Cryptocurrencies as an Asset Class: An Empirical Assessment ∗
Daniele Bianchi†
First draft: September 2017.
This draft: December 5, 2017
Abstract It has long been debated whether cryptocurrencies are just a passing fad, a disruptive innovation, or simply share features with standard securities. In this paper, I use a novel data set of prices, traded volumes, and market capitalization for a large set of cryptocurrencies to empirically investigate both their relationship with standard asset classes and the main driving factors behind market activity. The main empirical results suggest that there is a significant relationship between returns on cryptocurrencies and commodities such as gold and energy. Also, while volatility correlates with traded volume, the latter is primarily driven by past returns and by a short-lived effect of aggregate market uncertainty. This is consistent with existing theoretical models in which trading activity is primarily driven by investors’ sentiment. Finally, impulse-response functions from a panel Vector Autoregressive (VAR) model show that macroeconomic factors do not significantly drive trading activity in cryptocurrency markets. Keywords: Cryptocurrencies, Bitcoin, Blockchain, Financial Markets, Investments JEL codes: G11, G12, G15, G19
∗
I am grateful to RoboAdvisorCoin for providing the data and in particular Marco Querini for guiding me through the dataset. I thank Matthias Buechner for helpful comments and suggestions. † Warwick Business School, University of Warwick, Coventry, UK.
[email protected]
1
1
Introduction
The explosive growth of peer-to-peer exchanges and the Blockchain technology has spurred the emergence of cryptocurrencies as a significant component of financial markets.1 At the time of writing, there are about a thousand actively quoted cryptocurrencies; the most wellknown is Bitcoin, which has been classed as a commodity in the U.S., therefore covered by the Commodity Exchange Act along with gold and oil according to the Commodity Futures Trading Commission (CFTC).2 It is believed that cryptocurrencies could potentially disrupt financial services and central banks, therefore posing a risk for the stability of prices and the financial system.3 Figure 1 displays the massive upward trend both in terms of market capitalization (left panel) and average traded volume (right panel) of the cryptocurrency market. As a sign of the accelerating professionalization of investments in bitcoins, two of the world’s largest futures exchanges, Chicago Mercantile Exchange (CME) group and the Chicago Board Options Exchange (CBOE), have been given a regulatory green light to list futures contracts on Bitcoin, a significant step in giving investors a mainstream instrument to trade.4 Although the institutionalization of cryptocurrencies as an asset class does not guarantee their growth is sustainable in the medium-to-long term, it certainly shows that the underlying distributed ledger technology offer potential benefits that go far beyond increasing payments efficiency and promoting financial inclusion and transparency. 1
A Blockchain is an open and distributed ledger that records all transactions, called blocks, which are sequentially linked and secured in a verifiable and permanent way using cryptography. Each block contains a link to a previous block, i.e. a hash pointer, a time-stamp to identify the timing of the transaction, and the transaction data. 2 Bitcoin has been introduced in 2008 in a white paper by Satoshi Nakamoto, a still unverified entity, who first introduced Bitcoin as a decentralized, peer-to-peer electronic cash system which does not need central authorities or trusted third-party to operate. 3 The European Central Bank explicitly addressed the raise of virtual currency schemes and the subsequent risks for financial stability and the reputation of central banks in a series of discussion papers in 2012 and 2015. Similarly, the International Monetary Fund in a discussion paper in 2017 proposed an economic framework for thinking through the channels by which digital currencies might respond to consumer needs for trust, security, privacy, and better services, change the competitive landscape, and affect regulation. 4 Throughout the paper I use the term cryptocurrencies, digital currencies, cryptos, and coins interchangeably.
2
Many of the available cryptocurrencies are initially distributed to early investors through a crowd sale similar to an IPO, the so-called Initial Coin Offering (ICO). In its general form, an ICO represents an unregulated crowd funding whereby businesses and start-ups bypass the rigorous and regulated capital-raising process required by venture capitalists or banks. Thus, cryptocurrencies share more similarities with an equity investment in a company than an investment in a traditional, i.e. fiat, currency. For instance, as its advocates claim, holding Bitcoin can be ultimately seen as an investment in the Blockchain technology rather than a simple speculation. In this paper, I use a novel data set which consists of a large panel of prices, traded volumes, and market capitalization on actively quoted cryptocurrencies to empirically investigate their relationship with standard asset classes, e.g. equity, bond, etc., in an attempt to shed some light on the underlying investment properties. Also, I make use of a panel Granger causality test and a panel Vector Autoregressive (VAR) model to investigate the driving factors behind market activity, both over the short and the medium horizon. I report a number of new findings. First, I show that there is a significant relationship between the returns on cryptocurrencies and both Gold and Energy commodities, i.e. crude oil and natural gas. Such relationship holds controlling for both lagged returns and past changes in average traded volumes. The positive correlation with gold supports the prevailing view that posits that cryptocurrencies, and Bitcoin in particular, are a “store of value” since their supply is limited and they are arguably uncorrelated with a given country’s economy. For instance, anecdotal evidence shows that the price of a Bitcoin climbed as much as 10% on Zimbabwe’s Golix exchange on November 15th 2017, after the country’s armed forces seized power amid a shortage of hard currency. By the same token, the intentional devaluation of the Chinese Yuan and the government policy of capital controls spurred increasing demand in cryptocurrencies by Chinese investors looking for a safe and anonymous way to send their money off-shore.5 5
The Wall Street Journal featured in its November 5th headlines, “Chinese Investors Buying Up Bitcoin
3
As a case in point, top panel of Figure 3 shows the relative traded volume expressed in Bitcoin against major fiat currencies. Until early 2017 most of bitcoins were bought by using Chinese Yuan (CNY), which arguably means that trading activity on active exchanges was primarily coming from Chinese investors. Yet, the intervention of the Chinese government to crack down ICOs and exchanges had significantly shrunk trading activity of bitcoins against CNY. As far as volatility is concerned, the results show that there are no spillover effects between cryptocurrencies and standard asset classes. Once conditioning on past traded volume and a linear time trend, none of the other asset classes volatility is significantly correlated with the realized volatility of cryptocurrencies. Nevertheless, the empirical analysis provides evidence that a key driving factor for volatility is indeed the average traded volume. This is consistent with existing economic theory which suggests that market activity should be tightly related to volatility as returns movements are primarily due to the arrival of new information and the process that incorporates both public and private information in equilibrium prices (see, e.g., Glosten and Milgrom 1985 and Kyle 1985). Second, I show that traded volume is primarily driven by past returns and aggregate market uncertainty, as proxied by the VIX index. However, the impulse-response functions from a panel Vector Autoregressive (VAR) demonstrate that while past performance significantly lead traded volume up to few weeks ahead, the changes in the VIX index have a rather short-lived effect. The causal relationship of past returns on traded volume is confirmed by a panel Granger causality test as proposed by Dumitrescu and Hurlin (2012); the results show that returns Granger-cause traded volume and not the opposite, i.e. no feedback effect. These results are consistent with existing empirical evidence on equity markets that shows that there is a positive and significant correlation between stock returns and trading as Yuan Falls”. The Financial Times after the crackdown of Chinese authorities of public exchanges dealing with cryptocurrencies ran a similar story in November 7th arguing that “Bitcoin proves hard to kill in China”, as investor looking to escape capital controls bypassed the public controls by operating in peer-topeer over-the-counter markets.
4
volume (see, e.g. Karpoff 1987, Gallant et al. 1992, Schwert 1989, Campbell et al. 1993 and Llorente et al. 2002). Other macroeconomic factors such as inflation expectations, the yield curve, and real exchange rates do not have a significant role, consistent with earlier findings by Yermack (2013) which shows that Bitcoin is not driven by major macroeconomic events. Third, I investigate the existence of momentum effects in cryptocurrencies both in the time series and in the cross-section. Momentum implies that cumulative past returns have some predictability power for future realized returns one or several steps ahead. Similarly, a trading strategy which goes long to past winners and short on past losers should generate sizeable and significant returns. The empirical results are somewhat mixed; although there is evidence of a short-lived momentum effect in the time series, the returns from a long-short portfolio on past returns do not generate statistically significant returns. The literature on Blockchain applications for financial markets and cryptocurrencies in particular is very thin. Earlier papers are from the field of computer science and are concentrated on outlining the technical aspects and computational issues of cryptocurrencies and the Blockchain. The existing studies in economics related to Bitcoin and cryptocurrencies are mostly focused on the operational features of cryptocurrencies, such as the likelihood of exchanges default (e.g. Moore and Christin 2013), the possibility of mining manipulation (e.g. Eyal and Sirer 2014), the reconstruction of transaction networks, (e.g. Kondor et al. 2014), the efficiency of a decentralized public ledger for conducting financial transactions (e.g. Evans 2014 and Dwyer 2015), and the implications for central banks and monetary policy (e.g., Bordo and Levin 2017). Few theoretical models and discussions have also been proposed to study the equilibrium pricing mechanism and the development of a decentralized market place. Chiu and Koeppl (2017) developed a general equilibrium monetary model to investigate the optimal design of a cryptocurrency system based on a Blockchain. They show that a standard “proofof-work” protocol generates a welfare loss of 1.4% of consumption. Such loss can be lowered
5
substantially by financing mining rewards through money supply growth rather than by increasing transaction fees. Similarly, Huberman et al. (2017) investigate the fees structure and show it might be the outcome of an equilibrium of a congestion queuing game derived from the throughput limited by the infrastructure. Finally, Cong et al. (2017) analyse the informational gains of decentralization and the implications for both industrial organization and the landscape of industry competition more generally. However, the number of empirical works investigating the properties of cryptocurrencies as an asset class is even smaller. Yermack (2013) argues that Bitcoin cannot be considered a currency. He shows that the Bitcoin is uncorrelated with the majority of fiat currencies and is much more volatile, therefore being of limited usefulness for risk management purposes and diversification. On the other hand, Dyhrberg (2016) shows that the behavior of Bitcoin volatility has several similarities to gold and the dollar indicating hedging capabilities and advantages as a medium of exchange. Differently from them, I exploit the cross-sectional heterogeneity of a large panel of cryptocurrencies to identify average aggregate linkages with other standard global asset classes, as well as investigate the driving factors of traded volume. The structure of the paper is the following. Section 2 provides a brief review of cryptocurrencies and the ICOs scheme. Section 3 describes the data. Section 4 investigates the aggregate the aggregate relationship in both returns and volatility between cryptocurrencies and standard asset classes, as well as investigates momentum effects. Section 5 focuses on the interplay between traded volume, past returns and macroeconomic activity. Section 6 concludes.
