article.

Cited by (41)

Models used to characterise blockchain features. A systematic literature review and bibliometric analysis
2023, Technovation
Blockchain has emerged as an innovative technology with potential to transform business management, through operational efficiency improvements. Nevertheless, several performance and vulnerability issues have been identified for the different typologies supporting the wide range of blockchain-based applications currently implemented in different domains.
A variety of analytical and empirical models are being used to evaluate the issues associated with the different blockchain typologies, enabling systematic analyses of the corresponding efficiency impact, and technical or economic threats.
A thorough systematic literature review of these models has been performed, followed by a detailed assessment on the way these models have been employed, and the target parameters and applications evaluated (336 research selected and analysed). We propose a co-classification of these models, allowing us to identify which ones are employed to a greater extent to address the different blockchain issues in scientific research. In a second step, a bibliometric analysis on the selected research is conducted, offering a complementary overview of the status of and trends in blockchain modelling, including the most prolific authors and leading contributing countries to the topic.
The main outcome and contribution of the paper is the provision of a broad overview on how blockchain issues have been analytically tackled, through the synthesis and meta-analysis of the models used in the scientific literature since the inception of blockchain technology. The results have two main direct applications, firstly supporting novel vulnerability and performance analyses of existing blockchain applications by providing historical information on the models used so far, as well as the key parameters and typology of the blockchain-based applications evaluated. Secondly, in the implementation of new applications, by allowing the recognition of key issues identified that are associated with the different blockchain typologies and to determine the most suitable models to analyse the weaknesses and risks of the alternative designs under evaluation for these new implementations.
A Tree Augmented Naïve Bayes-based methodology for classifying cryptocurrency trends
2023, Journal of Business Research
As the popularity of blockchain technology and investor confidence in Bitcoin (BTC) increased in recent years, many individuals started making BTC and other cryptocurrency investments, in expectation of high returns. However, as recent market movements have shown, the lack of regulation and oversight makes it difficult to guard against high volatility and potentially significant losses in this sector. In this study, we propose a data-driven Tree Augmented Naïve (TAN) Bayes methodology that can be used for identifying the most important factors (as well as their conditional, interdependent relationships) influencing BTC price movements. As the model is parsimonious without sacrificing accuracy, sensitivity, and specificity—as evident from the average accuracy value—the proposed methodology can be used in practice for making short-term investment decisions.
Forecasting cryptocurrencies prices using data driven level set fuzzy models
2022, Expert Systems with Applications
The paper develops fuzzy models to forecast cryptocurrencies prices using a data-driven fuzzy modeling procedure based on level set. Data-driven level set is a novel fuzzy modeling method that differs from linguistic and functional fuzzy modeling in how the fuzzy rules are built and processed. The level set-based model outputs the weighted average of output functions of active fuzzy rules. Output functions map the activation levels of the fuzzy rules directly in model outputs. Computational experiments are done to evaluate the level set method in one-step-ahead forecasting of the closing prices of cryptocurrencies. Comparisons are made with the autoregressive integrated moving average, multi layer neural network, and the naïve random walk as a benchmark for Cardano, Binance Coin, Bitcoin, Ethereum, Chainlink, Litecoin, Tron, Stellar, Monero and Ripple. The results suggest that the random walk outperforms most methods addressed in this paper, confirming the Meese–Rogoff puzzle for the case of digital coins, i.e. the difficulty to surpass the naïve random walk in predicting exchange rates. However, when performance is measured by the direction of price change, the level set-based fuzzy modeling performs best amongst the remaining methods.
Past, present, and future of the application of machine learning in cryptocurrency research
2022, Research in International Business and Finance
