Chapter 3 The Role of Investment Research in view of MiFID II An EmpiricalAnalysis of Information Asymmetry Idiosyncratic Risk and Liquidity
Notes This table shows summary statistics for four months of daily data surrounding the implemen-tation of MiFID II (October 2017 November 2017 February 2018 and March 2018) Analyst Coverageis proxied by the number of available earnings-per-share estimates per firm in the IBES database
European CountriesAT 2175 087 087 2 19 719BE 3480 140 227 1 31 658CZ 174 007 234 2 11 609DE 13875 557 792 1 40 1249DK 3247 130 922 1 32 1060ES 4894 197 1119 1 34 1415FI 5216 210 1328 1 28 804FR 14470 581 1909 1 32 1010GB 22322 897 2806 1 33 1214GR 1698 068 2874 1 17 601HU 348 014 2888 3 9 633IE 2926 118 3006 3 31 1364IT 7596 305 3311 1 29 755LU 1467 059 3370 2 25 1104MT 261 010 3380 1 5 267NL 5112 205 3586 1 31 1149PL 5099 205 3791 1 17 522PT 1131 045 3836 2 24 984SE 14525 583 4419 1 30 556
Chapter 3 The Role of Investment Research in view of MiFID II An EmpiricalAnalysis of Information Asymmetry Idiosyncratic Risk and Liquidity
37
TABLE 32 Summary Statistics for Analyst Coverage
Notes This table shows summary statistics about analyst coverage derived from four months of dailydata surrounding the implementation of MiFID II (October 2017 November 2017 February 2018 andMarch 2018) for 1646 US and 1281 EU firms
Jurisdiction EU USMean Std Dev Mean Std Dev
Analyst Coverage () 981 738 1090 779Quantile 1 ndash LOW 317 144 361 147Quantile 2 ndash MEDIUM 838 212 851 222Quantile 3 ndash HIGH 1899 491 1976 630
Firms () 1281 1646Firm-Day Observations () 110016 138920
TABLE 33 The Effect of MiFID II on European Financial Markets
Notes This table shows the difference-in-differences estimator for multiple variables of interest sur-rounding the implementation of MiFID II The control group consists of 1646 US firms (not affectedby MiFID II) and the treatment group consists of 1281 EU firms (affected by MiFID II) Illiquidity ismeasured by the illiquidity measure proposed by Amihud (2002) The estimator is multiplied by 106The estimator for the relative bid-ask spread is multiplied by 100 for readability
Analyst Coverage (ALL) (LOW) (MEDIUM) (HIGH)Mean of
Diff-in-Diffs(Treatmentsvs Controls)
t-statistic
Mean ofDiff-in-Diffs(Treatmentsvs Controls)
t-statistic
Mean ofDiff-in-Diffs(Treatmentsvs Controls)
t-statistic
Mean ofDiff-in-Diffs(Treatmentsvs Controls)
t-statistic
Panel A Four Months (Mminus2 vs M+2)
Analyst Coverage minus003 minus007 016 120 003 015 minus009 minus017
Illiquidity 001lowastlowastlowast 424 003lowastlowastlowast 339 001lowast 187 000 009
log(Trading Volume) minus029lowastlowastlowast minus305 minus038lowastlowastlowast minus258 minus020 minus150 minus009 minus071
rel Bid-Ask Spread 007lowastlowastlowast 529 009lowastlowastlowast 316 008lowastlowastlowast 401 003lowastlowastlowast 345
rel Bid-Ask Spread(1) 006lowastlowastlowast 521 009lowastlowastlowast 319 007lowastlowastlowast 365 003lowastlowastlowast 365
Idiosyncratic Risk (Ψ) 170lowastlowastlowast 1682 165lowastlowastlowast 924 145lowastlowastlowast 840 193lowastlowastlowast 1142
Panel B Six Months (Mminus3 vs M+3)
Analyst Coverage minus034 minus084 minus010 minus071 minus018 minus085 minus041 minus077
Illiquidity 001lowastlowast 241 001lowast 170 000 090 000 147
log(Trading Volume) minus035lowastlowastlowast minus368 minus059lowastlowastlowast minus390 minus015 minus113 minus021lowast minus174
rel Bid-Ask Spread 003lowastlowast 250 006lowastlowast 206 004lowastlowast 206 000 minus001
rel Bid-Ask Spread(1) 003lowastlowast 227 006lowastlowast 208 minus004lowast 192 000 minus043
Idiosyncratic Risk (Ψ) 153lowastlowastlowast 1409 322lowastlowastlowast 3298 143lowastlowastlowast 766 185lowastlowastlowast 1036lowast p lt 010 lowastlowast p lt 005 lowastlowastlowast p lt 001 (1)Including a control variable for liquidity
Chapter 3 The Role of Investment Research in view of MiFID II An EmpiricalAnalysis of Information Asymmetry Idiosyncratic Risk and Liquidity
38
TABLE 34 The Impact of Analyst Coverage on Information Asymmetry