2
A Primer on Cryptocurrencies
While the focus of this paper is not to understand the protocols behind cryptocurrencies in details, it is useful to briefly review their key features as fundamentals and price processes are
6
possibly influenced by the underlying technology and the transactions validation procedure. Just like any commodity-backed currency or fiat money, cryptocurrencies have value by virtue of the ability to use them to purchase goods and services. However, trust in a given cryptocurrency is not based on its intrinsic value or on the belief of a central bank or government solvency, but rather is established by having a competition to the right of updating the existing transactions record through cryptography. Thus, there is no centralized authority which controls the amount of coins in circulation and prohibits holders to freely trade those coins. In the absence of trusted third-party and intermediaries, cryptocurrencies rely on a distributed verification, updating and storage of a record of transactions, called blocks, that have been conducted between the users. These blocks are sequentially linked and secured in a verifiable and permanent way by forming an open and distributed ledger, i.e., the Blockchain, such that network participants can publicly verify the balance a given user owns while the identity of the sender and receiver remains anonymous. The Blockchain requires that transaction taking place in different nodes of the network must be dynamically consistent, that is, if a user attempt to reverse a past transaction he needs to solve for an alternative Blockchain consistent with his proposal. Thus, the introduction of a confirmation lag before completing each transaction makes double-spending, i.e., the attempt to reverse back a transaction while keeping the bought good or service, very costly. This consensus protocol implies that the “longest” history will be accepted as the valid public record such that the responsibility to verifying the validity of transactions is left to the entire network without the need of central authorities such as clearing houses or banks. The competition to update the Blockchain can take various forms. For most currencies, this is done through a process called “mining”. Miners are transaction validators which compete to solve a computationally costly problem. The winner of this mining process has
7
the right to update the record and is rewarded by newly minted coins. Therefore, mining often represents the only way to create new coins in cryptocurrency schemes. The rewarding scheme for miners is scheduled by the protocol.6 Bitcoin represents the best known and most successful cryptocurrency circulating so far. The Bitcoin system is based on a peer-to-peer network similar to BitTorrent, the famous protocol for sharing files over the internet. The exchange rate is not pegged to any fiat money and is primarily determined by demand and supply in the market. At the time of writing, there are more than twenty active exchange platforms for buying bitcoins, five of them with more than $200mln in assets.7 Other financial institutions such as hedge funds, venture capital, financial advisory, and derivative exchanges are also increasingly engaged in cryptocurrency trading. As of November 2017 there are now more than 120 investment funds solely focusing on cryptocurrencies with about $2.3bn of total assets under management.8 The aggregate supply is not controlled by a single entity but is governed by an algorithm and anyone can access the source code via the internet on Github.9 Although still dominant in terms of market capitalization and traded volume, Bitcoin could be seen as a special sub-type in the full spectrum of digital currencies. There are currently more than a thousand of cryptocurrencies and the number is increasing steadily. The availability of the Blockchain source code has led to a proliferation of hundreds of alternative cryptocurrencies, which are commonly referred to as “altcoins”. On the one hand, some of these alternative coins have been introduced to solve the fundamental issues of the original Bitcoin platform, such as reducing the computational costs, increasing the 6
For instance, the reward for mining Bitcoin was initially set at 50 bitcoins per block mined. The protocol specifies that every 210,000 blocks mined that reward would be cut in half, until eventually becomes zero after the supply reaches the planned limit of 21mln coins. 7 Active exchanges which have more than $200mln of assets are Poloniex, Kraken, GDAX, Bittrex, Bitfinex, BitMEX, BTC-e, and Bitstamp, all of them trading not only Bitcoin but also other cryptocurrencies against major standard currencies such as U.S. dollar, the Euro and the British Pound. 8 CNBC on October 27th 2017 run a story according to which “While several leading Wall Street banking executives remain skeptical about Bitcoin, more seasoned money managers are moving into digital assets management.”. They reported estimates from financial research firm Autonomous Next according to which more than 90 funds focused uniquely on bitcoins have been launched since the beginning of 2017. 9 https://github.com/bitcoin/bitcoin
8
number of transactions that can be processed for a time stamp, increasing the block size, or increasing the anonymity of the public ledger. On the other hand, others have adopted different and innovative technologies: the main example is Ethereum which is based on a decentralized Turing-complete virtual machine that features smart contracts functionality. As a whole, cryptocurrencies can be grouped based on three main categories: the validating system, the underlying algorithm, and the supply of coin. The validating system concerns the method used to validate the transactions. The very first cryptocurrencies, (see, e.g., Bitcoin, Litecoin, Dogecoin) are based on a “proof-of-work” (POW) protocol, which represents a computationally expensive computer calculation that needs to be performed in order to validate transactions by means of hashing.10 As an alternative validating scheme, several other cryptocurrencies (see, e.g., BitShares, Dash, Factom) are based on a “proof-ofstake” (POS) protocol, which unlike POW automatically takes into account the circulating number of coins distributed in the network. More specifically, POS carry out a process which is called “forging” instead of “mining”, whereby all users known ex-ante the node of the network which will validate the transaction and update the public ledger. This increases the efficiency of the validating process and prevents the so-called “51% attack”, whereby those controlling more than 51% of the computer power of the network can dictate the general functioning of the scheme. The underlying algorithm represents the second main characteristics which differentiate cryptocurrencies. Two main mathematical procedure for calculating and processing the data can currently be identified: SHA-256 (e.g., Bitcoin) and Scrypt (e.g., Litecoin), where the latter can be thought of as an extension of the former. While SHA-256 requires a special equipment for processing transactions validation, Scrypt allows “miners” to perform 10
In cryptography, hashing is an application of a hash function which encrypts an input in a predetermined way. From a technical point of view, the mining process is an operation of inverse hashing, a problem that cannot be solved in other ways than by brute force, so the cryptographic hash algorithm of block data results in less than a given threshold. The latter is called “difficulty” and determines the competitive nature of mining. In this respect, the more computing power is added to the network, the higher becomes the number of calculations needed to solve the encryption and create a new block.
9
their activities on a regular hardware. Notably, the fast development of the cryptocurrency market makes quite possible that these algorithms will be replaced shortly by more advanced procedures. Finally, cryptocurrencies can be differentiated on the basis of the total supply of coins. In most cases the final supply is predetermined by the algorithm and the inflation rate decreases over time (e.g., Bitcoin, Litecoin, Dogecoin, etc). However, there are cases where the supply of coins is flexible and designed to maintain a fixed predetermined inflation rate.
2.1
Initial Coin Offering
New cryptocurrencies can be issued by developers to fund projects or businesses through an Initial Coin Offering (ICO) whereby people can invest in a project simply buying part of the cryptocurrencies token in advance based on a white paper or business plan made public for investors to evaluate. Although similar to an Initial Public Offering (IPO) in spirit, ICOs are unregulated means used to bypass the rigorous capital-raising process required by venture capitalists or financial intermediaries. The purpose is to raise money to pay the project developers, initially distribute coins fairly amongst a large group, and entice early adopters by offering a new cryptocurrency at a perceived discount well-below its potential future value. Thus, these sales share many similarities with traditional securities offerings where investors purchase stock as a bet on the future success of a company or a project. However, to simply assume that cryptocurrencies are similar to securities just because they can be sold through crowd-sales, fails to understand the many differences that belie between the two. One key distinction is that there is typically no traditional management structure behind a cryptocurrency. Most platforms are based on decentralized applications and do not depend on management decisions but are based on a defined and publicly known set of protocols. Thus, unlike when purchasing a stock, investing in a cryptocurrency is not a bet on the management’s ability to generate profits, but rather a bet on how useful
10
and popular a particular network application will become. Such distinction is not always clear cut though; in fact, investors buying a newly created cryptocurrency can also be seen as making a bet on a development team’s ability to make a particular platform or project financially successful. A second main distinction is information asymmetry; standard asset classes such as equity, bonds and real estate are primarily based on the principle that the full disclosure of financial fundamentals helps to close the information gap between buyers and sellers. For instance, companies planning to engage in an IPO are required to disclose information about their operations and their financial health. Within the universe of cryptocurrencies the concept of asymmetric information takes a radically different meaning. Most popular decentralized applications are open source so that investors has full access to the operations. This is not to say that there is no asymmetric information, but rather that investors are likely most interested in learning how the protocol behind that application works rather than understanding how the initial development team operates. Therefore investors may receive more protection from a technology audit by a computer scientist of the proposed protocol than from a financial audit by an accountant of the developer’s bank accounts. In this respect, the existing paradigm imposed by standard regulated securities is of little benefit to cryptocurrency investors. The third key distinction between cryptocurrencies and standard securities is that cryptos do not grant their holders any traditional legal rights. Holding a newly issued coin does not legally entitle one to share in profits realized by the project or to vote on key decisions about that project’s future. This makes ICOs almost like crowd-funding whereby people sponsor a project without having legal rights on intervening in the project development itself. In exceptional cases rights might be granted by computer code and are not backed by any actual legal obligation, and it is therefore unclear whether courts will, or even can, enforce any rights.
11
3
Data
The data sample refers to daily prices, traded volume and market capitalization in U.S. dollars for 1251 cryptocurrencies which are quoted from April 2016 to September 2017. Daily data are converted in weekly observations to mitigate the noise that characterize the highly volatile daily returns while keeping a sizable number of observations. The final sample for the main empirical analysis is obtained by pre-filtering cryptocurrencies based on their market capitalization and the average weekly traded volume. In this respect, I focus on large-cap currencies with a substantial amount of trading volume to reduce liquidity concerns. More specifically, I eliminate those currencies with an average traded volume which is below the sample median, as well as I eliminate those which have a market capitalization below the top 5th percentile of the cross-sectional distribution at the end of the sample. After the filters the final sample is N = 14 cryptocurrencies which accounts for more than 85% of the total market capitalization as of September 2017. A full description of the currencies used in the main empirical analysis is provided in Appendix A. Although the sample is relatively short, it is fairly representative of the development of the cryptocurrencies market. As a case in point, Figure 2 reports both the market capitalization (left panel) and the aggregate average traded volume (right panel) in billions of U.S. dollars. It is evident that both the market size and the amount of traded volume has not significantly developed until early 2017. [Insert Figure 1 about here] Notably, the sample includes different market phases, that is, the dramatic increase in valuations which made the headlines for the whole cryptocurrency market, as well as the substantial market drop that occurred in the summer of 2017, where almost every cryptocurrency has suffered massive losses. On the one hand, this drop can be partly explained by the looms of the “hard-fork” of the original Bitcoin which generated Bitcoin Cash in early August of 2017. On the other hand, the unregulated nature of the ICOs ecosystem made most of ICOs 12
starting to cash out after the massive boom in valuations generating a temporary sell-off which decreased prices.11 Similarly, Figure 2 shows that the structure of the cryptocurrency market has not really developed until late 2016, which implies that a significant fraction of currencies were neither traded nor liquid until recently. Left panel reports the capitalization of Bitcoin relative to the overall market, and right panel its relative average traded volume.
[Insert Figure 2 about here]
Until early 2017 Bitcoin has been by far the dominant player representing about 90% of the whole market on itself, then decline to a bit more than a half towards the end of the sample. Similarly, transactions on Bitcoin went from representing almost 80% of the overall traded volume in early 2016 to about 50% at the end of the sample. This means that by increasing the sample size going backward effectively means that one investigates almost exclusively Bitcoin’s volume-returns linkages. As a whole, Figure 2 illustrates that the sample under investigation is fairly representative of the development of the cryptocurrency market. By the same token, the market structure of cryptocurrencies only developed since late 2016. Figure 3 reports the geographical diversity of the Bitcoin investors (top panel) and the market concentration of Bitcoin exchanges (bottom panel). Top panel shows that Chinese investors have been dominating the market activity in early stages. In fact, the Bitcoin equivalent of traded volume in Chinese Yuan made up almost the entire amount of Bitcoin trading. Such a proportion steadily decreased only after the initial crackdown of the Chinese government on the ICOs market and the official shut-down of the exchanges that occurred in September 2017. [Insert Figure 3 about here] 11
Fortune magazine featured in its July 17 2017 arguing that part of the price collapse was the “recent spate of ICOs in which founders Blockchain companies have raised huge amounts of Bitcoin and Ethereum, and then dumped some of their windfall on the market.” (see http://fortune.com/2017/07/17/bitcoin-crash/).
13
Similarly, bottom panel of Figure 3 shows that the trading activity before the end of 2016 was essentially concentrated around three main exchanges, which more than double towards the end of the sample. In this respect, the market micro-structure is in full development, as shown by a higher number of equally relevant exchange platforms at the end of the sample. Such increasing number of exchange platforms could be possibly the results of an endogenous growth process led by increasing profits due to the compounding effect of increasing prices and market activity.
3.1
Descriptive Statistics
Unless otherwise specified, the main empirical results are based on the fourteen cryptocurrencies outlined above and described in Appendix A. Table 1 reports a set of weekly descriptive statistics for realized returns, market capitalization and average traded volume for these currencies. [Insert Table 1 about here] It is evident that cryptocurrencies are not comparable to any asset class in terms of their sample moments; in a nutshell cryptocurrencies consistently show high expected returns (in the range of 2% to 6% on a weekly basis), even higher volatility (which ranges from 9.7 to 43.73, again on a weekly basis), and a much larger probability of experiencing a significant gain rather than a large loss, as indicated by massive kurtosis and large positive skewness. As pointed out by Yermack (2013) the high volatility that characterize Bitcoin somewhat limits its usefulness in terms of risk diversification and hedging. Also, high volatility contradicts the idea that cryptocurrencies represent a “store of value”, meaning they can be saved and retrieved in the future without having lost much value in real terms. In fact, abrupt price fluctuations, captured by the sample standard deviation, are inconsistent with the idea that cryptocurrencies are a savings instrument for the short-term, let alone for the medium-tolong term. 14
Volatility in cryptocurrencies can be driven by many different factors. Regulatory interventions, technological innovations and security breaches are only few of the factors that possibly influence price fluctuations. Ultimately, the value of a coin reflects investor confidence in the protocol design and the prospects of being an actual method of payment. In this respect, price instability is largely driven by news and the perceptions of the intrinsic value of the cryptocurrency as a increasingly adopted method of payment for given goods or services. These factors generate demand pressure, which in turn generate wild price fluctuations given that, unlike fiat money, cryptocurrencies are in small and finite supply and therefore are highly dependent on demand shocks. Panel B of Table 1 reports the sample mean, median, and standard deviation of the market capitalization in millions of U.S. dollars. The fourth row reports the ratio between the last and the first observations in the sample. The market capitalization increased about a hundred times on average across cryptocurrencies over the sample period. This is the spectacular growth which led most commentators to claim that the market of digital currencies is one of the biggest bubbles in the history of financial markets.12 The end-of-sample market capitalization of Bitcoin and Ethereum alone is about $65bn and $28bn, respectively, and keeps increasing. This is the size of a large cap stock in the U.S. Nevertheless, the current market capitalization on average across cryptocurrencies is about $850mln, which is more the size of a small cap stock indeed. This shows that although the market for cryptos is steadily increasing in size, it is still in its infancy. Panel C of Table 1 reports the sample mean, median, and standard deviation of the weekly average traded volume in millions of U.S. dollars. As above, the fourth row reports the ratio between the last and the first observations in the sample. Unsurprisingly, the dollar amount of trading activity on cryptocurrencies massively increased throughout the sample. 12
On October 9th 2017, The Guardian featured an article by Kenneth Rogoff in which ha argued that the price of Bitcoin is completely detached to fundamentals and will collapse under regulators’ pressure (see https://www.theguardian.com/technology/news-blog/2017/oct/09/bitcoin-price-bubblegovernment-cryptocurrency).