Cryptocurrency has captured the interest of financial scholars and become a major research topic in blockchain. In cryptocurrency research, the use of machine learning algorithms is enabled by the presence of many types of data and abundant resources. However, there is currently no comprehensive review on cryptocurrencies using machine learning. Therefore, we collect papers on cryptocurrency-related using machine learning in the web of science database, and summarize these papers according to the algorithm, and draw the following conclusions: (1) The application of machine learning for cryptocurrencies research is increasing year over year; (2) Predicting cryptocurrency price trends and income fluctuations is the most relevant research topic; (3) The machine learning algorithm utilized in cryptocurrency research is not unique, and the practise of combining multiple machine learning approaches has emerged; (4) Concerns such as overfitting and interpretability still persist with machine learning methods. Finally, we suggest future research directions.
An interpretable decision-support systems for daily cryptocurrency trading
2022, Expert Systems with Applications
Citation Excerpt :
Their results commonly showed that employing more features (in addition to crypto-related predictors) such as stock market, google trends, etc., help increase the accuracy of the models. On the other hand, among the studies that fall under group 2b, Kurbucz (2019) and, Mallqui and Ricardo (2019) are among the studies that employed feature reduction techniques. Kurbucz (2019) specifically focused on determining the prediction ability of the transaction network’s most frequent edges on the price trend of Bitcoin.
Cryptocurrencies, especially Bitcoin (BTC), have become an important commodity for both individual and corporate investors within the last decade. The limited supply, high volatility, and random price fluctuations have increased investors' interest in BTC, especially in daily trading. Although BTC has been yielding a high rate of returns, price fluctuations and constant speculations make the investors wary of unexpected price movements. Predictive modeling suffers from the complexity of the datasets (i.e., the high number of features employed to forecast BTC movements) as well as the black-box nature of most machine learning algorithms (which is especially problematic for corporate investors since they are obligated to disclose their investment decisions to their clients). Therefore, the main goal of the current study is to assist individual and corporate investors in making transparent and interpretable daily BTC trading decisions by developing a predictive analytics framework. To address the complexities posed by the datasets, a comprehensive tri-level feature selection approach is proposed. The selected features are then, fed into the Classification & Regression Tree (C&RT) to build a highly parsimonious, transparent, and interpretable prediction model. The resultant model was not only evaluated on the test (holdout) sample but was also tested on challenging time periods, including the first half of 2020 (the start of the pandemic era) to exhibit the viability and reliability of the proposed framework. Finally, a decision support tool is developed for the practical implementation of the model. The tool can be used by short-term investors not only due to its highly simplistic, transparent, and interpretable structure, but also its higher accuracy, sensitivity, and specificity results when compared to the extant literature.
The price and cost of bitcoin
2022, Quarterly Review of Economics and Finance
Explaining changes in bitcoin’s price and predicting its future have been the foci of many research studies. In contrast, far less attention has been paid to the relationship between bitcoin’s mining costs and its price. One popular notion is the cost of bitcoin creation provides a support level below which this cryptocurrency’s price should never fall because if it did, mining would become unprofitable and threaten the maintenance of bitcoin’s public ledger. Other research has used mining costs to explain or forecast bitcoin’s price movements. Competing econometric analyses have debunked this idea, showing that changes in mining costs follow changes in bitcoin’s price rather than preceding them, but the reason for this behavior remains unexplained in these analyses. This research aims to employ economic theory to explain why econometric studies have failed to predict bitcoin prices and why mining costs follow movements in bitcoin prices rather than precede them. We do so by explaining the chain of causality connecting a bitcoin’s price to its mining costs.

View all citing articles on Scopus

View full text

Predicting the price of Bitcoin by the most frequent edges of its transaction network

Highlights

Abstract

Introduction

Section snippets

Bitcoin dataset

Results and discussion

Conclusions and future works

Declaration of Competing Interest

Acknowledgments

Econom. Lett.

Finance Res. Lett.

Finance Res. Lett.

Econom. Lett.

Econom. Lett.

Econom. Lett.

Forecasting bitcoin price with graph chainlets

Bitcoin price dataset

Random decision forests