Evolutionary dynamics of cryptocurrency transaction networks: An empirical study
RRESEARCH ARTICLE
Evolutionary dynamics of cryptocurrencytransaction networks: An empirical study
Jiaqi Liang * , Linjing Li * , Daniel Zeng The State Key Laboratory of Management and Control for Complex Systems, Institute of Automation,Chinese Academy of Sciences, Beijing, China, School of Computer and Control Engineering, University ofChinese Academy of Sciences, Beijing, China, Department of Management of Information Systems, TheUniversity of Arizona, Tucson, AZ, United States of America * [email protected] (JL); [email protected] (LL) Abstract
Cryptocurrency is a well-developed blockchain technology application that is currently aheated topic throughout the world. The public availability of transaction histories offers anopportunity to analyze and compare different cryptocurrencies. In this paper, we present adynamic network analysis of three representative blockchain-based cryptocurrencies: Bit-coin, Ethereum, and Namecoin. By analyzing the accumulated network growth, we find that,unlike most other networks, these cryptocurrency networks do not always densify over time,and they are changing all the time with relatively low node and edge repetition ratios. There-fore, we then construct separate networks on a monthly basis, trace the changes of typicalnetwork characteristics (including degree distribution, degree assortativity, clustering coeffi-cient, and the largest connected component) over time, and compare the three. We find thatthe degree distribution of these monthly transaction networks cannot be well fitted by thefamous power-law distribution, at the same time, different currency still has different networkproperties, e.g., both Bitcoin and Ethereum networks are heavy-tailed with disassortativemixing, however, only the former can be treated as a small world. These network propertiesreflect the evolutionary characteristics and competitive power of these three cryptocurren-cies and provide a foundation for future research.
Introduction
Network analysis, such as those reported in [1–4], has attracted increasing attention in eco-nomics and finance since it provides further insights than traditional methods. Although alarge volume of financial data, e.g., stock price, is available for network related research andanalysis, information about transaction details is usually considered sensitive and not availablefor research. Cryptocurrency, where a continuously growing list of records stored in a chain is accessible, provides opportunities to analyze transaction networks in detail.
A cryptocurrency is a digital currency in which blockchain techniques are used to securethe transactions and control the generation of new units of currency (the so-called coins),operating independently without a central authority. Specifically, cryptocurrency relies on a
PLOS ONE | https://doi.org/10.1371/journal.pone.0202202 August 17, 2018 1 / 18 a1111111111a1111111111a1111111111a1111111111a1111111111
OPEN ACCESS
Citation:
Liang J, Li L, Zeng D (2018) Evolutionarydynamics of cryptocurrency transaction networks:An empirical study. PLoS ONE 13(8): e0202202.https://doi.org/10.1371/journal.pone.0202202
Editor:
Alejandro Raul Hernandez Montoya,Universidad Veracruzana, MEXICO
Received:
January 19, 2018
Accepted:
July 30, 2018
Published:
August 17, 2018
Copyright: © Data Availability Statement:
All relevant data areavailable from the Harvard Dataverse database(doi: 10.7910/DVN/IO50XZ, 10.7910/DVN/XIXSPR,10.7910/DVN/M9K5OJ).
Funding: rivate key to prove ownership and a public history of transactions to prevent double-spend-ing [5]. Since Bitcoin [6], the first cryptocurrency, emerged in 2009, many other alternativeshave emerged with modified rules of transaction and usage, e.g., Namecoin provides decentral-ized name registration [7], while Ethereum allows the automatic transfer of digital assetsaccording to the so-called “smart contract” [8]. By introducing new types of assets and newtransaction management methods, cryptocurrency has the potential to replace traditional fiat-currency. At the time of writing, there are over 900 currencies in the cryptocurrency marketand the total market cap has exceeded $578 billion [9, 10]. Thus, it is the right time to investi-gate and compare them, so as to fully understand cryptocurrency and provide a foundation forfuture research.Public availability of cryptocurrency transactions provides a basis for analyzing its transac-tion networks. For networks, especially the so-called complex networks, reported investiga-tions mainly focus on descriptive statistics, network evolution, statistical mechanics ofnetwork topology and dynamics [11]. There are also studies on the robustness against failuresand attacks, spreading processes and synchronization [12]. The descriptive statistics are majorly adopted to depict the behavior of Bitcoin users [13, 14]. In the field of Namecoin,
Kalodner et al. [15] analyzed the uses and the transfers of the namespace, they also devisedsome principles on mechanism design. Relating to the evolution of networks, most networksencountered in practice have the tendency to densify over time [16], however, Bitcoin networkdensifies only in its first five years [17] and Namecoin network only densifies in the first year[18]. Motivated by empirical data, complex networks have some typical structure features,including small worlds, clustering, and degree distribution fitted by the power law. Baumannet al. [14] found that the Bitcoin system was a “small world” network and followed a scale-freedistribution. Kondor et al. [19] further illustrated that the transaction networks are character-ized by disassortative degree correlation in the trading phase, they applied linear preferentialattachment to interpret it. Regarding research on multiple currencies, Anderson et al. [20]studied the characteristics of three representative cryptocurrencies separately, but no compari-son of network characteristics was provided. Walsh et al. [21] identified eight key characteris-tics of system design and divided currencies into four archetypes, but there was no in-depthnetwork analysis.In this