15
The conventional wisdom posits that cryptocurrencies are not necessarily correlated with other asset classes and therefore represent a good instrument for investors looking for hedging macroeconomic shocks or spillover effects from the returns in classic financial instruments. Figure 4 somewhat contradict this view by showing the time-varying correlation between a value-weighted market index for cryptocurrencies and other asset classes. The market index is constructed by using those currencies with a market capitalization in the top 20th percentile of the cross-sectional distribution at the end of the sample and for which the average traded volume in U.S. dollars is non-zero. The sample contains 42 cryptocurrencies which accounts for more than 95% of the total market capitalization at the end of the sample.
[Insert Figure 4 about here]
Top-left panel shows the correlation between the weekly returns of the value-weighted portfolio of cryptos and the returns of the FTSE Global All-Cap value-weighted index, which summarizes the performance of around 7,400 large, mid and small cap stocks and cover both developed and emerging markets.13 Although small in magnitude, the correlation against equity is negative for a large fraction of the sample. One possible explanations lies in a standard “flight-to-quality” phenomenon whereby as economies built with fiat currencies show signs of strength or weakness, investors may allocate more or less of their assets into cryptocurrencies which, by construction, are not directly determined by macroeconomic aggregates. In other words, a downward trend in financial markets may increase the investors’ incentives to lock in resources in cryptocurrencies, generating a positive demand shock which increases prices in equilibrium under fixed supply. Top-right panel of Figure 4 reports the correlation between the value-weighted portfolio of cryptocurrencies and the Global Broad Market index provided by BofA-Merrill Lynch, 13
The choice of a global equity index, instead of, say, the S&P500, is consistent with the fact that crypocurrencies are traded globally regardless if they are mined in China, Europe or the U.S. Thus, a country-specific benchmark is potentially misleading.
16
which tracks the performance of investment grade public debt issued in the major markets, including “global” bonds. Consistent with economic theory, the dynamics of correlation with bonds is practically the reversed of equities. Mid panels report the conditional correlation between cryptocurrencies and real estate (left panel) and equity volatility (right panel). The former is calculated as the returns on the MSCI World REITs index which represents a free float-adjusted value-weighted index that captures both large and mid cap representation across more than 20 developed markets around the world. Weekly returns on equity volatility are computed from the S&P 500 VIX Short-Term Futures index which replicate a constant one-month rolling long position in first and second month VIX futures contracts. There is no clear trend in the correlation between cryptocurrencies and real estate markets. Notably, the shape of the correlation between cryptocurrencies and equity volatility is similar to the correlation with bond returns. Again, this may be justified based on the fact that increasing equity volatility can be associated with weakening aggregate financial conditions and increasing market uncertainty, and ultimately with a “flight-to-safety” phenomenon. Similarly, investments in gold can be driven by investors looking for a safe bet during market turmoil and downturns. Although it is not risk-adjusted, the positive correlation between cryptocurrencies and returns on gold – measured as the S&P GSCI Gold index which tracks the COMEX gold futures – reported in the right panel of Figure 4 somewhat confirms the view expressed in the media that championed the idea that cryptocurrencies might be helpful for diversification purposes. Notably, the bottom-left panel shows that returns on Energy commodities, proxied by the S&P GSCI Energy index – which includes crude oil and supporting contracts, as well as natural gas – are negatively correlated with returns on cryptocurrencies, similar to equity. On comment is in order. Although there is some correlation in the returns between cryptocurrencies and gold, this does not necessarily means that they are actually similar in spirit. As a matter of fact, although significant, the size of the conditional correlation is almost negligible. 17
4
Empirical Analysis
Figure 4 tells a story which is consistent with the idea that cryptocurrencies might be mildly correlated with financial markets, somewhat contradicting the conventional wisdom that posits that cryptos are a global hedging instrument. Although indicative, this evidence is speculative at best and therefore should be taken with a grain of salt. In this section, I investigate average, i.e., population, contemporaneous linkages by using a random-coefficient panel regression framework which explicitly accounts for both a unit specific average response and the cross-sectional heterogeneity in the regression betas (see Swamy 1970). A randomcoefficient specification generalizes a common panel regression with individual random-effect, or random intercept, whereby the sensitivity of the response variable to the explanatory is restricted to be the same across groups. Let xt = (x1t , · · · , xKt )′ denote the K-dimensional vector of explanatory variables and yit the response variable for the ith group at time t. In the random-coefficient regression model both the intercept and the marginal effect of the explanatory variable vary across groups, i.e., individual cryptocurrencies,
yit = (α + α0i ) + (β + β 0i )′ xt + ǫit ,
i = 1, ..., N,
t = 1, ..., T,
(1)
where ǫit ∼ N (0, σ 2 ), α0i ∼ N (0, τ 2 ) and β 0i ∼ N (0, Σ) are the group-specific regression parameters which describe how the intercept and the betas deviate from the population mean values α and β, respectively. The covariance structure of the stochastic individual effects is left unstructured such that Cov (αi0 , β ′i0 ) = δ ′ .14 The term (β + β 0i )′ represents a combination of the systematic effect of the explanatory variables on the response, β, and 14 The random-coefficient model in (1) is estimated through a Generalized Least Squares (GLS) estimator which is defined as a weighted-average of the group-specific Ordinary Least Squares (OLS) estimates,
ˆ α ˆ, β
′
=
N X
Wi bi ,
i=1
where bi are the OLS group-specific estimates of the intercept and slope parameters, and the weights are
18
a random component summarizing how effects vary across groups β 0i with the variance Σ describing the nature of such variation. One comment is in order. In a standard random-effect model, it is only the overall level of the response captured by the intercept that could randomly vary across groups. However, Pesaran and Smith (1995) pointed out that under a wrong hypothesis of constant slope parameters across groups the estimates might be biased, that is, hypothesis testing results could be misleading. As a matter of fact, the estimate of the slope parameters obtained in an homogeneous model tend to converge to the cross-sectional average of the true slope coefficients; if such average is close to zero one could wrongly conclude that there is no relationship between the response and the explanatory variables. As a robustness, Appendix B also reports the results for a restricted random-intercept panel regression specification whereby the slope parameters are constant across groups. Table 2 reports the Generalized Least Squares (GLS) estimates with robust standard errors based on 5,000 bootstrap iterations reported in parenthesis. The response variable is the returns on each cryptocurrency outlined in Section 3 and detailed in Appendix A. The set of explanatory variables is defined by the returns on the FTSE Global All-Cap index (Equity), the BofA-Merrill Lynch Global Broad Market index (Bond), the MSCI World REITs index (Real Estate), the S&P GSCI Energy index (Energy), the S&P GSCI Gold index (Gold), the S&P 500 VIX Short-Term Futures index (Options), and the log-change of the nominal effective exchange rate of the U.S. dollar (FX), which represents a trade-weighted average of nominal bilateral rates between the USD and a basket of foreign currencies. The lagged returns of both the response and the explanatory variables, as well as the lagged values of defined as Wi =
N X i=1
Ω + Vˆi
−1
!−1
(Ω + Vi ) ,
where
Ω=
and Vˆi the covariance matrix of the OLS betas estimates (see Swamy 1970).
19
τ2 δ
δ′ Σ
the average weekly traded volume, are included as additional control variables.
[Insert Table 2 about here]
The results indicate that only returns on commodities such as Energy and Gold have some systematic correlation with cryptocurrencies. Column eight provides evidence that spillover effects in returns of both Gold and Energy remain significant after controlling for both contemporaneous and lagged returns on other asset classes. The null hypothesis that all the slope parameters are constant across groups, i.e. H0 : βˆ1 = βˆ1 = · · · = βˆN , is rejected for the complete regression specification with a chi-square statistics equal to χ2(K(N −1)) = 154.9 (p-value = 0.000). Although the literature is limited, the significant relationship between returns on cryptocurrencies and commodities is consistent with some of the existing empirical evidence. By using a classic asymmetric GARCH modeling framework, Dyhrberg (2016) shows that there are several similarities between Bitcoin and Gold indicating the former may be useful for risk averse investors to hedge their exposure on other asset classes. Again, Dyhrberg (2016), reinforced the idea that Bitcoin can be used as an hedge against negative shocks in the stock market. The reason why Gold and cryptos may share a “value storage” feature is intuitive: they both have a limited supply growth and equilibrium prices primarily depend on aggregate demand. Over the past half century, new gold mining supply has added anywhere from 1% to 2% to the existing stock of previously mined gold. In the same spirit, the supply inflation rate for most cryptocurrencies is steadily decreasing. In both cases the mining output is scarce, such that market prices largely depend on demand pressure. It is more difficult to rationalize the negative and significant relationship between Energy markets and returns on cryptocurrencies. One possible explanation lies on the fact that crude oil is positively correlated with global macroeconomic conditions. (see, e.g. Hamilton 2009 20
and Kilian 2009). In particular, increasing Energy prices, due to increasing demand, are normally associated with improving aggregate economic conditions. As economies shows signs of strength, the marginal benefit of lock in resources in a volatile investment such as cryptocurrencies might be less stringent. Despite the fact that there are similarities between Gold and cryptocurrencies – they are both in limited supply, their prices can be highly volatile and each can be seen as an alternative investment for those lacking faith in fiat currency and monetary policy – by no means they are based on the same premises. Investing in cryptos do not represent a safe bet per se, but rather it is a bet on the underlying Blockchain technology or project.
4.1
Volatility and Volume
I now investigate volatility spillover effects between cryptocurrencies and standard asset classes. As in Section 4.1, I make use of a random-coefficient panel regression framework as outlined in Equation (1). The realized volatility for both cryptocurrencies and other asset classes is calculated as the square root of the sum of daily squared demeaned returns. I include a linear time trend, the lagged realized volatility, and the average traded volume as additional control variables. Table 3 reports the GLS estimates with robust bootstrap standard errors reported in parenthesis.
[Insert Table 3 about here]
Perhaps surprisingly, the empirical results show that risk in cryptocurrency markets does not correlate with risk in other asset classes. Yet, market activity, as proxied by the average traded volume, significantly correlates with the volatility of cryptocurrencies. This is consistent with existing evidence in equity markets. Economic theory suggests that variables such as trading volume, the number of transactions, or market liquidity should be tightly related to volatility as returns movements are primarily due to the arrival of new information and 21
the process that incorporates both public and private information in equilibrium prices (see, e.g., Glosten and Milgrom 1985 and Kyle 1985). Similarly, Kim and Verrecchia (1991) show that the relationship between price changes and trading volume arises because the size of the trades is positively related to the quality of information in the presence of competition and asymmetric information among traders. Finally, Foster and Viswanathan (1993) and Holden and Subrahmanyam (1992) show that the same type of positive correlation between volume and volatility can be generated in strategic models, although might be attenuated by strategic playing of monopolist informed traders or market makers. On the empirical side, a significant part of the literature has documented a strong correlation between traded volume and return volatility in other asset classes. Among others, Gallant et al. (1992) and Andersen (1996) show that there exists a positive and significant correlation between market activity and returns volatility. Given the evidence of Table 3, I now further investigate any possible causal relationship between volume and volatility on the basis of a Granger (1988)’s causality, a statistical notion of causality based on the relative forecasting power of two series. More specifically, I test for causality and feedback effects between traded volume and volatility by implementing a panel Granger causality test as proposed by Dumitrescu and Hurlin (2012). This allows to test feedback effects and causality within a panel data framework whereby the cross-sectional heterogeneity of the causal relationships between response and explanatory variables is explicitly acknowledged, which in turn improves the power of tests (see, e.g., Holtz-Eakin et al. 1988). Let RVit and V olumeit denote the realized volatility and average traded volume (in U.S. dollars). For each observed individual i = 1, ..., N and time periods t = 1, ..., T , I consider the following linear model:
RVit = ai +
P X
γip RVit−p +
Q X q=1
p=1
22
biq V olumeit−q + ǫit ,
(2)
For simplicity, the individual fixed effects ai are assumed to be time-invariant and lag orders are identical for all cross-section units of the panel. Both the autoregressive parameters γip and the regression coefficients biq are allowed to differ across groups. In this context, volume is said to “Granger-cause” volatility if past values V olumeit−q contain information useful to predict volatility above and beyond past information contained in volatility on itself, and vice-versa. Thus, the null hypothesis is that the regression parameters are jointly not different from zero H0 : bi = 0, ∀i = 1, ..., N , with bi = (bi1 , ..., biQ ). Notice that the slope parameters can differ across groups under the alternative hypothesis. As a result, the test is not a test of Granger-causality for all of the cryptocurrencies as it allows for the presence of non-causality for a sub-group N1 under the alternative, i.e.,
H1 : bi = 0,
∀i = 1, ..., N1 ,
and
bi 6= 0,
∀i = N1 + 1, ..., N,
where N1 is unknown a priori and satisfies the condition 0 < N1 /N < 1. When N1 = N there is no causality for any of the individuals under the alternative. Conversely, when N1 = 0 there is causality for all of the groups in the sample (see Dumitrescu and Hurlin 2012 for details). Table 4 reports the results for different lags specifications.
[Insert Table 4 about here]
Top rows show that we can consistently reject the null hypothesis that volume do not Granger-cause volatility, that is, market activity significantly drive volatility. This is consistent with the fact that, by construction, equilibrium prices in cryptocurrency markets are primarily driven by demand pressure as the supply is either fixed or growth at a predetermined, exponentially decreasing, pace. As a result, shocks to traded volume, especially when trading is in large quantities, could lead equilibrium prices to go up and down sharply. Bottom rows in Table 4 suggest that there is no feedback effect between volume and volatility. As a matter of fact, the null hypothesis that volatility does not Granger-cause volume 23
cannot be significantly rejected for any of the lags specifications.