Notes This table shows fixed-effects regression results on log(Spread) for the whole sample (ALL)and three sub-samples based on quantiles of analyst coverage (LOW MEDIUM HIGH) We measureIlliquidity by computing the illiquidity measure proposed by Amihud (2002) Market Capitalizationis measured in US Dollar
log(Spread)Analyst Coverage (ALL) (LOW) (MEDIUM) (HIGH)
Variables of Interest
log(Analyst Coverage) minus026lowastlowastlowast minus012lowastlowastlowast minus025lowastlowastlowast minus042lowastlowastlowast
(minus6510) (minus1056) (minus1248) (minus2262)
Illiquidity 170lowastlowastlowast 192lowastlowastlowast 163lowastlowastlowast 222lowastlowastlowast
(3791) (3612) (1402) (805)
Market Capitalization 000lowastlowastlowast minus000 minus000lowastlowastlowast 000lowastlowastlowast
(1026) (minus103) (minus2388) (2127)
MiFID IItimeslog(Analyst Coverage) minus003lowastlowastlowast 008lowastlowastlowast minus014lowastlowastlowast minus116lowastlowastlowast
(minus337) (383) (minus302) (minus2634)
Control Variables
MiFID II 002 minus007lowastlowast 032lowastlowastlowast 322lowastlowastlowast
(128) (minus232) (329) (2498)
Year [2017=02018=1] 006lowastlowastlowast 005lowastlowastlowast 005lowastlowastlowast 008lowastlowastlowast
(823) (381) (437) (655)
Country [US=0EU=1] 165lowastlowastlowast 156lowastlowastlowast 168lowastlowastlowast 171lowastlowastlowast
(21319) (11015) (12855) (13415)
Constant minus350lowastlowastlowast minus364lowastlowastlowast minus339lowastlowastlowast minus311lowastlowastlowast
(minus36544) (minus21935) (minus8116) (minus5722)
Observations 241598 80558 80277 80763R2 310 283 309 298
t statistics in parentheseslowast p lt 010 lowastlowast p lt 005 lowastlowastlowast p lt 001
Chapter 3 The Role of Investment Research in view of MiFID II An EmpiricalAnalysis of Information Asymmetry Idiosyncratic Risk and Liquidity
39
TABLE 35 The Impact of Analyst Coverage on Idiosyncratic Risk
Notes This table shows fixed-effects regression results on the idiosyncratic risk measure Ψ for thewhole sample (ALL) and three sub-samples based on quantiles of analyst coverage (LOW MEDIUMHIGH) We measure Illiquidity by computing the illiquidity measure proposed by Amihud (2002)Market Capitalization is measured in US Dollar
Relative Idiosyncratic Risk (Ψimonth)Analyst Coverage (ALL) (LOW) (MEDIUM) (HIGH)
Variables of Interest
log(Analyst Coverage) minus019lowastlowastlowast minus005lowastlowastlowast minus024lowastlowastlowast minus029lowastlowastlowast
(minus3143) (minus291) (minus785) (minus1031)
Illiquidity 112lowastlowastlowast 092lowastlowastlowast 097lowastlowastlowast minus059(1671) (1169) (553) (minus143)
Market Capitalization minus000lowastlowastlowast minus000lowastlowastlowast minus000lowastlowastlowast minus000lowastlowastlowast
(minus4001) (minus2991) (minus2823) (minus3339)
MiFID IItimeslog(Analyst Coverage) minus015lowastlowastlowast 003 009 minus053lowastlowastlowast
(minus1289) (099) (135) (minus820)
Control Variables
MiFID II 190lowastlowastlowast 139lowastlowastlowast 124lowastlowastlowast 343lowastlowastlowast
(6799) (3127) (852) (1792)
Year [2017=02018=1] minus225lowastlowastlowast minus193lowastlowastlowast minus214lowastlowastlowast minus264lowastlowastlowast
(minus20820) (minus9294) (minus12162) (minus14879)
Country [US=0EU=1] 010lowastlowastlowast 028lowastlowastlowast 022lowastlowastlowast minus018lowastlowastlowast
(864) (1333) (1132) (969)
Constant 358lowastlowastlowast 339lowastlowastlowast 377lowastlowastlowast 407lowastlowastlowast
(25037) (13739) (5984) (5036)
Observations 242170 80742 80578 80850R2 215 162 206 264
t statistics in parentheseslowast p lt 010 lowastlowast p lt 005 lowastlowastlowast p lt 001
Chapter 3 The Role of Investment Research in view of MiFID II An EmpiricalAnalysis of Information Asymmetry Idiosyncratic Risk and Liquidity
40
TABLE 36 Medium-Term Development of Analyst Coverage since the Introduc-tion of MiFID II (2018 ndash 2019)
Notes This table shows the change in analyst coverage between March 2018 to March 2019 for EUstocks affected by MiFID II We measure the change in analyst coverage be computing the change inthe amount of available earnings-per-share estimates available in the IBES If the drop in coverageis based on the delisting of the respective stock we drop the observation from our sample