paper, we apply statistics and network analysis methods to explore the dynamiccharacteristics of three transaction networks. We download transaction data from the respec-tive blockchain explorers. To the best of our knowledge, these are the largest datasets adoptedin cryptocurrency analysis to date. We analyze the growth pattern of the accumulated networkand find that unlike most networks, these cryptocurrency networks do not always densify overtime. Then based on the datasets, we find that the monthly repetition ratios measured by eithernode or edge are relatively low. As such, studying the whole accumulated network, as done inmost previous work [18, 19], is not the appropriate way to understand the network dynamics.Hence we focus on coining the dynamics through computing the values of typical networkmeasures on a monthly basis, and make a comparison among the three networks.The main contributions of our research are: 1) We find that the growth pattern of crypto-currency transaction networks is different from that of most other networks reported in the lit-erature in the way that they do not always follow neither the densification law nor the constantaverage degree assumption over time; 2) Monthly network, instead of accumulated network, isproposed as an appropriate object to understand the dynamics of the network; 3) we conduct the first empirical comparison among three representative cryptocurrency networks and point out the similarities and differences to help understand the peer-to-peer technology on a net-work level. Different from previous researches on complex networks, we find that the degree
Evolutionary dynamics of cryptocurrency transaction networksPLOS ONE | https://doi.org/10.1371/journal.pone.0202202 August 17, 2018 2 / 18
Competing interests:
The authors have declaredthat no competing interests exist. istribution of the cryptocurrency transaction networks cannot be well fitted by the famouspow-law distribution.The remainder of this paper is organized as follows. In the next section, we provide ourdatasets, the necessary background to understand the transaction networks and our methodol-ogy used to analyze the networks. The Results section presents our findings for Bitcoin, Ether-eum, and Namecoin networks. We offer our conclusions in the last section.
Methodology
In this section, we first introduce the datasets used for analysis then explain how to construct atransaction network from corresponding dataset, the transaction network is the basis for thesubsequent dynamic analysis. Finally, we introduce the measures used for the networkanalysis.
Datasets
Among the complete list of cryptocurrencies, we choose three representatives for our analysis:
Bitcoin, Namecoin, and Ethereum. Bitcoin is chosen as it is the first and by far the largest cryp-tocurrency; Namecoin is the first cryptocurrency that works as a decentralized domain namesystem; and Ethereum is the first cryptocurrency that supports “smart contract” and is alsoone of the most active cryptocurrencies. The data on transactions are from the blockchainexplorers [22–24]. We believe, but cannot fully verify, that the data should be the same as whatone could get as a cryptocurrency client. Even if there are tiny differences, they are likely tohave only a negligible effect on our statistical results. We downloaded the complete list oftransactions of each currency from its inception through 31 October 23:59:59 2017 UTC. Asummary of the datasets is provided in Table 1.
Transaction network
Blockchain is a distributed public ledger that records transactions ever verified in the network.It is implemented as a chain of blocks, each block containing a hash of the previous block upto the genesis block of the chain. And each block holds batches of valid transactions in theform of owner X transferring Y coins to payee Z. In the cryptocurrency system, payers andpayees can create an unlimited number of addresses. A transaction in cryptocurrency systemis a kind of regular bank transaction in the sense that it allows multiple sending addresses andmultiple receiving addresses existing in a transaction.Take the Bitcoin system as an example, it specifies how many Bitcoins are sent or receivedfrom an address, but there are no details of who sends how many Bitcoins to whom. Fig 1Ashows an example of the transaction with two sending addresses and two receiving addresses
Table 1. Summary of cryptocurrencies studied in this paper.Cryptocurrency Bitcoin Namecoin EthereumTime of genesis block
Size on disk https://doi.org/10.1371/journal.pone.0202202.t001
Evolutionary dynamics of cryptocurrency transaction networksPLOS ONE | https://doi.org/10.1371/journal.pone.0202202 August 17, 2018 3 / 18 hich was added on the blockchain on May 1, 2011, and the relevant details can be queried onthe corresponding crawling website through the identifier. Specifically, the time in the upperright corner indicates when the transaction was added to the blockchain, and the value on thefirst row is the transaction identifier, i.e., a hash value. Next, “1BTC” and its neighboring value(hash value of transaction address) denote that the address sent 1 Bitcoin. Therefore, Fig 1Ashows a transaction that two sending addresses contribute 1 Bitcoin and 135 Bitcoins respec- tively, the two receiving addresses receive 135.67 Bitcoins (for payment) and 0.33 Bitcoins (maybe a transaction fee or a new address for remaining bitcoins). Fig 1B is an example infor-mation extracted from Bitcoin transactions where the value of the arrow represents corre-sponding value in Bitcoins that are flowing. Here a i ’s represent addresses and t i ’s denote Fig 1. Illustration of transaction network construction. (A) An example of Bitcoin transaction details. (B) Exampleinformation extracted from Bitcoin transactions, and the information in the orange box correspond to the transaction in(A). (C) The Bitcoin transaction network as a directed graph. https://doi.org/10.1371/journal.pone.0202202.g001