4.2
Momentum Effects
One of the most studied phenomenon in financial markets is the relation between an asset’s future performance and its past returns, the so-called “momentum” effect. Momentum strategies have become central to the academic debate on the efficiency of equity, commodity, and foreign exchanges markets, and spurred an increasing number of competing theories to explain the existence of bandwagon effects in assets’ returns. Early evidence of momentum in stock returns, either in the U.S. or internationally, or both, can be found in Jegadeesh and Titman (1993), Rouwenhorst (1998), Liew and Vassalou (2000), Griffin et al. (2003), Chui et al. (2010), and Fama and French (2012), among others. Asness et al. (1997) and Bhojraj and Swaminathan (2006) show that momentum effects are also present for aggregate stock market indexes across different countries. Empirical evidence that assets with highly positive past performances outperform those with low past returns can be found not only in equity markets but also relatively to other asset classes, such as foreign exchange markets (see, e.g., Kho 1996, LeBaron 1999, Okunev and White 2003, and Menkhoff et al. 2012), commodities (see, e.g., Erb and Harvey 2006, Miffre and Rallis 2007 and Narayan et al. 2015), and more generally across different markets (see, e.g., Moskowitz et al. 2012, Asness et al. 2013, and Daniel and Moskowitz 2016). I examine the existence of time-series momentum by first investigating the predictability of one-week ahead returns on the basis of past cumulative performances across different time horizons. I include both a time-trend and the lagged value of the average weekly traded volume as additional control variables. Methodologically, I make use of a random-coefficient regression model as in Equation (1), and regress the return on the ith cryptocurrency in week t + 1, on its own cumulative returns from t to t − h. Panel A of Table 5 reports the
24
estimation results for h = 27, 16, 12, 9, 4.
[Insert Table 5 about here]
The results show there is evidence of a mild persistence in returns for one to six months. However, only the past one-month cumulative return contain some significant predictive power on the one-week ahead returns. The null hypothesis that the predictive betas are constant across groups can be rejected only for h = 4.15 Another way to look at the predictive power of past performance is to simply focus on the direction of the cumulative past returns rather than their magnitude. I test the hypothesis that is the sign of past returns that actually matters in predicting future returns by using the same panel regression setting used so far. Panel B of Table 5 reports the estimation results for h = 27, 16, 12, 9, 4. Robust bootstrap standard errors are reported in parenthesis. Again, the conclusion that can be drawn from the regression estimates are similar to Panel A, that is, past performance seems to have a predictive effect only in the very short term. Time-series momentum implies that a simple trading strategy which goes long on past winners and short on past losers should generate positive and significance returns. I now investigate this implication by building a trading strategy based on past performances of different cryptocurrencies. For each trading strategy I compute a single time series of weekly returns following the methodology proposed by Jegadeesh and Titman (1993): for each cryptocurrency I compute the past k−week cumulative returns, skipping the most recent week’s return to avoid short-term reversals (see, e.g., Jegadeesh 1990, Lo and MacKinlay 1990, and Grinblatt and Moskowitz 2004). I keep constant the one-week holding period return while changing the “look-back period”, meaning the number of past weeks I use to construct the trading signal. To mitigate legitimate concerns about liquidity and the actual applicability of such trading strategy I focus on those currencies with a market capitalization in the top 15
A separate set of unreported results show that, except for h = 4, the currency-specific betas are virtually equal to zero for all cryptocurrencies.
25
20th percentile of the cross-sectional distribution at the end of the sample and for which the average traded volume in U.S. dollars is non-zero. The sample contains 36 cryptocurrencies which accounts for more than 95% of the total market capitalization as of September 2017. Using these trading signals based on past performances, I follow Asness et al. (2013) and construct momentum portfolios by ranking cryptocurrencies and sorting them into three equal groups – high, middle, and low – where each currency is value-weighted by their beginning-of-the-week market capitalization. Table 6 reports the mean return with its statistical significance, the standard deviation, and the Sharpe ratio of the low (P1), middle (P2), and high (P3) portfolios as well as the spread P3-P1. Each of the these statistics is computed for different look-back periods, i.e. k = 27, 12, 4 weeks.
[Insert Table 6 about here]
Momentum portfolios generates average realized returns which are positive and significant. However, the spread P3-P1 shows that a long-short strategy based on past performances does not generate a return which is statistically different from zero. Also, portfolio returns turn out to be highly volatile, which raises concerns about the suitability of momentum strategies in cryptocurrency markets for even moderately risk averse investors. Last row of Table 6 reports the intercepts, or the Jensen’s alphas, from a time-series regression of each of the portfolio returns on the return on a benchmarking market index which constructed as a value-weighted average of the cryptocurrencies originally used to construct the momentum portfolios. Weights are computed by the lagged market capitalization at each time t. Although most of the returns on momentum portfolios are statistically significant, only P3 with k = 4 weeks as look-back period outperforms the benchmark with a statistically significant alpha equal to 1.371. Yet, for none of the specifications the spreads between high and low momentum portfolios generate returns statistically different from zero. Combining the results on time-series momentum (Table 5) and momentum portfolios 26
(Table 5), the empirical evidence suggests that, if anything, there is some predictive power of past performance for future returns only in the very short term.
5
Volume, Returns, and Macroeconomic Factors
Anecdotal evidence shows that the initial fuel for Bitcoin investments has been the compounding effect of the intentional devaluation of the Chinese Yuan (CNY) and the capital controls set by the Chinese central government. Both these factors spurred the demand of bitcoins by Chinese investors looking for a safe and anonymous way to send their money off-shore. As a case in point, top panel of Figure 3 reports the relative weight of each foreign currency in terms of Bitcoin trading. Until early 2017 the majority of bitcoins were exchanged against CNY. The steady decrease throughout the sample is the effect of both government crackdown – which ended up shutting down all of the exchanges in September 2017 – and the corresponding increasing demand from other countries. In this section, I formally investigate the relationship between the traded volume and macroeconomic factors in an attempt to shed some light on the nature of trading activity in cryptocurrency markets. I consider four key factors in addition to a measure of market uncertainty and past performance of cryptocurrencies. The set of macroeconomic indicators comprise a measure of inflation expectations, the slope of the yield-curve, the Credit Default Swaps (CDS) premia on sovereign bonds for major economies, and the real effective exchange rate for the U.S. dollar (USD REER). The real effective exchange rate is calculated as the weighted average of the U.S. dollar relative to a basket of other major currencies, adjusted for inflation.16 Aggregate market uncertainty is proxied by the VIX index which measures investors’ expectations of near-term volatility of the S&P500 index. The VIX index is constructed by using the implied volatilities of a wide range of both calls and puts S&P500 16
The weights are determined by comparing the relative trade balance of the U.S. economy against each country within the basket.
27
options and is commonly referred to as “fear index” or “fear gauge” (see, e.g., Whaley 2000). The short-term expectations for future inflation are approximated by using the swap rate of the 1-year inflation swap contracts for different countries. Under no-arbitrage, the fixed payment in an inflation swap contract approximates the expected value of inflation up to a risk-premium component which depends, among others, on inflation uncertainty and investors’ risk aversion and preferences (see, e.g., Fama and Schwert 1977, Buraschi and Jiltsov 2005, H¨ordahl and Tristani 2007, and Chernov and Mueller 2012). Different measures of expected inflation can be obtained by using alternative approaches such as survey data (see, e.g., Evans 1998, Thomas 1999, Schmeling and Schrimpf 2011), or econometric modeling (see, e.g., Stock and Watson 1999 and Stock and Watson 2010). However, the advantage of inflation swaps is that they trade in active markets. This makes swap rates available at a high frequency and directly reflect expectations of actual market participants (see, e.g., Faust et al. 2013). In order to have an aggregate measure of expected inflation I extract the first principal component of the weekly 1-year swap rates for both the U.S., the Eurozone and the United Kingdom. These represent the biggest and most liquid markets for inflation swaps. The first principal components accounts for about 70% of the cross-sectional variation in the swap rates. The use of the yield spread as a macroeconomic indicator in the finance literature at least since Ferson and Harvey (1991). In order to construct an aggregate measure of the yield spread, I take the differential between 10-year and 1-year government bond yield for the US, the Eurozone, the United Kingdom, China, and Japan, and extract the first principal component which accounts for about 60% of the cross-sectional variation in the data. Similarly, I compute a cross-country CDS spread by extracting the first principal component from the sovereign 1-year CDS spread for major economies such as France, Germany, Spain, Italy, USA, United Kingdom, China and Japan. The first principal component accounts for about 75% of the variation in the cross-section of 1-year sovereign CDS spread.
28
Table 7 reports the GLS estimates from a random-coefficient panel regression as above. The response variable is the average weekly traded volume in U.S. dollars for each of the cryptocurrencies outlined in Section 3 and detailed in Appendix A. The explanatory variables are the first-order differences of the first principal components of the inflation swap rates, the yield spreads, and the CDS premia. In addition one-period changes in the VIX, the USD REER, the past returns for each currency, as well as the lagged values of both the dependent and the independent variables are included as additional control variables.
[Insert Table 7 about here]
Column one to six shows that most macroeconomic indicators significantly correlates with market activity. However, when considering all factors jointly only two regressors remain significant: the change in the VIX index and past returns. The relationship between market activity and aggregate uncertainty is consistent with the idea that increasing dispersion in investors’ expectations and aggregate risk aversion may accelerate market activity as a whole (see, e.g., Karpoff 1986, Kandel and Pearson 1995, Lo and Wang 2006, Banerjee and Kremer 2010, and Carlin et al. 2014, among others). Kandel and Pearson (1995) show that aggregate uncertainty affects trading volume and asset prices even when there are no changes in price around anticipated announcements. Banerjee and Kremer (2010) show that disagreement and uncertainty exacerbates volatility and leads to higher trading volume. Similarly, Carlin et al. (2014) find that increased disagreement and uncertainty are associated with larger trading volume. Table 7 also suggests that there is a positive correlation between returns and traded volume. Such relationship has been previously stated in the empirical finance literature. Karpoff (1987), Gallant et al. (1992), and Schwert (1989) show that there is a positive and significant correlation between daily stock returns and contemporaneous changes in trading volume. Similarly, Campbell et al. (1993) and Llorente et al. (2002) provide evidence that
29
trading volumes tend to be higher when stock prices are rising. In the following, I further investigate the existence of any longer-run relationship between macroeconomic factors and market activity, as well as causal effects between volume and past returns.
5.1
Impulse-Response Functions
Although informative, the results provided in Table 7 do not tell much as to how, and if, there is a significant effect of today macroeconomic factors on future average traded volume. To address this issue, I explore the transmission mechanism of structural macroeconomic shocks on traded volume by estimating the impulse response functions from a panel Vector Autoregressive (VAR) model. Unlike standard VARs, a panel VAR allows to explicitly account for dynamic individual inter-dependencies with minimal restrictions. Shocks identification can then be performed by explicitly considering the heterogeneity presents in potentially highly interdependent markets. More specifically, just like standard VARs all variables are assumed to be endogenous and interdependent. However, a cross-sectional dimension is added to the econometric specification which gives panel VARs three characteristic features: first, lags of all endogenous variables of all units enter the model for cryptocurrency i, i.e. dynamic inter-dependencies. Second, the reduced-form error terms are allowed to be correlated in the cross-section of cryptocurrrency returns, i.e. static inter-dependencies. Third, the model intercepts are allowed to be currency specific, i.e. cross-sectional heterogeneity (see Canova and Ciccarelli 2013 for details). The vector of endogenous variables contains the average weekly traded volume and the weekly realized returns for each cryptocurrency, the changes in the first principal components of the inflation swap rates, the yield spreads, the CDS premia, one-period changes in the VIX, and the USD REER. Top panels of Figure 5 show the impulse response function (IRF) of traded volume to inflation (left panel) and the yield spread (right panel). The dark blue line represents the average impulse response and the light-blue shaded area represents the
30
95% credibility intervals obtained by a double non-parametric bootstrap scheme, which is a combination of temporal re-sampling and cross-sectional re-sampling (see Kapetanios 2008 for details). [Insert Figure 5 about here] Neither shocks to inflation expectations nor shocks to the yield curve significantly affect market activity in cryptocurrency markets. However, the sign of the impulse response in the short term is consistent with the conventional wisdom and anecdotal evidence. The average effect of a shock to inflation expectations is positive indicating that increasing inflation might be a relevant factor in increasing the propensity to trade in cryptocurrencies. Similarly, a positive shock to the yield curve decreases trading in cryptos, which is consistent with the idea that the marginal propensity to invest in cryptocurrencies decrease with improving economic conditions. As a whole, top panels of Figure 5 suggest that trading in cryptocurrencies can be counter-cyclical. Although shocks to the risk of default of major economies do not affect the propensity to trade in cryptocurrencies, bottom-right panel of Figure 5 shows that aggregate market uncertainty appears to significantly drive traded volume up to two weeks ahead. Figure 6 reports the response of traded volume on a shock to returns (left panel) and the response of returns to a shock in market activity (right panel). Although positive, the response of realized returns on a one-unit shock to traded volume is not significantly different from zero. This somewhat contradicts some of the existing literature; for instance, Gervais et al. (2001) investigate the role of trading activity in terms of the information it contains about future prices, and find that individual stocks whose trading volume is usually large (small) over period of a day or a week, tend to experience large (small) returns over the subsequent month. Similarly, Saatcioglu and Starks (1998) find that volume significantly leads stock price changes in emergency markets.