All Firms Low CoverageN = 1 271 N = 389 (Q025)
∆Coverage (2018rarr 2019) Total Total
ge +3 72 566 12 308+2 86 677 26 668+1 155 1220 74 19020 311 2447 163 4190minus1 275 2163 77 1979minus2 139 1093 13 334le minus3 192 1511 1 026No more Coverage 41 323 23 591
Total 1271 10000 389 10000
41
Chapter 4
Innovative Finanzierungen uumlberInitial Coin Offerings Strukturund bisherige Performance
Chapter 4 has been published as a journal articleHaumlfner David and Dirk Schiereck (2020) Innovative Finanzierungen uumlber InitialCoin Offerings Struktur und bisherige Performance in Zeitschrift fuumlr KMU undEntrepreneurship (ZfKE) 682 pp 73-98 ISSN 1860-4633DOI httpsdoiorg103790zfke68273
42
Chapter 5
Statistical Properties ofCryptocurrency Order Books
51 Introduction
Securities exchanges or multilateral trading platforms allow us to analyze the be-havior of interacting market participants In these financial markets all market par-ticipants trade the same asset and pursue the same goal (profit) which should in-centivize each market participant to act rational This setting allows to study notonly general economic and financial theories but also theories of individual humanbehavior By analyzing market microstructure data apparently universal laws havebeen discovered in the literature that seem to hold true independent of the tradedasset or the exchange (eg Bouchaud Meacutezard and Potters 2002 and Naeligs and Skjel-torp 2006) These statistical laws help to better describe the competitive behaviorof market participants and some of these laws have revealed noticeable similaritiesacross different assets time periods and regions (Potters and Bouchaud 2003 Cont2001 and Mantegna and Stanley 1999)In this study we investigate whether properties in established securities markets canalso be found in cryptocurrency markets Cryptocurrencies did not yet exist whensome of the statistical laws were first discovered and verifying these properties incryptocurrency markets would strongly boost their validity and the robustness ofthose characteristics adding important evidence to the existing microstructure liter-atureGroundbreaking work has been done by OrsquoHara (1997) who analzyes the devel-opment of microstructure theory and the evolution of the literature to that pointand more recently the impact of high-frequency trading on market microstructure(OrsquoHara 2015)Studying market microstructure adds to a deeper understanding of financial mar-kets in general and the real economy benefits from efficient markets in many wayseg risk sharing and an efficient capital allocation As market microstructure andthe market design directly affect market efficiency inefficient markets on the microlevel lead to higher transaction costs and imply that someone can earn money on
Chapter 5 Statistical Properties of Cryptocurrency Order Books 43
someone elses expense If a market is inefficient on a microscopic level direct andindirect transaction costs will inevitably increase Further an ideal price discoverymay be disturbed which can lead to gaps between the price and the perceived valueof an asset This again affects the real economy as an investor is only willing to holdan asset if he is confident about its value This is crucial for companies to be ableto use financial markets as a reliable source of capital These relations between themarket microstructure and the real economy are especially relevant for emergingcryptocurrency markets Recent developments in the acceptance of cryptocurren-cies as a source of capital for companies ndash eg via Initial Coin Offerings (ICOs) ndash orfor portfolio diversification makes analyzing cryptocurrencies on a micro level moreimportant than ever beforeThe remainder of this study is structured as follows In Section 52 and Section 53we first describe the operating principle of a Limit Order Book (LOB) and give abrief overview of the cryptocurrencies examined in this study Further a detaileddescription of the data collection process is provided and the data set used for