Evolutionary dynamics of cryptocurrency transaction networksPLOS ONE | https://doi.org/10.1371/journal.pone.0202202 August 17, 2018 4 / 18 ransactions. t is a transaction with two inputs ( a and a ) and two outputs ( a and a ). Thetransaction was added to the blockchain on May 1, 2011. t is a transaction with five inputsand two outputs ( a and a ) happened on the same day. note that two inputs (i.e. a and a ) of t are connected to the aforementioned outputs of t . An input to a transaction is either theoutput of a previous transaction or incentives (including newly generated bitcoins and transac-tion fees) for users. Regarding the number of transaction inputs, it can be a single input from aprevious larger transaction, or multiple inputs combining smaller amounts. For security pur-poses, a transaction may have multiple outputs: one for the transfer of the rest, if any, back tothe sender, and the other is used for the payment.Public availability of cryptocurrency transactions and the input-output relationshipbetween transactions provide a basis for transaction network research. The transaction net-work represents the flow of cryptocurrency between addresses over time. In a transaction net-work, each node represents an address. Without the specific value of cryptocurrency flow frominputs and outputs, there is an edge with a timestamp between any sending address and receiv-ing address existing in a transaction. For instance, Fig 1C shows the network constructed from transactions in Fig 1B. Network measures
In the first part of our analysis, several descriptive statistics are calculated to analyze the accu-mulated network growth. The number of edges and nodes are adopted to represent the net-work size. Many networks encountered in practice densify over time with the average degreeincreasing, which means the number of edges grow superlinearly with respect to the numberof nodes. This property is quantified by e ( t ) * n ( t ) a , where e ( t ) and n ( t ) denote the number ofedges and nodes of the graph at time t respectively, and a > Ratio tRep ¼ j E t \ E t (cid:0) jj E t [ E t (cid:0) j ; ð Þ where E t refers to the set of edges or nodes in a network at time t , \ is the intersection operator and [ is the union operator as in ordinary set theory, and | (cid:1) | gives the number of elements when applying to a specific set.For the monthly networks, we further analyze the dynamic characteristics to investigate thetopologic properties. We select four most representative measures for analysis, includingdegree distribution, degree assortativity, average clustering coefficient, and properties of theLCC. The network measures adopted are briefly introduced in the following. Degree distribution captures the individual connectivity of nodes [11]. The in(out)-degreeof a node represents the number of transactions it involves as output(input), and the degreedistribution is the probability distribution of these degrees over the whole network. Empiri-cally, observed complex networks tend to show a heavy-tailed distribution following a power- law distribution p ( k ) * k − γ , where k is the value of the degree and the coefficient γ has beenfound to be the characteristic of a complex network [25]. Degree assortativity measures the node preference—that nodes with similar degreestend to be connected to each other [26]. Its strength, expressed as the degree assortativity
Evolutionary dynamics of cryptocurrency transaction networksPLOS ONE | https://doi.org/10.1371/journal.pone.0202202 August 17, 2018 5 / 18 oefficient, denoted by r , is defined as: r ¼ M (cid:0) P i j i k i (cid:0) M (cid:0) P i j i þ k i ð Þ (cid:2) (cid:3) M (cid:0) P i j i þ k i (cid:0) (cid:1) (cid:0) M (cid:0) P i j i þ k i ð Þ (cid:2) (cid:3) ; ð Þ where j i and k i are the degrees of the nodes at the ends of the i -th edge, and M is the number ofedges. Clustering coefficient represents the tendency of the nodes in a graph to cluster together,and the overall level of clustering is measured as the average of the local clustering coefficientof all nodes: C ¼ N X v j D v j d v ð d v (cid:0) Þ = ; ð Þ where | Δ v | denotes the number of triangles containing node v . To calculate | Δ v |, we ignore thedirectionality of the graph, and d v is the degree of node v in the undirected graph. Watts and Strogatz [25] applied the clustering coefficient to discover small-world phenomenon within several networks.
The largest connected component (LCC) is a maximal subgraph in which any two nodescan be connected by a path. LCC is an important factor in understanding the network struc-ture [11]. In this paper, we adopt relative size and the diameter of the LCC. The relative size iscalculated by dividing the number of nodes that connect to the LCC by the number of nodesin the whole network. The diameter is the longest shortest path among all the nodes that formthe LCC.
Results
The analysis of cryptocurrency networks is conducted from three perspectives. In the firstpart, we explore the accumulated network growth. Then we select the appropriate investiga-tion object for analysis. In the last part, we focus on analyzing the dynamics of the monthlynetworks and making comparisons. The analysis program is implemented in Python with theaid of powerlaw [27], Networkit [28], and statsmodel [29] packages.
Accumulated network growth
In this part, we investigate the network growth from cryptocurrencies’ inception till 31 Octo-ber, 2017. For each month m , we construct a network using all transactions published up tomonth m . We analyze two aspects: network size (number of nodes and edges) and averagedegree.The number of edges and nodes can be adopted to represent the size of the network, andthey indicate the adoption rate and competitiveness of currency. As shown in Fig 2, the growthprocess can be divided into two phases.• Initial phase . The system had low activity. Users just tried the currency experimentally andcompared it with other currencies to find relative advantages. When a currency becamemore popular, more users would adopt it. Therefore, the network exhibited growing ten-dency with excessive fluctuations. • Trading phase . With a certain number of adopters, growth slowed and did not change sig- nificantly. A reason is that the currency is constantly being accepted and rejected as a resultof competition with other cryptocurrencies in the market.