[Insert Figure 6 about here]
31
Instead, left panel shows that trading activity is significantly driven by past performances up to four weeks ahead. This is consistent with existing literature in empirical finance and behavioral economics which shows that investments can be widely affected by investors sentiment and past performances. For instance, Huddart et al. (2009) document that trading volumes in equity markets are higher when current prices are above the usual range defined by the past price trajectory over the previous fifty-two weeks of trading days. Similarly, Kliger and Kudryavtsev (2010) show that abnormal trading volumes are significantly higher (lower) if the general stock market index rises (falls).
5.2
A Further Discussion on Volume and Returns
In this section I further investigate causal effects between realized returns and traded volume. As in Section 4.1 feedback effects and causality are tested by using a panel Granger causality test as proposed by Dumitrescu and Hurlin (2012). Table 8 reports the test for different lags specifications. The results confirm that there is no feedback relationship between returns and volume; while the null hypothesis that volume does not lead returns can not be rejected for any lags, there is strong evidence that market activity in cryptocurrency markets is driven by past performances. [Insert Table 8 about here] As a whole, the evidence provided by Table 7 and 8, as well as by Figure 5 and 6, suggest that trading activity in cryptocurrency markets may be primarily driven by past performances rather than by macroeconomic events. These results are consistent and complement earlier works on the effect of investors’ sentiment on the time series of stock returns, such as Kothari and Shanken (1997), Neal and Wheatley (1998), Shiller Robert (2000), and Baker and Wurgler (2000), among others.
32
6
Conclusion
The Blockchain technology behind Bitcoin has led to a proliferation of alternative coins, called cryptocurrencies, and, in turn, to a massive increase in the number of exchanges and investors who actually trade these cryptocurrencies. Just as the value of a U.S. dollar investment fluctuates based on countless factors, such as national interest rates, trade deficit with other countries, and government policy, cryptocurrencies trade at prices which are based on the perceived value of the platforms and projects they are associated with. In this respect, instead of an investment in a country’s economy, holding a cryptocurrency can be seen as an investment in the network and the technology behind it. In this paper, I use a novel data set of prices, traded volumes, and market capitalization for a large set of cryptocurrencies to empirically investigate both their relationship with standard asset classes and the main driving factors behind market activity. The main empirical results suggest that there is a significant relationship between returns on cryptocurrencies and commodities such as gold and energy. Also, while volatility correlates with traded volume, the latter is primarily driven by past returns and by a short-lived effect of aggregate market uncertainty. This is consistent with existing theoretical models in which trading activity is primarily driven by investors’ sentiment. Finally, impulse-response functions from a panel Vector Autoregressive (VAR) model show that macroeconomic factors do not significantly drive trading activity in cryptocurrency markets.
References Andersen, T. G. 1996. Return volatility and trading volume: An information flow interpretation of stochastic volatility. The Journal of Finance 51:169–204. Asness, C. S., J. M. Liew, and R. L. Stevens. 1997. Parallels between the cross-sectional predictability of stock and country returns. The Journal of Portfolio Management 23:79– 87. 33
Asness, C. S., T. J. Moskowitz, and L. H. Pedersen. 2013. Value and momentum everywhere. The Journal of Finance 68:929–985. Baker, M., and J. Wurgler. 2000. The equity share in new issues and aggregate stock returns. the Journal of Finance 55:2219–2257. Banerjee, S., and I. Kremer. 2010. Disagreement and learning: Dynamic patterns of trade. The Journal of Finance 65:1269–1302. Bhojraj, S., and B. Swaminathan. 2006. Macromomentum: returns predictability in international equity indices. The Journal of Business 79:429–451. Bordo, M. D., and A. T. Levin. 2017. Central bank digital currency and the future of monetary policy. Tech. rep., National Bureau of Economic Research. Buraschi, A., and A. Jiltsov. 2005. Inflation risk premia and the expectations hypothesis. Journal of Financial Economics 75:429–490. Campbell, J. Y., S. J. Grossman, and J. Wang. 1993. Trading volume and serial correlation in stock returns. The Quarterly Journal of Economics 108:905–939. Canova, F., and M. Ciccarelli. 2013. Panel Vector Autoregressive Models: A Survey. In VAR Models in Macroeconomics–New Developments and Applications: Essays in Honor of Christopher A. Sims, pp. 205–246. Emerald Group Publishing Limited. Carlin, B. I., F. A. Longstaff, and K. Matoba. 2014. Disagreement and asset prices. Journal of Financial Economics 114:226–238. Chernov, M., and P. Mueller. 2012. The term structure of inflation expectations. Journal of financial economics 106:367–394. Chiu, J., and T. V. Koeppl. 2017. The economics of crypto currencies-Bitcoin and beyond. Working Paper . Chui, A. C., S. Titman, and K. J. Wei. 2010. Individualism and momentum around the world. The Journal of Finance 65:361–392. Cong, L. W., Z. He, and J. Zheng. 2017. Blockchain Disruption and Smart Contracts . Daniel, K., and T. J. Moskowitz. 2016. Momentum crashes. Journal of Financial Economics 122:221–247. Dumitrescu, E.-I., and C. Hurlin. 2012. Testing for Granger non-causality in heterogeneous panels. Economic Modelling 29:1450–1460. Dwyer, G. P. 2015. The economics of Bitcoin and similar private digital currencies. Journal of Financial Stability 17:81–91. Dyhrberg, A. H. 2016. Bitcoin, gold and the dollar - A GARCH volatility analysis. Finance Research Letters 16:85–92. Engle, R. 2002. Dynamic conditional correlation: A simple class of multivariate generalized autoregressive conditional heteroskedasticity models. Journal of Business & Economic Statistics 20:339–350. Erb, C. B., and C. R. Harvey. 2006. The strategic and tactical value of commodity futures. Financial Analysts Journal 62:69–97. Evans, D. S. 2014. Economic aspects of Bitcoin and other decentralized public-ledger currency platforms. Working Paper . 34
Evans, M. D. 1998. Real rates, expected inflation, and inflation risk premia. The Journal of Finance 53:187–218. Eyal, I., and E. G. Sirer. 2014. Majority is not enough: Bitcoin mining is vulnerable. In International conference on financial cryptography and data security, pp. 436–454. Fama, E. F., and K. R. French. 2012. Size, value, and momentum in international stock returns. Journal of financial economics 105:457–472. Fama, E. F., and G. W. Schwert. 1977. Asset returns and inflation. Journal of financial economics 5:115–146. Faust, J., J. H. Wright, et al. 2013. Forecasting inflation. Handbook of economic forecasting 2:3–56. Ferson, W. E., and C. R. Harvey. 1991. The variation of economic risk premiums. Journal of Political Economy 99:385–415. Foster, F. D., and S. Viswanathan. 1993. The effect of public information and competition on trading volume and price volatility. The Review of Financial Studies 6:23–56. Gallant, A. R., P. E. Rossi, and G. Tauchen. 1992. Stock prices and volume. The Review of Financial Studies 5:199–242. Gervais, S., R. Kaniel, and D. H. Mingelgrin. 2001. The high-volume return premium. The Journal of Finance 56:877–919. Glosten, L. R., and P. R. Milgrom. 1985. Bid, ask and transaction prices in a specialist market with heterogeneously informed traders. Journal of financial economics 14:71–100. Granger, C. W. 1988. Some recent development in a concept of causality. Journal of econometrics 39:199–211. Griffin, J. M., X. Ji, and J. S. Martin. 2003. Momentum investing and business cycle risk: Evidence from pole to pole. The Journal of Finance 58:2515–2547. Grinblatt, M., and T. J. Moskowitz. 2004. Predicting stock price movements from past returns: The role of consistency and tax-loss selling. Journal of Financial Economics 71:541–579. Hamilton, J. 2009. Understanding Crude Oil Prices. The Energy Journal pp. 179–206. Holden, C. W., and A. Subrahmanyam. 1992. Long-lived private information and imperfect competition. The Journal of Finance 47:247–270. Holtz-Eakin, D., W. Newey, and H. S. Rosen. 1988. Estimating vector autoregressions with panel data. Econometrica: Journal of the Econometric Society pp. 1371–1395. H¨ordahl, P., and O. Tristani. 2007. Inflation risk premia in the term structure of interest rates . Huberman, G., J. D. Leshno, and C. C. Moallemi. 2017. Monopoly without a monopolist: An economic analysis of the bitcoin payment system. Working Paper . Huddart, S., M. Lang, and M. H. Yetman. 2009. Volume and price patterns around a stock’s 52-week highs and lows: Theory and evidence. Management Science 55:16–31. Jegadeesh, N. 1990. Evidence of predictable behavior of security returns. The Journal of finance 45:881–898.
35
Jegadeesh, N., and S. Titman. 1993. Returns to buying winners and selling losers: Implications for stock market efficiency. The Journal of finance 48:65–91. Kandel, E., and N. D. Pearson. 1995. Differential interpretation of public signals and trade in speculative markets. Journal of Political Economy 103:831–872. Kapetanios, G. 2008. A bootstrap procedure for panel data sets with many cross-sectional units. The Econometrics Journal 11:377–395. Karpoff, J. M. 1986. A theory of trading volume. The Journal of Finance 41:1069–1087. Karpoff, J. M. 1987. The relation between price changes and trading volume: A survey. Journal of Financial and quantitative Analysis 22:109–126. Kho, B.-C. 1996. Time-varying risk premia, volatility, and technical trading rule profits: Evidence from foreign currency futures markets. Journal of Financial Economics 41:249– 290. Kilian, L. 2009. Not all oil price shocks are alike: Disentangling demand and supply shocks in the crude oil market. The American economic review 99:1053–1069. Kim, O., and R. E. Verrecchia. 1991. Trading volume and price reactions to public announcements. Journal of accounting research pp. 302–321. Kliger, D., and A. Kudryavtsev. 2010. The availability heuristic and investors’ reaction to company-specific events. The Journal of Behavioral Finance 11:50–65. Kondor, D., M. P´osfai, I. Csabai, and G. Vattay. 2014. Do the rich get richer? An empirical analysis of the Bitcoin transaction network. PloS one 9. Kothari, S. P., and J. Shanken. 1997. Book-to-market, dividend yield, and expected market returns: A time-series analysis. Journal of Financial Economics 44:169–203. Kyle, A. S. 1985. Continuous auctions and insider trading. Econometrica 53:1315–1335. LeBaron, B. 1999. Technical trading rule profitability and foreign exchange intervention. Journal of international economics 49:125–143. Liew, J., and M. Vassalou. 2000. Can book-to-market, size and momentum be risk factors that predict economic growth? Journal of Financial Economics 57:221–245. Llorente, G., R. Michaely, G. Saar, and J. Wang. 2002. Dynamic volume-return relation of individual stocks. The Review of Financial Studies 15:1005–1047. Lo, A. W., and A. C. MacKinlay. 1990. When are contrarian profits due to stock market overreaction? The Review of Financial Studies 3:175–205. Lo, A. W., and J. Wang. 2006. Trading volume: Implications of an intertemporal capital asset pricing model. The Journal of Finance 61:2805–2840. Menkhoff, L., L. Sarno, M. Schmeling, and A. Schrimpf. 2012. Currency momentum strategies. Journal of Financial Economics 106:660–684. Miffre, J., and G. Rallis. 2007. Momentum strategies in commodity futures markets. Journal of Banking & Finance 31:1863–1886. Moore, T., and N. Christin. 2013. Beware the middleman: Empirical analysis of Bitcoinexchange risk. In International Conference on Financial Cryptography and Data Security, pp. 25–33. 36
Moskowitz, T. J., Y. H. Ooi, and L. H. Pedersen. 2012. Time series momentum. Journal of Financial Economics 104:228–250. Narayan, P. K., H. A. Ahmed, and S. Narayan. 2015. Do Momentum-Based Trading Strategies Work in the Commodity Futures Markets? Journal of Futures Markets 35:868–891. Neal, R., and S. M. Wheatley. 1998. Do measures of investor sentiment predict returns? Journal of Financial and Quantitative Analysis 33:523–547. Okunev, J., and D. White. 2003. Do momentum-based strategies still work in foreign currency markets? Journal of Financial and Quantitative Analysis 38:425–447. Pesaran, M. H., and R. Smith. 1995. Estimating long-run relationships from dynamic heterogeneous panels. Journal of econometrics 68:79–113. Rouwenhorst, K. G. 1998. International momentum strategies. The Journal of Finance 53:267–284. Saatcioglu, K., and L. T. Starks. 1998. The stock price–volume relationship in emerging stock markets: the case of Latin America. International Journal of forecasting 14:215–225. Schmeling, M., and A. Schrimpf. 2011. Expected inflation, expected stock returns, and money illusion: What can we learn from survey expectations? European Economic Review 55:702–719. Schwert, G. W. 1989. Why does stock market volatility change over time? The journal of finance 44:1115–1153. Shiller Robert, J. 2000. Irrational exuberance. Princeton, New Jersey, Princeon University . Stock, J. H., and M. W. Watson. 1999. Forecasting inflation. Journal of Monetary Economics 44:293–335. Stock, J. H., and M. W. Watson. 2010. Modeling inflation after the crisis. Tech. rep., National Bureau of Economic Research. Swamy, P. A. 1970. Efficient inference in a random coefficient regression model. Econometrica pp. 311–323. Thomas, L. B. 1999. Survey measures of expected US inflation. The Journal of Economic Perspectives 13:125–144. Whaley, R. E. 2000. The investor fear gauge. The Journal of Portfolio Management 26:12–17. Yermack, D. 2013. Is Bitcoin a real currency? An economic appraisal. National Bureau of Economic Research .