thesubsequent empirical analysis is described and descriptive statistics are presentedIn Section 54 we reconstruct the limit order books of three different cryptocurren-cies from our data We compute the aggregated limit order book volume also re-ferred to as the slope of the order book As shown in our descriptive analysis wefind empirical evidence for volume peaks in the LOB at specific price levels relativeto the best price We hypothesize that these peaks do not appear at random but arerather caused by ldquolazyrdquo investors disregarding the granularity of the price grid Wetest the empirical significance of our observation and find these peaks to be statisti-cally highly significant and to appear across all examined LOBsWe further find that the slope of the LOB varies over time ie the aggregated LOBchanges its shape This finding raises the question whether information is hiddenin the slope of the order book We examine the explanatory power of the slope inSection 55 Similar to Naeligs and Skjeltorp (2006) who study stock LOBs we inves-tigate three groups of models using the slope of the order book to explain changesin prices trading volume and the correlation between price changes and tradingvolume Our results suggest that the slope of the LOB can explain changes in the de-pendent variables We further find evidence suggesting that limit orders placed faraway from the best price still carry price-relevant information Section 56 concludesand discusses implications of our findings
52 Structure of a Limit Order Book
An asset can only be traded when a buyer and a seller find to each other Whilethe buyer wants to pay as little as possible for the respective asset the sellers wantssell for the highest possible price Exchanges serve as a platform to enable trading
Chapter 5 Statistical Properties of Cryptocurrency Order Books 44
by helping potential buyers and sellers to find a trading partner Since the devel-opment of stock exchanges two alternative systems have emerged differing in theway of how trading is organized In quote-driven markets market makers will postbuy and sell quotes for which they are willing to trade an asset1 A market makeris responsible to provide the market with liquidity by matching incoming ordersagainst other orders or by buying or selling from his own inventory While marketmakers provide liquidity ndash which is beneficial to all market participants ndash quote-driven markets lack transparency as market participants do not know the identityof their counterpart2 The second way to organize trading on an exchange is via alimit order book Exchanges operating via a LOB are commonly referred to as order-driven markets In an order-driven market all valid buy and sell orders are listedin the (electronic) LOB which each market participant has access to providing fulltransparency to all market participants Order driven markets became increasinglypopular in the past and many of the largest stock exchanges currently operate withLOBs (eg NYSE Euronext Deutsche Boumlrse Nasdaq the Tokyo Stock Exchange andthe London Stock Exchange) The matchmaking of a LOB is based on a set of pre-defined mechanical rules and strictly follows a first-come first serve principle andmeets a price-time priority Further the order processing mechanics of LOBs is au-tomated and follows transparent rules On the contrary the flow of incoming ordersis solely based on traders internal decision making processesFrom a scientific point of view the arrival of limit orders in an order-driven marketis a complex field of study as the order placement is not bound to any rules andpurely stems from the decision of an individual to indicate the willingness to buyor sell a specific quantity of the respective asset during a specific point in time for aself-set price Thus studying limit order books yields the most detailed insight intodynamic behavior in financial markets As every decision by a market participant isdirectly linked to a potential financial loss or gain market participants are stronglyincentivized to act rational providing an ideal experimental setting to study eco-nomic behaviorFigure 51 depicts a stylized version of a LOB Each block on the demand side re-