Evolutionary dynamics of cryptocurrency transaction networksPLOS ONE | https://doi.org/10.1371/journal.pone.0202202 August 17, 2018 6 / 18 hen referring to the specific growth rates and duration time, Bitcoin grew over 10,000times bigger in its first two-and-a-half years, while Namecoin and Ethereum grew over 100times bigger in their first year. A reason for Bitcoin’s long duration is that during 2009 to 2010,cryptocurrency was a new concept and Bitcoin was the only cryptocurrency in the market. Allusers who wanted to try cryptocurrency had to choose Bitcoin.Then we investigate the average degree over time to find the network’s tendency to becomedense. Growth patterns in Fig 3 show the differences among the three networks. For Bitcoin,the average degree increased over time until September 2015. Subsequently, the decrease lastedfor almost two years, probably because it had issues, such as hard to mine and large price fluc-tuations, and its competitor Ethereum offered a new option, “smart contract,” for users inter-ested in cryptocurrencies. Bitcoin has shown an increase since July 2017. For Namecoin, except for the increase in the initial phase, the average degree remained constant with some fluctuations due to competition among currencies. For Ethereum, the average degree contin-ued increasing except for a decrease in October 2016. We suspect that the network instabilityis caused by a number of denial-of-service (DoS) attacks in late September and the two-stage“hard fork” to secure the network [30].To gain more insight, we plot the number of nodes versus the number of edges for eachcryptocurrency network on a logarithmic scale and fit a line reflecting the overall growth pat-tern of the network, as shown in Fig 4. The fitting parameters are shown in Table 2. For Bit-coin, the exponent is a = 1.15, which is clearly greater than 1, indicating a large deviation fromlinear growth with increasing average degree. For Ethereum and Namecoin, the exponent isclose to 1, corresponding to the constant average degree over time. We also check the latest 1/3of the data. Surprisingly, the Bitcoin network exponent is less than 1, the Ethereum networkexponent is larger than 1, and the Namecoin exponent is close to 1, which coincides with the Fig 2. The size of accumulated transaction networks with respect to various cryptocurrencies in log coordinate.
The number of nodes and edges are used to represent the size of networks. The three networks have similar growthpattern with rapid growth first and slower growth later. https://doi.org/10.1371/journal.pone.0202202.g002
Evolutionary dynamics of cryptocurrency transaction networksPLOS ONE | https://doi.org/10.1371/journal.pone.0202202 August 17, 2018 7 / 18 indings in Fig 3. The difference between the results of all data and the last 1/3 of the data indi-cates that the overall trend does not represent the real-time situation.The above analysis on accumulated networks indicts that the cryptocurrency transactionnetworks do not always follow the densification law and the constant average degree assump-tion, which is different from most networks investigated in [16]. We must point out that thereare several previous researches on cryptocurrency which have reported similar findings.Chang et al. [18] recognized that Namecoin only densifies in the first year while Holtz et al.[17] verified that Bitcoin densifies in the first five years. However, our conclusion is more validand general since our conclusion is based on a quantitative analysis on three cryptocurrenciesand our dataset covers a longer history.
Fig 3. The average node degree of accumulated networks over time.
The average degree of the three networks is notconstant. https://doi.org/10.1371/journal.pone.0202202.g003
Fig 4. The number of edges e ( t ) versus the number of nodes n ( t ) in accumulated transaction networks in log coordinate. The red lines show fitted power-law distribution for the networks. In the figure’s equation, x represents the number of nodesand ^ y represents the fitting number of edges, and the exponents are 1.15, 1.00, 1.05, respectively. https://doi.org/10.1371/journal.pone.0202202.g004 Evolutionary dynamics of cryptocurrency transaction networksPLOS ONE | https://doi.org/10.1371/journal.pone.0202202 August 17, 2018 8 / 18 hy do the cryptocurrency networks not obey the densification law? Security is the mostprobable explanation. In cryptocurrency system, to securely receive, store, and send coins, auser can spread his coins in multiple wallets, corresponding to multiple nodes in the network,to reduce risks. Therefore, in a transaction network, one user may have multiple nodes corre-sponding to multiple addresses. While in other real networks, a user usually has only onenode.
The object for dynamic analysis
Since the nodes and edges of the networks are changing all the time, we checked the monthlyrepetition ratio as shown in Fig 5. As to the nodes, Bitcoin and Namecoin have repetition ratioless than 0.1, while the value of Ethereum is less than 0.25. As to the edges, Bitcoin and Name-coin have repetition ratio less than 0.1, and the value of Ethereum is less than 0.2. Thus, afterthe initial phase, both the node and edge repetition ratio reach relatively low values, indicatingthat a lot of nodes and edges do not survive from one time window to the next and networkreconfiguration takes place all the time. The low survival ratio of both nodes and edges can be
Table 2. Fitting parameters of the power law. total lastest 1/3 of the data a R
95% confidence level a R
95% confidence level
BTC 1.15 0.998 [1.145, 1.163] 0.86 0.987 [0.830, 0.900]ETH 1.00 0.995 [0.970, 1.029] 1.38 0.998 [1.332, 1.429]NMC 1.05 0.989 [1.027, 1.078] 0.99 0.982 [0.938, 1.049]Here a is the exponent of e ( t ) * n ( t ) a , and R is the coefficient to measure the goodness of fit. It ranges from 0 to 1, the better the power law fits the data, the closer thevalue of R is to 1. https://doi.org/10.1371/journal.pone.0202202.t002 Fig 5. Monthly repetition ratios over time. (A) The ratios of edges. (B) The ratios of nodes. After the initial phase, allratios reach relatively low values. https://doi.org/10.1371/journal.pone.0202202.g005
Evolutionary dynamics of cryptocurrency transaction networksPLOS ONE | https://doi.org/10.1371/journal.pone.0202202 August 17, 2018 9 / 18 llustrated by the aforementioned security reason as well. Specifically, in order to enhancesecurity, users may constantly change their addresses. These surviving addresses may be theaddresses of fixed payees, such as donors and miners, and these addresses are used to receivecryptocurrencies which are inconvenient to replace.Due to the above characteristics of the cryptocurrency transaction networks, it is better toanalyze the transaction networks in separate time intervals, rather than in the accumulatedmanner as done by most of the previous works. In the following of this paper, we set the timeinterval as one month and construct the monthly transaction networks to understand thedynamics of the transaction networks.