37
Appendix A
Cryptocurrencies for the Main Empirical Analysis
The initial sample refers to data on prices, traded volume (in U.S. dollars) and market capitalization for 1251 cryptocurrencies which are traded as of September 2017. The sample period is from April 2016 to September 2017. Daily data are converted in weekly observations to keep a large enough number of observations while mitigating the noise that characterize the highly volatile daily returns. The final cross section of cryptocurrencies is obtained by pre-filtering for market capitalization and average weekly traded volume. In particular, I eliminate those digital currencies which have an average market capitalization below the top 5th percentile at the end of the sample. Similarly, I eliminate from the main sample those cryptocurrencies which have a sample average traded volume below the median. The final sample is made by fourteen cryptocurrencies which accounts for more than 85% of the total market capitalization as of September 2017. In the following I briefly describe each of those fourteen currencies in alphabetical order.
A.1
BitShares
BitShares (BTS) is an open-source Financial Smart Contracts platform that enables trading of digital assets and has market-pegged assets that track the value of their underlying asset, e.g. bitUSD tracking the U.S. dollar. As of September 2017 the market capitalization of BitShares is about $200mln, with a circulating and maximum supply equal to 2.6 and 3.6 billion coins, respectively.
38
A.2
Bitcoin
Bitcoin (BTC) has been created in 2009 following the ideas set out in a white paper by Satoshi Nakamoto, whose true identity has yet to be verified. The original idea of Bitcoin is to provide a payment method which lower transaction fees than traditional online payment systems and is operated by a consensus network, therefore without a trusted third-parties or a central authority like happens for government-issued currencies. There are no physical bitcoins circulating, only balances kept on a public ledger that, along with all transactions, is verified by a massive network of computers. Unlike fiat currencies, bitcoins are not guaranteed by any central bank or government agencies. As of September 2017 the market capitalization of BitShares is about $70bn, with a circulating and maximum supply equal to 16 and 21 million coins, respectively.
A.3
Bytecoin
Bytecoin (BCN) has been launched in 2012 and is the first cryptocurrency based on the CryptoNote protocol with an open source code designed for anonymous cash settlement. BCN protects the user’s privacy by securing the transactions as the identities of the sender, the receiver and the amount of transaction are all concealed. As of September 2017 the market capitalization is $254mln, with circulating and maximum supply 183 and 184 billion coins, respectively.
A.4
Dash
Dash (DAHS), formerly known as Darkcoin and XCoin, is an open source peer-to-peer cryptocurrency that offers instant transactions (InstantSend), private transactions (PrivateSend) and operates a self-governing and self-funding model that enables the Dash network to pay individuals and businesses to perform work that adds value to the network. In this respect, unlike Bitcoin’s, where all jobs on the network are performed by miners, i.e. single-tier 39
network, Dash utilizes a two-tier network whereby certain functions, such as creating new blocks, are handled by the miners while second tier “masternodes” perform PrivateSend, InstantSend, and governance functions. Dash’s decentralized governance and budgeting system makes it a decentralized autonomous organization (DAO). As of September 2017 the market capitalization is $2.5bn, with circulating and maximum supply 7.7 and 18.9 million coins, respectively.
A.5
Dogecoin
Although launched as a “joke currency” in 2013, Dogecoin (DOGE) quickly developed its own online community. In particular, the currency has gained traction as an payment method in social media by which users grant Dogecoin tips to other users for providing interesting or noteworthy content. There are few mainstream commercial applications but almost the entire amount of transactions are made on online communities such as Twitter and Reddit. As of September 2017 the market capitalization is $130mln, with circulating supply equal to 111 billions coins. There is no maximum supply set by the protocol for this coin.
A.6
Ethereum
Ethereum (ETH) has been initially proposed in 2013 as an open-source distributed computing platform which features smart contracts, i.e. scripting, functionality based on a decentralized Turing-complete virtual machine whereby scripts can be executed using an international newtork of public nodes. In 2016 Ethereum was forked into two separate blockchains thereby creating Ethereum Classic. In the main empirical analysis we will use the forked version ETH being the most traded version. As of September 2017 the market capitalization is $30bn, with circulating supply equal to 96 millions coins. The maximum supply of ETH is not fixed although was projected to increase by 14.75% in 2017, with an algorithm in place to gradually decline to 1.59% by 2065. However, a new implementation of based on proof-of-
40
stake rather than proof-of-work protocols is expected to reduce the inflation rate to between 0.5% to 2%.
A.7
Factom
Factom (FCT) represents an open-source blockchain-based protocol which allow to store any type of data, from financial transactions to simple business analysis, which makes the platform a distributed and decentralized record-keeping system. More specifically, FCT provides a document authentication solution that secure physical documents on the Blockchain. As of September 2017 the market capitalization is $140mln, with circulating supply equal to 8.7 million coins.
A.8
GameCredits
GameCredits (GAME) blockchain-based currency which aims at serving gamers worldwide and empower game developers. Its purpose is to let gamers and developers buy and sell games and in-game items fast, safely and privately through a distributed and decentralized platform. The processing time for each block is 90 seconds and miners reward is currently at 12.5 coins. As of September 2017 the market capitalization is $150mln, with circulating and maximum supply equal to 64 and 84 million coins, respectively.
A.9
Litecoin
Litecoin (LTC) was released via an open-source client on GitHub on 2011. It represents a peer-to-peer cryptocurrency based on an publicly available cryptographic protocol and which decentralized and not managed by any central authority. While in many regards nearly identical to Bitcoin, LTC has almost zero payment cost and facilitates payments faster than Bitcoin by having a decreased block generation time, increased maximum number of coins, and used a different hashing algorithm. As of September 2017 the market capitalization is 41
$2.6bn, with circulating and maximum supply equal to 54 and 84 million coins, respectively.
A.10
Monero
Monero (MOX) is an open-source cryptocurrency introduced in 2014 which is based on the CryptoNote protocol and possesses significant algorithmic differences relating to blockchain obfuscation. By providing a high level of privacy, units of currencies are indistinguishable from one another, meaning that every unit of the currency can be substituted by another unit. This makes Monero different from public-ledger cryptocurrencies, where addresses previously associated with undesired activity can be blacklisted and refused by network members. The privacy of transactions is protected by hiding both the sending address, the amount of transactions and the receiving address. As of September 2017 the market capitalization is $1.4bn, with circulating supply equal to 15 million coins.
A.11
NEM
NEM (XEM) is a peer-to-peer cryptocurrency and blockchain platform launched in 2015. Unlike most altcoins, NEM is based on a proof-of-importance (POI) algorithm, multisignature accounts, encrypted messaging, and an Eigentrust reputation system. Being based on a POI algorithm, it does not require a lot of computing power and energy to mine. The NEM blockchain software is private and has a licensed commercial version called Mijin, which makes NEM a first attempt of public/private combination in blockchain implementation. As of September 2017 the market capitalization is $2bn, with circulating supply equal to 9 billion coins.
A.12
Ripple
Ripple (XRP) is the native currency of the Ripple network that only exists within the Ripple system. The latter is an open-source, distributed peer-to-peer real-time gross settlement 42
system, currency exchange and remittance network. The system is designed such that XRP is a scarce asset with decreasing available supply. Not dependent on any third party for redemption, XRP is the only currency in the Ripple network that does not entail counterparty risk, while other currencies exchanged in the Ripple network are liabilities, and exist in the form of balances. One of the specific functions of XRP is as a bridge currency, for e.g. when transacting between two rarely traded currency pairs. As of September 2017 the market capitalization is $8bn, with circulating and maximum supply equal to 38 and 100 billion coins, respectively.
A.13
Siacoin
Siacoin (SC) is the native currency on the Sia network which provides decentralized, encrypted, peer-to-peer storage. In this respect, the idea behind Sia is that instead of everything being hosted on central servers, blockchain technology enables to decentralize file storage and make it open source. Companies are then involved and host their own private decentralized cloud and sell that as a service to their customers. Sia essentially provides the platform that will have users paying for storage and providers running their own private decentralized cloud. As of September 2017 the market capitalization is $150mln, with circulating supply equal to 30 billion coins.
A.14
Steem
Steem (STEEM) is a native currency for the Steemit social media platform which allows publishers to monetize their online content and community, based on battle-tested blockchain technology. Steem began with a highly inflationary supply model, doubling roughly every year. Due to increasing demand pressure the inflation rate of Steem was changed to 9.5% per year, reducing by 0.5% per year. As of September 2017 the market capitalization is $400mln, with circulating and maximum supply equal to 246 and 263 million coins, respectively.
43
B
Robustness
In this section I re-examine the main empirical analysis by using a restricted version of the panel regression framework outlined in Section 4. Let xt = (x1t , · · · , xKt )′ denote the Kdimensional vector of explanatory variables and yit the response variable for the ith group at time t. A standard random-intercept (or random-effect) panel regression which is defined as follows,
yit = (α + α0i ) + β ′ xt + ǫit ,
i = 1, ..., N,
t = 1, ..., T,
(A.1)
where ǫit ∼ N (0, σ 2 ), α0i ∼ N (0, τ 2 ) and the slope parameter is common across groups, i.e. β 0i = β, ∀i = 1, ..., N . Table B.1 reports the estimates with robust standard errors reported in parenthesis. The response variable is the set of weekly returns for the cryptocurrencies outlined in Section 3 and detailed in Appendix A. The explanatory variables are the weekly returns on the FTSE Global All-Cap index (Equity), the BofA-Merrill Lynch Global Broad Market index (Bond), the MSCI World REITs index (Real Estate), the S&P GSCI Energy index (Energy), the S&P GSCI Gold index (Gold), the S&P 500 VIX Short-Term Futures index (Options), and the log-change of the nominal effective exchange rate of the U.S. dollar (FX) which represents a trade-weighted average of nominal bilateral rates between the USD and a basket of foreign currencies. I include both lagged returns of both the cryptos and the other assets, as well as the lagged values of the average weekly traded volumes as additional control variables. [Insert Table B.1 about here] The results substantially support the main empirical analysis of Section 4. Returns on the Energy commodity sector show a significant negative average correlation with the returns on cryptocurrencies. The opposite holds for the weekly returns on on Gold. This is consistent with the conventional wisdom that posits that cryptocurrencies are a “store-of-value”
44
asset being in limited supply, in the spirit of precious metals and gold in particular. Not surprisingly, the correlation with other asset classes is virtually zero. This supports the view of cryptocurrencies as a viable diversification investment. I now re-examine the evidence on spillovers effects in volatilities as in Section 4.1. The realized volatility for both cryptocurrencies and the other asset classes is computed as the square root of the sum of daily demeaned squared returns. I include a linear time trend and used lagged values of the realized volatilities of both the cryptos and the other assets as additional control variables. Table B.2 reports the estimation results with robust standard errors are reported in parenthesis.
[Insert Table B.2 about here]
The estimates widely confirm the results from Section 4. Consistent with existing evidence, both theoretical, see e.g. Glosten and Milgrom (1985) and Kim and Verrecchia (1991), and empirical, see e.g. Gallant et al. (1992) and Andersen (1996), there is a significant positive correlation between returns volatility and average traded volume. For the sake of completeness, I also re-examine the effect of aggregate macro-financial indicators on trading activity. Table B.3 shows the estimation results with robust standard reported in parenthesis. The response variables are the average weekly traded volume in U.S. dollars for the set of cryptocurrencies outlined in Section 3 and detailed in Appendix A. The explanatory variables are detailed in Section 5 of the main text. I include lagged values of both responses and covariates, as well as the lagged values of the average weekly traded volumes as additional control variables. [Insert Table B.3 about here] The empirical results confirm that the average traded volume for cryptocurrencies mainly correlates with past performance and the aggregate level of market uncertainty as proxied by the VIX index. 45
Table 1. Descriptive Statistics This table shows the descriptive statistics for the sample of cryptocurrencies used in the main empirical analysis. The initial sample refers to daily data on prices, volume and available supply for 1251 cryptocurrencies which are traded as of September 2017. In order to mitigate concerns about actual tradability and market liquidity we consider only those cryptocurrencies within the top 5% of market capitalization and for which we have at least a year of good-quality trading data. Thus, the final sample is fourteen currencies which accounts for more than 85% of the total market capitalization as of September 2017. The sample period is from April 2016 to September 2017, weekly. Panel A: reports the unconditional mean, median, standard deviation, skewness and kurtosis for the sample returns (expressed in percentages), as well as the weekly Sharpe ratio. Panel B: reports the sample mean, median, and standard deviation of the market capitalization computed as the market price at the end of the week times the number of coins circulating at the end of the same week, expressed in millions of U.S. dollars. The fourth row reports the ratio between the last and the first observations in the sample. Finally, Panel C: reports the sample mean, median and standard deviation of the weekly average traded volume expressed in millions of U.S. dollars. The fourth row reports the ratio between the last and the first observations in the sample. The sample period for returns, market capitalization and the average traded volume is from April 2016 to September 2017, weekly.