sembles one unit of the traded asset For simplicity we assume that each limit orderrefers to exactly one unit of the traded asset We define the best bid b(t) as the high-est price level at which at least one market participant is willing to buy one unit ofthe asset a(t) describes the lowest price for which at least one market participantis willing to sell In this steady state of the order book no trade would occur as
1While cryptocurrency markets are not order-driven and there are no market makers in Bitcoinmarkets (Marshall Nguyen and Visaltanachoti 2019) Holste and Gallus (2019) find empirical evi-dence for market maker quotes at cryptocurrency exchanges Market maker type of traders issuelimit orders with attractive prices however the volume of their offers is rather small and they shouldtherefore not be regarded as liquidity providers
2Anonymity is no distinct feature of quote-driven markets When trading is organized by operatinga limit order book the identity of the counterpart if often unknown to traders and only referred to bya unique number Solely the exchange operator knows about the true identity of the buyer and seller
Chapter 5 Statistical Properties of Cryptocurrency Order Books 45
-
-
-
-
-
-
-
-
-
-
-
MinimumTick Size
Demand(Bid)
Supply(Ask)
Bid-AskSpread
b(t)
a(t)
Price
FIGURE 51 Structure of a Limit Order Book Own Representation based on Preiset al (2006)
the lowest sell price is higher than the highest ask price Consequently a(t)minus b(t)describes the implicit bid-ask spread in this modelObserving the state of the LOB shown in Figure 51 imagine the case that a new mar-ket buy order arrives comprised of two units at a price level of a(t) + 1 This marketorder would be matched against the best limit sell order available in the order bookIn the current state of the order book there exists exactly one unit at a price level ofa(t) and two units at a price level of a(t) + 1 Thus the unit at price level a(t) will bematched against the incoming order The next best price is at a(t) + 1 However asthere are two units available it is not clear which one should be matched against theremaining unit from the market order In this case a time priority is met ie the limitsell order at price a(t) + 1 which has been submitted first will be matched againstthe market order Ceteris paribus this would lead to an increase in the bid-ask spreadby one tick as there is no volume left at a(t) after the market order has been matchedOne would describe the arrival of a market sell order in an analogous mannerIf a limit order arrives at a price level which can not directly be matched it will beadded to the LOB This scenario is shown for a limit buy order with a size of one unitat a price level of b(t)minus 2 in Figure 51 As there exist higher bids in the order bookthis order will not be executed immediately but rather added to the order book (in-dicated by the arrow) and remain in the order book until either all higher bid ordersare already matched or canceled and a sell order at price level b(t)minus 2 arrives or theorder is canceled by the trader who submitted the orderIt is important to note that besides the arrival of the market buy order or the arrival
Chapter 5 Statistical Properties of Cryptocurrency Order Books 46
of the limit buy order all subsequent processes are based on predefined mechani-cal rules The order book follows two simple principles to decide how an order ismatched price priority and time priority with the first dominating the latter Dueto the seemingly nondeterministic arrival of orders the volume at each price levelin the order book is nondeterministic as well and provides an interesting researchfield on its own Recent studies also discuss the endogeneity of the size of the bid-ask spread As shown above the bid-ask spread changes based on the arrival ofnew orders and is thus based on the behavior of the market participants Howeverwhen the size of the bid-ask spread almost always equals the minimum tick size ndasheg when the traded asset is very liquid ndash the question arises whether the