Monthly network analysis
In this part, we present our main results on dynamic characteristics of cryptocurrency basedon the monthly networks.
Degree distribution.
We first check whether the degree distribution of the three represen- tatives can be fitted by the power law. In the case of cryptocurrency networks, often the initial, small values of the data do not follow the power-law distribution, thus we ignore these datawhen fitting. Further, we use the Kolmogorov-Smirnov (KS) test to assess the goodness-of-fit.We find that almost all degree distributions cannot be accepted as the power law strictly underthe 95% confidence level. However, the degree distribution is still a clear heavy-tailed distribu-tion, which means that the majority of addresses have low degrees, while small but not negligi-ble addresses have relatively high degrees. As shown in Fig 6, we divide the phases as follows.• The number of adoption users is small, and there exist large errors in the fitting. Specifically,Bitcoin has a longer duration for the reason discussed in the “Accumulated network analy-sis” section.• With a certain number of adoption users, the data are approximately fitted by the power law,though the acceptance rate using the KS test of power-law fit on the degree distribution islow. And the exponent fluctuates within a certain range. Specifically, the ranges of the expo-nent γ are: [2.0, 3.0] for Bitcoin, [1.5, 3.0] for Ethereum, and [1.5, 3.5] for Namecoin. Notethat the coefficient for the power law typically lies in the range [2, 3] as reported in [31].• Due to fierce competition with other currencies, the range of data that satisfies the power-law distribution narrows. During and after the transition phase, different currencies havedifferent features as follows: for Bitcoin and Ethereum, after the transition stage, the data areagain approximately fitted by the power law and the exponents do not change significantly;for Namecoin, in the transition stage, there exists a phenomenon that the number of nodeswith large degree is large too, thus it does not fit the power law.Our analysis suggests that when adoption users reach a certain amount, the distributionapproximately fits with the power law. However, under market competition, the scale ofNamecoin network’ nodes stabilize at a relatively small level of ten thousand, while the othertwo networks have millions of nodes. At the same time, due to the specific function of domainregistration, there are some enthusiasts who insist on using Namecoin, leading to the phenom-enon that the number of nodes with large degree is large too. In summary, under the effect ofmarket competition, failed currencies do not fit well with the power law, while successful cur- rencies approximately fit with the power law with fixed exponents. Degree assortativity.
We use the in-assortativity r ( in , in ) and out-assortativity r ( out , out )to further investigate how the nodes are mixing by the degree in the network. A positive valuefor r (assortative mixing) indicates that high-degree nodes are preferentially attached to other Evolutionary dynamics of cryptocurrency transaction networksPLOS ONE | https://doi.org/10.1371/journal.pone.0202202 August 17, 2018 10 / 18 igh-degree nodes. A negative value for r indicates disassortative mixing, i.e., high-degreenodes are prone to connect to low-degree ones. Finally, r = 0 (neutral mixing) indicates thenetwork is non-assortative.As shown in Fig 7, except for the initial phase, the ranges of the in-assortativity and out-assortativity are [-0.05, 0] for Bitcoin and [-0.1, 0] for Ethereum, which suggests that the twonetworks are disassortative. The coefficients of Namecoin stay in the range of [-0.1, 0.1], mak-ing it difficult to judge its degree assortativity. In general, small values of r are hard to interpret,thus we measure the quantity h k nn i ¼ P k k P c ð k j k Þ , i.e., the average degree of nearest neigh-bors of nodes with degree k , for the in-degree and out-degree of the last month (October2017).In networks without degree correlations, the degrees of connected nodes do not depend oneach other. Therefore for such networks, we expect the h k nn i of the in-degree and out-degree Fig 6. Samples of degree distributions of monthly networks.