Panel A: Returns
46
Mean Median St Dev Skewness Kurtosis SR
BitShares Bitcoin 3.534 2.901 -0.599 1.682 22.04 9.708 1.144 -0.418 6.653 5.442 0.160 0.299
Bytecoin 4.259 0.192 27.40 1.513 6.673 0.155
Dash 5.173 4.379 15.76 1.292 7.666 0.328
Dogecoin 2.180 0.958 18.26 0.534 10.359 0.119
Ethereum 4.565 1.538 18.94 1.593 8.649 0.241
Factom 3.505 3.203 19.64 0.580 5.031 0.178
GameCredits Litecoin 4.691 3.609 1.152 0.750 24.16 15.57 0.705 0.543 4.552 6.552 0.194 0.232
Monero 5.833 1.893 20.77 0.994 4.367 0.281
NEM 6.605 0.106 21.37 0.905 4.904 0.309
Ripple Siacoin 4.423 4.821 -1.478 -2.835 21.75 25.10 2.338 0.536 9.134 3.979 0.203 0.192
Steem 0.721 -2.802 43.73 3.486 22.558 0.016
Ripple 2663.09 252.70 3960 35.58
Siacoin 74.35 9.31 123.2 98.41
Steem 135.25 49.96 144.2 933.32
Ripple Siacoin 293.2 22.22 12.20 1.302 696.0 52.20 25.96 77.29
Steem 6.343 1.140 14.90 26.31
Panel B: Market Capitalization (mln USD) Mean Median Std Last/First
BitShares Bitcoin 112.7 22879 12.51 14496 195.2 19000 15.57 9.98
Bytecoin 96.65 10.03 155.1 31.54
Dash 545.7 89.45 741.2 63.13
Dogecoin 82.52 25.14 101.3 5.20
Ethereum 7710.46 1105.27 10900 42.25
Factom 65.42 26.74 76.12 13.54
GameCredits 52.14 12.48 73.73 50.26
Litecoin 807.90 196.31 1020 18.78
Monero 33.05 166.31 434.2 117.21
NEM 569.57 57.71 866.2 169.92
Panel C: Average Traded Volume (mln USD) Mean Median Std Last/First
BitShares Bitcoin 65.71 2520 1.326 787.0 160.00 3420 66.60 19.45
Bytecoin 4.761 0.039 12.80 1512
Dash 80.42 10.50 111.0 285.4
Dogecoin 21.73 1.230 56.00 28.57
Ethereum 1410 138.0 2480 15.57
Factom 11.12 5.081 20.40 1.272
GameCredits 4.895 0.999 7.481 299.3
Litecoin 506.1 24.70 925.0 164.8
Monero 60.73 24.00 108.0 64.60
NEM 16.61 1.496 28.60 49.36
Table 2. Cryptocurrencies and Other Asset Classes This table shows the estimates of a random-coefficient panel regression model which investigates systematic correlations between the cross-section of cryptocurrency returns and other asset classes. The set of cryptos used in the regression is outlined in Section 3 and detailed in Appendix A. The set of weekly returns on other asset classes comprises the FTSE Global All-Cap index (Equity), the BofA-Merrill Lynch Global Broad Market Index (Bond), the MSCI World REITs index (Real Estate), the S&P GSCI Energy index (Energy), the S&P GSCI Gold index (Gold), the S&P 500 VIX Short-Term Futures Index (Volatility), and the logchange of the nominal effective exchange rate of the U.S. dollar (FX) which represents a trade-weighted average of nominal bilateral rates between the USD and a basket of foreign currencies. Lagged returns are included in addition to past changes in the average traded volume as control variables. The sample period is from April 2016 to September 2017, weekly. Robust standard errors are generated by 5,000 bootstrap iterations and reported in parenthesis. * p < 0.10, ** p < 0.05, *** p < 0.01
Panel A: Panel Regression Estimates Equity
M1 -0.0451 (1.004)
Bond
M2
M3
M4
M5
M6
M8
1.309 (1.046)
Real Estate
0.643 (0.590)
Energy
-0.408** (0.181)
Gold
-0.417** (0.189) 0.654** (0.338)
0.870** (0.437)
FX
-0.321 (1.126)
Volatility Obs. N Wald (χ2 )
M7
1063 14 66.64
1063 14 66.20
1063 14 63.60
1063 14 61.35
1063 14 64.00
47
1063 14 62.47
0.0136 (0.058) 1063 14 64.39
1063 14 68.40
M9 -1.627 (1.396) -5.827 (3.870) 0.625 (0.950) -0.919*** (0.259) 0.501** (0.242) -4.634* (2.719) -0.0442 (0.083) 1063 14 154.9
Table 3. Volatility Spillover and Volume This table shows the estimates of a random-coefficient panel regression model which investigates contemporaneous correlations in the volatilities of cryptocurrencies and other asset classes conditionally on the average weekly traded volume. Realized volatility for both cryptocurrencies and other asset classes is computed as the square root of the sum of daily squared demeaned returns. Lagged values of the realized volatilities are included in addition to a time trend and past changes in the average traded volume as control variables. The sample period is from April 2016 to September 2017, weekly. Robust standard errors are generated by 5,000 bootstrap iterations and reported in parenthesis. * p < 0.10, ** p < 0.05, *** p < 0.01
Panel A: Panel Regression Estimates Equity
M1 0.781 (0.827)
Bond
M2
M3
M4
M5
M6
M7
M8
0.222 (1.871)
Real Estate
-0.360 (0.623)
Energy
0.255 (0.450)
Gold
0.513 (0.465) -0.800 (0.645)
-0.350 (0.610)
FX
1.039 (1.655)
Volatility
-0.029 (0.060)
Volume Obs. N Wald (χ2 )
M9
1063 14 166.2
1063 14 157.6
1063 14 146.8
1063 14 152.0
1063 14 149.6
1063 14 147.1
1063 14 163.6
0.066*** (0.009) 1063 14 223.0
1063 14 162.3
M10 1.393 (1.286) 0.900 (2.551) -0.777 (0.981) 0.742 (0.521) -1.849* (0.987) 3.161 (1.912) -0.190 (0.172) 0.067*** (0.010) 1063 14 352.9
Table 4. Traded Volume and Volatility This table shows the results of a panel Granger causality test between traded volume and volatility using the methodology proposed by Dumitrescu and Hurlin (2012). This test allows to test for causality in heterogeneous panel data models whereby the cross-sectional heterogeneity of the causal relationships between volume and volatility is explicitly acknowledged. The table shows the test results for different lags of the autoregressive parameters and the regression slopes as in Equation (2). The sample period is from April 2016 to September 2017, weekly.
Panel A: Panel Granger Causality Test
H0 : Volume do not Granger-cause volatility
H0 : Volatility does not Granger-cause volume
1 z-stat 3.224 p-value 0.001
2 2.842 0.004
3 2.331 0.019
# Lags 4 5 1.692 2.741 0.097 0.006
6 3.022 0.002
7 2.629 0.008
8 3.674 0.000
z-stat 1.509 p-value 0.131
1.155 0.248
-0.057 0.959
0.007 0.994
-0.931 0.352
-0.441 0.658
-0.821 0.409
48
-0.653 0.515
Table 5. Time-Series Momentum and Returns Predictability This table shows the results of a random-coefficient panel predictive regression framework whereby the response variable is the one-week ahead returns on cryptocurrencies and the explanatory variable is the onelagged cumulative returns over the past k weeks. The average weekly traded volume which is also included as additional control variable. Panel A: reports the results obtained by considering the cumulative returns skipping the one-lag weekly returns. Panel B: reports the results obtained by considering the sign of the cumulative returns over the last k weeks skipping the one-lag returns. Robust standard are generated by 5,000 bootstrap iterations and reported in parenthesis. The sample period is from April 2016 to September 2017, weekly. * p < 0.10, ** p < 0.05, *** p < 0.01
Panel A: Cumulative Lagged Returns 27 Cumulative Returns -0.055 (0.009) Volume (-1) 0.013 (0.014) Obs. 685 N 14 Wald (χ2 ) 41.58
Lags (Weeks) 16 12 9 0.003 0.015 0.014 (0.010) (0.012) (0.017) 0.013 0.012 0.013 (0.012) (0.012) (0.012) 839 895 937 14 14 14 44.26 33.26 28.96
4 0.041** (0.021) 0.008 (0.012) 1,007 14 25.12
Panel B: Signed Cumulative Lagged Returns
sign(Cumulative Returns) Volume (-1) Obs. N Wald (χ2 )
27 0.551 (1.923) 0.013 (0.014) 685 14 31.58
Lags (Weeks) 16 12 9 0.826 1.603 1.721 (1.846) (1.041) (1.147) 0.012 0.012 0.013 (0.012) (0.012) (0.012) 839 895 937 14 14 14 29.68 33.94 39.07
49
4 2.281** (1.041) 0.009 (0.011) 1,007 14 44.59
Table 6. Momentum Portfolios This table shows the performance statistics of a set of portfolios constructed following the methodology proposed by Jegadeesh and Titman (1993): for each cryptocurrency we compute the past k−week, i.e. lookback period, cumulative returns, skipping the most recent week’s return to avoid short-term reversals. We utilize the weekly returns for those currencies with a market capitalization in the top 20th percentile of the cross-sectional distribution at the end of the sample and for which the average traded volume in U.S. dollars is non-zero. We construct momentum portfolios by ranking cryptocurrencies and sorting them into three equal groups – high, middle, and low – where we value weight each currency in the portfolios by their beginningof-the-week market capitalization. The table reports the mean return with its statistical significance, the standard deviation, and the Sharpe ratio of the low (P1), middle (P2), and high (P3) portfolios as well as the spread P3-P1. Also, we report the alphas from a time-series regression of each of the portfolio returns series on the return on a market index constructed as a value-weighted average of the all cryptocurrencies used to construct the momentum portfolios. The sample period is from April 2016 to September 2017, weekly. * p < 0.10, ** p < 0.05, *** p < 0.01
Panel A: Momentum Portfolios 27-Week Look-back period P1 P2 P3 P3-P1 Mean 1.422* 1.676*** 1.861* 0.439 Stdev 5.661 3.301 7.081 6.831 Sharpe 0.251 0.507 0.261 0.061 Alpha 0.445 1.414 1.361 0.916
12-Week Look-back period P1 P2 P3 P3-P1 1.378*** 0.870 1.631** 0.253 4.039 4.760 6.501 6.616 0.341 0.183 0.251 0.038 0.956 0.409 1.129 0.173
50
4-Week Look-back period P1 P2 P3 P3-P1 1.044** 1.096 1.802*** 0.758 4.453 6.411 5.889 7.394 0.234 0.171 0.306 0.102 0.761 0.199 1.371** 0.616
Table 7. Traded Volume and Macroeconomic Factors This table shows the estimates from a random-coefficient panel regression model whereby the dependent variable is the one-period change in the average weekly traded volume in U.S. dollars and the independent variable is a set of aggregate macro-financial indicators. These indicators are defined as the first-order differences of the first principal components of the inflation swap rates (Inflation), the yield spreads (Yield Spread), the CDS premia as explained above (CDS), the VIX, and the changes in the USD REER, and realized returns on cryptocurrencies. The sample period is from April 2016 to September 2017, weekly. Robust standard errors are generated by 5,000 bootstrap iterations and reported in parenthesis. * p < 0.10, ** p < 0.05, *** p < 0.01
Panel A: Panel Regression Estimates Inflation
M1 9.056* (5.222)
Yield Spread
M2
M3
M4
M5
16.62* (8.673)
CDS
0.120 (0.168)
VIX
0.707*** (0.229)
USD REER
-2.278 (4.568)
Returns Obs. N Wald (χ2 )
M6
1063 14 56.29
1063 14 55.64
1063 14 59.26
1063 14 54.01
1063 14 53.01
0.424** (0.208) 1063 14 65.37
M7 1.246 (9.735) 15.03 (16.016) 0.0600 (0.227) 0.722*** (0.233) -3.008 (4.748)
1063 14 65.13
M8 1.588 (9.539) 13.96 (15.683) 0.0948 (0.223) 0.674*** (0.236) -3.043 (4.819) 0.339** (0.167) 1063 14 77.58
Table 8. Traded Volume and Past Returns This table shows the results of a panel Granger causality test between traded volume and past returns using the methodology proposed by Dumitrescu and Hurlin (2012). This test allows to test for causality in heterogeneous panel data models whereby the cross-sectional heterogeneity of the causal relationships between volume and volatility is explicitly acknowledged. The table shows the test results for different lags of the autoregressive parameters and the regression slopes as in Equation (2). The sample period is from April 2016 to September 2017, weekly.