bid-askspread is in fact perceived as exogenous rather than endogenous by market partici-pants (Biais Hillion and Spatt 1995) In this case the granularity of the price grid ndashwhich is set by the minimum tick size ndash may prevent the observability of the endo-geneity of the size of the bid-ask spreadWhile determining the price level of a limit order market participants have to taketwo antagonizing criteria into account Opting to place a limit order close to or atthe bid-ask spread leads first to a higher probability that the order will be fulfilledand secondly a decrease in the mean time passing until the order will be matchedOn the other hand placing a limit order close to the bid ask-spread yields the risk ofhaving the limit order executed at an unfavorable price abandoning the chance oftaking advantage of price movements towards the desired direction (eg the possi-bility to sell at a higher price or buy at a lower price) Thus every market participanthas to select a price level for his limit order which takes into account his individualprice and time preferences
53 Data Collection and Description
In order to analyze order book characteristics in cryptocurrency markets we col-lect trading data from one of the largest US-based cryptocurrency exchanges Theexchange generates a turnover of 1 bn USD and offers one of the largest onlinetrading platforms for cryptocurrencies to a peak of more than 10 mn worldwideusers At the time of data collection the cryptocrurrencies Bitcoin (BTC) BitcoinCash (BCH) Ethereum (ETH) and Litecoin (LTC) could be traded against each otherand against US Dollar (USD) and Euro (EUR)Bitcoin is a peer-to-peer electronic cash system first proposed by Nakamoto (2008)The main advantage of Bitcoin is the redundancy of intermediaries as transactionscan be made peer-to-peer and are tamper-proof as all transactions are recorded andstored in a blockchain Bitcoins can either be created via a process known as min-ing or purchased on an exchange While a traditional money transfer typically in-volves a third party (eg a bank) no such entity is necessary to transfer Bitcoin
Chapter 5 Statistical Properties of Cryptocurrency Order Books 47
In order to send Bitcoin from one party to another only a bitcoin wallet and an in-ternet connection is necessary In the Bitcoin network many miners define theexact structure of Bitcoin using an algorithm thus the currency is not controlled byone single authority This organizational structure leads to continuous majority de-cisions where computational power is used to vote Bitcoin is transparent as eachtransaction is stored and traceable in the blockchain A new transaction is verifiedonly if the majority of miners in the system confirms the transaction Although thetransaction history is visible the counterparties of a transaction remain anonymousas solely the address of the Bitcoin wallet of the sending and receiving entity canbe observed With approximately ten minutes the average transaction time in theBitcoin network is many times faster than a traditional bank transaction which stilltakes some business days to be completed However Bitcoin is currently not able tocompete with the speed of large credit card operatorsAs of September 2019 Bitcoin has a total market capitalization of approximately 189billion USD which is about 70 of the total cryptocurrency market The marketcapitalization constantly changes by either a change in the BTCUSD price or an in-crease in the available amount The maximum amount of Bitcoin is mathematicallylimited to 21 mn BTC of which approximately 18 mn BTC have been mined as of2019 Computing power needed to mine the remaining amount increases dynam-ically with the total computing power in the network Bitcoin has many possibleapplications as it can be used as means of payment a protection against inflation oras a value storage With a maximum of 21 mn BTC its limited availability makesBTC a scarce resource Further Bitcoin is decentralized and available through theinternet making it portable and tradeable in small units with low storage costs ndash onemajor advantage over storing values physically