Data are sampled from Bitcoin (top row), Ethereum (middlerow), and Namecoin (bottom row). Example data for power-law fitting are approximate fit (first column), poor fit (mediumcolumn), and inconsistent fit (last column). The legends show the fitting exponent γ in p ( k ) * k − γ with respect to indegreeand outdegree distribution. https://doi.org/10.1371/journal.pone.0202202.g006 Evolutionary dynamics of cryptocurrency transaction networksPLOS ONE | https://doi.org/10.1371/journal.pone.0202202 August 17, 2018 11 / 18 hould be constant. Regarding the last month’s cryptocurrency transaction network, as can beseen from Fig 8, we find that for the Bitcoin and Ethereum networks, h k nn i is a decreasingfunction, which indicates that nodes with high degree are prone to connect with low-degreenodes, thus the networks are disassortative. For the Namecoin network, the curve is nearlyconstant, which means that the degrees of connected nodes do not rely on each other, so theNamecoin network is non-assortative. A possible reason is that for highly heterogeneous(scale-free) networks, the maximum entropy principle leads to disassortativity [32]. Thus thecause of the difference in assortativity is also the market competition.In summary, based on the analysis of the networks’ in/out-assortativity and h k nn i of in/out-degree, we find that Bitcoin and Ethereum networks are disassortative, while the Namecoinnetwork is non-assortative, which is consistent with the observation that the degree distribu-tions are heavy-tailed for the Bitcoin and Ethereum networks. Average clustering coefficient.
In order to find the evidence for a small-world network,we further compare the average clustering coefficients of networks to a random network withthe same degree sequence [33]. In cryptocurrency networks, small-world means the currencies
Fig 7. Evolution of the degree assortativity.
In the figure of Bitcoin, we magnify the y-axis of the data in the yellowbox and display it in the bottom right corner. After the initial phase, the coefficients of Bitcoin and Ethereum arenegative, and the coefficient of Namecoin converges to a certain range near 0. https://doi.org/10.1371/journal.pone.0202202.g007
Evolutionary dynamics of cryptocurrency transaction networksPLOS ONE | https://doi.org/10.1371/journal.pone.0202202 August 17, 2018 12 / 18 an be transferred between most nodes in the network by a small number of hops or steps ifthey want.As shown in Fig 9, for these three networks, in the initial phase, there is no significant dif-ference between the clustering coefficient of the cryptocurrency network and the coefficientof the random network with the same degree sequence, and even sometimes the value of therandom network is larger than the value of transaction network. In the latter phase, the threenetworks behave differently. Specifically, for Bitcoin, the clustering coefficient reaches a sta-tionary value around C (cid:25) C (cid:25) The phenomenon at the initial stage maybe results from transactions taking place between addresses belonging to a few enthusiasts who try to play the system by moving cryptocurren-cies between their addresses. The possible reason for the later phase of Bitcoin is that it is disas-sortative, which means newly added nodes tend to attach to high degree nodes, resulting thenodes tend to cluster together and form a small world. Ethereum’s abnormalities in 2016 werecaused by the network instability. And Namecoin network is non-assortative, that is, there isno correlation between pairs of linked nodes, thus the network does not exhibit this property.
Properties of the largest connected component (LCC).
Last but not least, we measurethe relative size and diameter of the LCC in the transaction network (Fig 10).For the Bitcoin network, after the initial phase, the LCC connects about 60% of the nodes in the network. For the Ethereum network, the percentage of LCC connecting nodes rises with a fluctuation and most recently connects about 40% of the nodes. And for the Namecoin net-work, the LCC connects less than 5% of the nodes in the network. Therefore, the relative sizesof the LCC of Bitcoin and Ethereum are relatively large, while the size of Namecoin network is
Fig 8. A comparison of the average degree of nodes’ neighbors and the degree of nodes.
The red line indicates thedegree of the node and the average degree of the node’s neighbors is equal. In networks without degree correlations,the h k nn i is constant. However, for Bitcoin and Ethereum, h k nn i is a decreasing function, while for Namecoin h k nn i isnot. https://doi.org/10.1371/journal.pone.0202202.g008 Evolutionary dynamics of cryptocurrency transaction networksPLOS ONE | https://doi.org/10.1371/journal.pone.0202202 August 17, 2018 13 / 18 elatively small. This result is consistent with our previous finding that the Bitcoin and Ether-eum networks are disassortative mixing in the ways that the new nodes with lower degreestend to connect to the nodes with higher degrees and vice versa, while the Namecoin networkis non-assortative as there is no correlation between nodes.The diameter of the Bitcoin LCC is around 100, indicating inefficient system transfer,which is possibly the result of anonymous users trying to hide their identity by moving theirown bitcoins as reported in [14]. The Ethereum LCC diameter is gradually increasing, possiblybecause the network is in its developing phase. The diameter of the Namecoin LCC fluctuates,which may be caused by the competition with other cryptocurrencies. Thus, during the periodof our analysis, the LCCs of these three networks do not have a sign of shrinking diameter, andthe possible reason may be the same as the reason that Bitcoin’s LCC has a larger diameter.
Discussion and conclusion
This paper analyzed the dynamic characteristics of the transaction networks of three represen-tative cryptocurrencies: Bitcoin, Ethereum, and Namecoin. We first analyzed the growth of the
Fig 9. Evolution of clustering coefficients.