Panel A: Panel Granger Causality Test 1 H0 : Volume does not Granger-cause returns z-stat -0.041 p-value 0.966
2 -1.044 0.296
3 -0.237 0.812
# Lags 4 5 -0.469 -0.734 0.638 0.465
6 -0.927 0.353
7 -1.112 0.263
8 -0.400 0.689
H0 : Returns do not Granger-cause volume
2.601 0.009
3.289 0.001
3.689 0.000
3.457 0.000
3.964 0.000
2.551 0.010
z-stat p-value
1.468 0.142
51
3.267 0.001
Table B.1. Robustness: Cryptocurrencies and Other Asset Classes This table shows the estimates of a random-intercept panel regression model which which restricts the slope parameters to be constant across units. The set of cryptos used in the regression is outlined in Section 3 and detailed in Appendix A. The set of weekly returns on other asset classes comprises the FTSE Global All-Cap index (Equity), the BofA-Merrill Lynch Global Broad Market Index (Bond), the MSCI World REITs index (Real Estate), the S&P GSCI Energy index (Energy), the S&P GSCI Gold index (Gold), the S&P 500 VIX Short-Term Futures Index (Volatility), and the log-change of the nominal effective exchange rate of the U.S. dollar (FX) which represents a trade-weighted average of nominal bilateral rates between the USD and a basket of foreign currencies. Lagged returns are included in addition to past changes in the average traded volume as control variables. The sample period is from April 2016 to September 2017, weekly. Robust standard errors are reported in parenthesis. * p < 0.10, ** p < 0.05, *** p < 0.01
Panel A: Panel Regression Estimates (Random Effects) Equity
M1 0.244 (0.714)
Bond
M2
M3
M4
M5
M6
M8
0.531 (0.695)
Real Estate
0.947 (0.648)
Energy
-0.368*** (0.098)
Gold
-0.362*** (0.098) 0.806*** (0.283)
0.851*** (0.282)
FX
-1.075 (0.765)
Volatility Obs. Groups R2
M7
1063 14 0.020
1063 14 0.027
1063 14 0.020
1063 14 0.016
1063 14 0.017
52
1063 14 0.016
0.00436 (0.034) 1063 14 0.018
1063 14 0.021
M9 -0.614 (0.684) -4.688 (3.235) 0.789 (0.794) -0.948*** (0.205) 0.794* (0.419) -3.362 (2.992) -0.0435 (0.037) 1063 14 0.093
Table B.2. Robustness: Volatility Spillovers and Volume This table shows the estimates of a random-intercept panel regression model which restricts the slope parameters to be constant across groups to investigate contemporaneous correlations in the volatilities of cryptocurrencies and other asset classes conditionally on the average weekly traded volume. Realized volatility for both cryptocurrencies and other asset classes is computed as the square root of the sum of daily squared demeaned returns. Lagged values of the realized volatilities are included in addition to a time trend and past changes in the average traded volume as control variables. The sample period is from April 2016 to September 2017, weekly. Robust standard errors are reported in parenthesis. * p < 0.10, ** p < 0.05, *** p < 0.01
Panel A: Panel Regression Estimates (Random Effects) Equity
M1 0.555 (0.614)
Bond
M2
M3
M4
M5
M6
M7
M9
0.0554 (1.217)
Real Estate
-0.124 (0.393)
Energy
0.303*** (0.145)
Gold
0.609*** (0.237) -1.033 (1.381)
-0.549 (0.347)
FX
0.537 (1.166)
Volatility
-0.032 (0.045)
Volume Obs. Groups R2
M8
1063 14 0.170
1063 14 0.183
1063 14 0.168
1063 14 0.173
1063 14 0.168
53
1063 14 0.170
1063 14 0.174
0.068*** (0.008) 1063 14 0.320
1063 14 0.176
M10 1.035 (0.866) 1.443 (1.957) -0.183 (0.781) 0.876** (0.402) -0.459 (0.578) 1.993 (1.136) -0.194 (0.151) 0.069*** (0.008) 1063 14 0.355
Table B.3. Robustness: Traded Volume and Macroeconomic Factors This table shows the estimates from a random-intercept panel regression model whereby the slope parameter is restricted to be constant across groups. The dependent variable is the one-period change in the average weekly traded volume in U.S. dollars and the independent variable is a set of aggregate macro-financial indicators. These indicators are defined as the first-order differences of the first principal components of the inflation swap rates (Inflation), the yield spreads (Yield Spread), the CDS premia as explained above (CDS), the VIX, and the changes in the USD REER, and realized returns on cryptocurrencies. The sample period is from April 2016 to September 2017, weekly. Robust standard errors are reported in parenthesis. * p < 0.10, ** p < 0.05, *** p < 0.01
Panel A: Panel Regression Estimates (Random Effects) Inflation
M1 9.097*** (1.804)
Yield Spread
M2
M3
M4
M5
16.61*** (5.497)
CDS
0.136* (0.073)
VIX
0.667*** (0.143)
USD EER
-2.490 (2.934)
Returns Obs. N R2
M6
1063 14 0.136
1063 14 0.138
1063 14 0.135
1063 14 0.144
54
1063 14 0.134
0.453*** (0.107) 1063 14 0.143
M7 1.454 (6.626) 14.82 (12.692) 0.0884 (0.151) 0.686*** (0.146) -3.231 (3.100)
1063 14 0.150
M8 0.269 (6.235) 15.15 (12.127) 0.112 (0.147) 0.641*** (0.151) -3.096 (3.102) 0.413*** (0.116) 1063 14 0.157
Figure 1. Market Capitalization and Traded Volume
01apr2016
5 4 3 2 0
1
Average Weekly Traded Volume (USD bl)
100 50 0
Market Capitalization (USD bl)
150
This figure shows the market capitalization (left panel) and the average weekly traded volume (right panel), both expressed in billions of U.S. dollars. Both indicators are computed by using those currencies in the top 20th percentile of the cross-sectional distribution of market capitalization at the end of the sample and for which the average traded volume in U.S. dollars is non-zero. The sample period is from April 2016 to September 2017, weekly.
01oct2016
01apr2017
01oct2017
01apr2016
(a) Aggregate Market Capitalization
01oct2016
01apr2017
01oct2017
(b) Aggregate Average Traded Volume
Figure 2. Bitcoin Market Capitalization and Traded Volume as a Fraction of the Total
01apr2016
01oct2016
01apr2017
01oct2017
1 .8 .6 .4 .2
.8 .7 .6 .5 .4
Fraction of Mkt Cap from BTC
.9
Fraction of Avg. Weekly Traded Volume from BTC
This figure shows the market capitalization (left panel) and the average weekly traded volume (right panel) of Bitcoin as a fraction of the total market capitalization and averaged traded volume, respectively. Both indicators are computed by using those currencies in the top 20th percentile of the cross-sectional distribution of market capitalization at the end of the sample and for which the average traded volume in U.S. dollars is non-zero. The sample period is from April 2016 to September 2017, weekly.
01apr2016
(a) Market Capitalization of BTC (% total)
01oct2016
01apr2017
01oct2017
(b) Average Traded Volume of BTC (% total)
55
Figure 3. Bitcoin Trading by Exchanges and Currencies This figure shows the Bitcoin trading by fiat currency (top panel) and by exchange (bottom panel). Trading activity is defined by weekly traded volume expressed in Bitcoin to ensure comparability across platforms and for different currencies. The sample period is from April 2016 to September 2017, weekly.
1 0.9 CNY
Total Volume in BTC (%)
0.8
EUR GBP
0.7
HKD IDR
0.6
JPY PLN
0.5
RUB
0.4
USD Other
0.3 0.2 0.1 0 Jun16
Sep16
Dec16
Mar17
Jun17
Sep17
(a) Bitcoin Trading by Currency (%)
1 0.9
Total Volume in BTC (%)
0.8
bit-x bitfinex bitflyer bitchina coinbase huobi kraken lakebtc okcoin others
0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 Jun16
Sep16
Dec16
Mar17
(b) Bitcoin Trading by Exchanges (%)
56
Jun17
Sep17
Figure 4. Time-Varying Pairwise Correlations with Other Asset Classes This figure shows the time-varying correlation between a value-weighted market index for cryptocurrencies and other asset classes. The market index is constructed by using those currencies with a market capitalization in the top 20th percentile of the cross-sectional distribution at the end of the sample and for which the average traded volume in U.S. dollars is non-zero. The sample contains 36 cryptocurrencies which accounts for more than 95% of the total market capitalization as of September 2017. Top-left panel shows the conditional correlation with FTSE Global All-Cap index, which is a value-weighted index of large, mid and small cap stocks which covers both developed and emerging markets. Top-right panel shows the correlation of the crypto value-weighted market index and the Global Broad Market index provided by BofA-Merrill Lynch, which tracks the performance of investment grade public debt issued in the major domestic and Eurobond markets, including “global” bonds. Mid-left panel shows the conditional correlation with the returns on the MSCI World REITs index which represents a free float-adjusted value-weighted index of both large and mid caps across more than 20 developed markets. Mid-right panel shows the correlation between the returns on the value-weighted index of cryptocurrencies and returns from the S&P 500 VIX Short-Term Futures index which replicate a constant one-month rolling long position in first and second month VIX futures contracts. Finally, bottom-right and bottom-left panels show correlation between market returns on cryptocurrencies and the returns on the S&P GSCI Gold index, which tracks the COMEX gold futures, and the returns on the S&P GSCI Energy index, which includes crude oil (and supporting contracts) and natural gas, respectively. The sample period is from April 2016 to September 2017, weekly. The conditional correlation is computed by using a Dynamic conditional correlation model (see, e.g. Engle 2002). 0.1
0.3
Equity (Global All-Cap USD)
0.2
0
0.1
Correlation
Correlation
-0.1
-0.2
0 -0.1
-0.3 -0.2
Bond (Global Mkt Index) -0.4
-0.3
-0.5
-0.4 Jun16
Sep16
Dec16
Mar17
Jun17
Sep17
Jun16
Sep16
Dec16
Mar17
Jun17
Sep17
Sep16
Dec16
Mar17
Jun17
Sep17
Sep16
Dec16
Mar17
Jun17
Sep17
0.5
0.3
Real Estate Investment Trusts (Global) 0.4 0.2 0.3 0.2
Correlation
Correlation
0.1
0
0.1 0 -0.1
-0.1
-0.2 -0.2 -0.3 -0.3
Volatility (Equity)
-0.4 Jun16
Sep16
Dec16
Mar17
Jun17
Sep17
Jun16
0.05
0.4
0
0.3
-0.05 0.2
Correlation
Correlation
-0.1 -0.15 -0.2
0.1 0
-0.25 -0.1 -0.3
Gold
-0.35
-0.2
Commodity (Energy)
-0.4
-0.3 Jun16
Sep16
Dec16
Mar17
Jun17
Sep17
Jun16
57
Figure 5. Response of Volume to a Shock in Macroeconomic Factors
5 0 −5 −15
−10
Response of Volume to Yield Spread
15 10 5 0 −5
Response of Volume to Inflation
20
This figure shows the response of the average weekly traded volume in U.S. dollars to a one-unit shock to a macro-financial factor. Impulse-responses are estimated from a first-order panel Vector Autoregressive model where the vector of endogenous variables contains the one-period change in the average traded volume in U.S. dollars and the realized returns for each cryptocurrency, the first-order differences of the first principal components of the inflation swap rates, the yield spreads, the CDS premia as explained above, the VIX, and the changes in the USD REER. Top panels show the impulse response function (IRF) of traded volume to inflation (left) and the yield spread (right). Bottom panels show the response of average traded volume to a shock to CDS spreads (left) and the VIX (right). The dark blue line represents the average impulse response and the light-blue shaded area represents the 95% credibility intervals which are obtained by implementing a double non-parametric bootstrap scheme. The latter is a combination of temporal re-sampling and crosssectional re-sampling (see Kapetanios 2008 for more details). The sample period is from April 2016 to September 2017, weekly.
1
2
3
4
5
6
7
8
1
2
3
Forecast horizon
5
6
7
8
(b) Response of Volume to Yield Spread
10 5 0 −5
−10
−5
0
Response of Volume to VIX
15
5
(a) Response of Volume to Inflation
Response of Volume to CDS
4
Forecast horizon
1
2
3
4
5
6
7
8
1
Forecast horizon
2
3
4
5
6
7
Forecast horizon
(c) Response of Volume to CDS
(d) Response of Volume to VIX
58
8
Figure 6. Volume-Returns Relationship
1 0 −1
−5
Response of Volume on Returns 0 5 10
Response of Returns on Volume
2
15
This figure shows the response of the average weekly traded volume in U.S. dollars to a one-unit shock to a macro-financial factor. Impulse-responses are estimated from a first-order panel Vector Autoregressive model where the vector of endogenous variables contains the one-period change in the average traded volume in U.S. dollars and the realized returns for each cryptocurrency, the first-order differences of the first principal components of the inflation swap rates, the yield spreads, the CDS premia as explained above, the VIX, and the changes in the USD REER. Left panel shows the response of traded volume to a shock to past returns, while right panel shows the response of returns to a shock in average traded volume. The dark blue line represents the average impulse response and the light-blue shaded area represents the 95% credibility intervals which are obtained by implementing a double non-parametric bootstrap scheme. The latter is a combination of temporal re-sampling and cross-sectional re-sampling (see Kapetanios 2008 for more details). The sample period is from April 2016 to September 2017, weekly.
1
2
3
4 5 Forecast horizon
6
7
8
1
(a) Response of Volume to Returns
2
3
4 5 Forecast horizon
6
7
(b) Response of Returns to Volume
59
8