If the average clustering coefficient of a network is rather higher than arandom network with the same degree sequence, the network is a small-world network. In the figure of Bitcoin, wemagnify the y-axis of the data in the red box and display it in the upper right corner. We find only Bitcoin exhibits thisfeature. https://doi.org/10.1371/journal.pone.0202202.g009
Evolutionary dynamics of cryptocurrency transaction networksPLOS ONE | https://doi.org/10.1371/journal.pone.0202202 August 17, 2018 14 / 18 ransaction networks as they reflect the relative competitiveness of the cryptocurrencies underinvestigation. By analyzing the accumulated network growth, we find similar growth patterns:they all grow with large fluctuations in the initial phase then their growth slows down and fluc-tuates within a narrow range. However, unlike most networks reported in the literature, thecryptocurrency transaction networks do not always densify over time. This phenomenon ispossibly due to the anonymity of the cryptocurrencies, where a user can create multipleaddresses to receive, store, and send cryptocurrencies.Through computing the repetition ratio, we found that the overall accumulated network isnot an applicable research object to investigate cryptocurrency properties. We thus conducteda monthly analysis on typical network measures and obtained the following insights on thesethree currencies. 1) Both Bitcoin and Ethereum networks may converge to heavy-tailed distri-bution in the long run, however, their degree distribution can only be approximately fitted by the power-law distribution. For Namecoin, its degree distribution cannot be fitted by the power-law distribution. 2) Bitcoin and Ethereum networks exhibit disassortative mixing, thatis, newly added nodes tend to connect to nodes with higher degree. 3) Only the Bitcoin net-work is a small-world network according to the analysis of clustering coefficients. 4) Bitcoin
Fig 10. The properties of the LCC.
The relative size (blue line) is the proportion of LCC nodes in all nodes, and thediameter (green line) reflects the connectivity of the LCC. Later stage, the relative sizes of the three networks are 60%,40% and 5% respectively, and the diameter of BTC is 100, while the other two are in fluctuations. https://doi.org/10.1371/journal.pone.0202202.g010
Evolutionary dynamics of cryptocurrency transaction networksPLOS ONE | https://doi.org/10.1371/journal.pone.0202202 August 17, 2018 15 / 18 nd Ethereum’s LCCs contain a relatively large proportion of nodes, however, the anonymityof Bitcoin results in a relatively large diameter.The causes of these differences might be the original design ideas and user adoption of thecryptocurrencies. Since Bitcoin is the oldest and the most dominant cryptocurrency in themarket, it was the unique choice for enthusiastic users, especially in the early days. At the sametime, price volatility is another reason to attract users to Bitcoin by treating it as an investmentalternative. Thus it is reasonable that Bitcoin network is heavy-tailed and organized as a smallworld. Ethereum, as a younger cryptocurrency, allows developers to write their own programsby replacing Bitcoin’s more restrictive language with “smart contract”, which attracts a greatdeal of user attention after its emergence. As a developing cryptocurrency, its network isheavy-tailed, but not a small world. The original design idea of Namecoin is to create a decen-tralized domain system, in which users can pay Namecoin to register and update the names oftheir domain. However, there are some other competitors in the market, say EmerCoin andNXT, which provide similar functionality. And this may be treated as the cause of its volatilitycharacteristics.
Our findings suggest that these network properties reflect the evolutionary characteristics and competitive power of different cryptocurrencies. In the future, we will relate these transac-tion network properties with the currency characteristics to guide the design of digital finan-cial products, policy regulations, and legislation.
Author Contributions
Conceptualization:
Jiaqi Liang, Linjing Li.
Data curation:
Jiaqi Liang.
Formal analysis:
Jiaqi Liang, Linjing Li.
Funding acquisition:
Daniel Zeng.
Investigation:
Jiaqi Liang, Linjing Li.
Methodology:
Jiaqi Liang.
Project administration:
Linjing Li.
Resources:
Daniel Zeng.
Software:
Jiaqi Liang.
Validation:
Jiaqi Liang, Linjing Li.
Visualization:
Jiaqi Liang.
Writing – original draft:
Jiaqi Liang.
Writing – review & editing:
Linjing Li, Daniel Zeng.
References Cimini G, Squartini T, Garlaschelli D, Gabrielli A. Systemic risk analysis on reconstructed economic andfinancial networks. Scientific Reports. 2015; 5:15758. https://doi.org/10.1038/srep15758 PMID:26507849 Diebold FX, Yılmaz K. On the network topology of variance decompositions: Measuring the connected-ness of financial firms. Journal of Econometrics. 2014; 182(1):119–134. https://doi.org/10.1016/j.jeconom.2014.04.012 Kristoufek L. What are the main drivers of the Bitcoin price? Evidence from wavelet coherence analysis.PLOS ONE. 2015; 10(4):e0123923. https://doi.org/10.1371/journal.pone.0123923 PMID: 25874694Evolutionary dynamics of cryptocurrency transaction networksPLOS ONE | https://doi.org/10.1371/journal.pone.0202202 August 17, 2018 16 / 18 . Schweitzer F, Fagiolo G, Sornette D, Vega-Redondo F, Vespignani A, White DR. Economic networks:The new challenges. Science. 2009; 325(5939):422–425. https://doi.org/10.1126/science.1173644PMID: 19628858 Narayanan A, Bonneau J, Felten E, Miller A, Goldfeder S. Bitcoin and Cryptocurrency Technologies: AComprehensive Introduction. Princeton University Press; 2016. Nakamoto S. Bitcoin: A peer-to-peer electronic cash system. 2008 [cited 2018 Jan 19]. Available from:https://bitcoin.org/bitcoin.pdf. Buterin V. A next-generation smart contract and decentralized application platform. 2014 [cited 2018Jan 19]. Available from: https://github.com/ethereum/wiki/wiki/White-Paper. Coin Market Cap [Internet]. Cryptocurrency Market Capitalizations. Available from: http://coinmarketcap.com/currencies/views/all/ Cited 19 Jan 2018.
Albert R, Baraba´si AL. Statistical mechanics of complex networks. Reviews of modern physics. 2002;74(1):47–97. https://doi.org/10.1103/RevModPhys.74.47
Boccaletti S, Latora V, Moreno Y, Chavez M, Hwang DU. Complex networks: Structure and dynamics.Physics Reports. 2006; 424(4):175–308. https://doi.org/10.1016/j.physrep.2005.10.009
Ron D, Shamir A. Quantitative Analysis of the Full Bitcoin Transaction Graph. In: Sadeghi AR, editor.Financial Cryptography and Data Security; 2013 Apr 1-5; Okinawa, Japan. Berlin: Springer; 2013.p. 6–24.
Baumann A, Fabian B, Lischke M. Exploring the Bitcoin Network. In: Proceedings of the 10th Interna-tional Conference on Web Information Systems and Technologies; 2014 Apr 3-5; Barcelona, Spain.Springer; 2014. p. 369-374.
Kalodner H, Carlsten M, Ellenbogen P, Bonneau J, Narayanan A. An empirical study of Namecoin andlessons for decentralized namespace design. In: Workshop on the Economics of Information Security;2015 Jun 22-23; Delft, Netherlands. Citeseer. 2015.
Leskovec J, Kleinberg J, Faloutsos C. Graphs over time: densification laws, shrinking diameters andpossible explanations. In: Proceedings of the eleventh ACM SIGKDD international conference onKnowledge discovery in data mining; 2005 Aug 21-24; New York, USA. ACM; 2005. p. 177-187.
Fortuna J, Holtz B, Neff J. Evolutionary Structural Analysis of the Bitcoin Network. 2013 [cited 2018 Jan19]. Available from: https://pdfs.semanticscholar.org/35b2/cac7bb85b6051f80b9eacfe292435c84a5c0.pdf.
Chang TH, Svetinovic D. Data Analysis of Digital Currency Networks: Namecoin Case Study. In: 201621st International Conference on Engineering of Complex Computer Systems (ICECCS) Nov 6-8;Dubai, United Arab Emirates. IEEE; 2016. p. 122-125.
Kondor D, Po´sfai M, Csabai I, Vattay G. Do the rich get richer? An empirical analysis of the Bitcoin trans-action network. PLOS ONE. 2014; 9(2):e86197. https://doi.org/10.1371/journal.pone.0086197 PMID:24505257
Anderson L, Holz R, Ponomarev A, Rimba P, Weber I. New kids on the block: an analysis of modernblockchains; 2016. Preprint. Available from: arXiv:1606.06530. Cited 19 Jan 2018.
Walsh C, OReilly P, Gleasure R, Feller J, Li S, Cristoforo J, et al. New kid on the block: a strategic arche-types approach to understanding the Blockchain. In: Proceedings of the International Conference onInformation Systems; 2016 Dec 11-14; Dublin, Ireland. Association for Information Systems; 2016.
Bitcoin Block Explorer—Blockchain [Internet]. Available from: https://blockchain.info/. Cited 19 Jan2018.
Ethereum BlockChain Explorer and Search [Internet]. Available: https://etherscan.io/. Cited 19 Jan2018.
Namecoin Explorer [Internet]. Available from: https://bitinfocharts.com/de/namecoin/explorer/. Cited 19Jan 2018.
Watts DJ, Strogatz SH. Collective dynamics of ‘small-world’ networks. Nature. 1998; 393(6684):440–442. https://doi.org/10.1038/30918 PMID: 9623998
Newman ME. Assortative mixing in networks. Physical review letters. 2002; 89(20):208701. https://doi.org/10.1103/PhysRevLett.89.208701 PMID: 12443515Evolutionary dynamics of cryptocurrency transaction networksPLOS ONE | https://doi.org/10.1371/journal.pone.0202202 August 17, 2018 17 / 18 Alstott J, Bullmore E, Plenz D. powerlaw: a Python package for analysis of heavy-tailed distributions.PLOS ONE. 2014; 9(1):e85777. https://doi.org/10.1371/journal.pone.0085777 PMID: 24489671
Staudt CL, Sazonovs A, Meyerhenke H. NetworKit: A tool suite for large-scale complex network analy-sis. Network Science. 2016; 4(4):508–530. https://doi.org/10.1017/nws.2016.20
Seabold S, Perktold J. Statsmodels: Econometric and statistical modeling with python. In: Proceedingsof the 9th Python in Science Conference; 2010 Jun 28-30; Austin, Texas. SciPy society Austin; 2010.
Castor A. Ethereum’s Hard Fork Drags Classic Along with it. 2016 Oct 17 [cited 2018 Jan 19]. In:BTCMANAGER.COM [Internet]. Available from: https://btcmanager.com/imminent-hard-fork-aligns-with-ethereum-classic-principles/.
Clauset A, Shalizi CR, Newman ME. Power-law distributions in empirical data. SIAM review. 2009;51(4):661–703. https://doi.org/10.1137/070710111
Johnson S, Torres JJ, Marro J, Munoz MA. Entropic origin of disassortativity in complex networks.Physical review letters. 2010; 104(10):108702. https://doi.org/10.1103/PhysRevLett.104.108702 PMID:2036645833.