eg by buying goldA big problem of the Bitcoin network is the transaction speed and energy consump-tion Currently Bitcoin can only verify seven transactions per second With an in-crease in popularity more transactions need to be verified per second creating apotential bottleneck for the mainstream adoption and usability in everyday transac-tions Moreover one single Bitcoin transaction consumed at least 300 kWh in 2018(De Vries 2018 p 804) while a bank transfer by credit card only needs 0001 to 0002kWh In 2017 the cryptocurrency Bitcoin Cash (BCH) was introduced to tackle thisscaling challenge Bitcoin Cash was created by a hard fork of the Bitcoin blockchainmaking Bitcoin Cash technically almost identical to Bitcoin except for an increasedblock size of the blockchain resulting in a higher number of transactions that canbe verified per second While Bitcoin Cash is faster than Bitcoin larger blocks areharder to process favoring larger miners which is diametral to the initial Bitcoinconcept of decentralization Since its creation Bitcoin Cash developed towards anindependent cryptocurrency and is the fourth largest cryptocurrency in terms ofmarket capitalization (53 bn USD) as of September 2019
Chapter 5 Statistical Properties of Cryptocurrency Order Books 48
Similar to Bitcoin and Bitcoin Cash Ethereum is based on the blockchain technol-ogy However Ethereum acts as both a cryptocurrency and a decentralized comput-ing platform that allows developers to execute decentralized computer programsEthereum was introduced in 2015 by Vitalik Buterin and is used as a means of pay-ment in the Ethereum network As of September 2019 Ethereum is the secondlargest cryptocurrency after Bitcoin with a market capitalization of approximately19 bn USD From a consumer perspective Ethereum can be used to do everythingthat is possible with fiat money Ethereum can be spent invested saved or it can betransferred to peers For companies Ethereum can be used to finance MampA activi-ties and other investment decisions without a financial intermediary Neither a banknor a payment processor is needed to transfer Ethereum ie bank fees for grantinga financing loan can be avoided completely by using Ethereum On a broader scaleEthereum has the potential to make the economy more efficient and increase produc-tivity as the present value of projects increases due to lower capital costs leading tomore profitable projects overallIn the Ethereum network the currency ETH is used to pay participating computersfor providing computational power ie Ether can also be mined While the finan-cial industry is assumed to be the primary user of the blockchain concept (Noferet al 2017) the Ethereum blockchain also receives increasing attention from moredistant industry sectors lately Eg a USD 30 mn real estate property was tok-enized with blockchain in Manhattan in 2018 (Wolfson 2018) The transaction wasbased on a theoretical Two Token Waterfall model proposed by Lippiatt and Oved(2018) utilizing the Ethereum blockchain demonstrating the wide range of possibleapplications of Ethereum The model provides a structural framework to tokenizereal assets and is based on two tokens representing debt and equity classes Con-sequently both classes combined represent the total capitalization of a transactionLippiatt and Oved (2018) state that this tokenized structure can increase liquidity ofreal assets The waterfall depicted in Figure 52 represents the flow of cash in the caseof liquidation of the tokenized asset From the flow of payments it appears that debttoken holders enjoy seniority Compared to traditional debt interest payments arenot paid on a recurring basis in this model but the accrued amount is paid at the timeof liquidation This benefits the equity token holder as fewer cash requirements arenecessary and the date of sale does not have to be predetermined allowing equityholders to sell during favorable market conditions In exchange for this flexibilityequity token holders must share their excess sales profit with debt token holders ona prenegotiated split Lippiatt and Oved (2018) utilize the distributed ledger tech-nology as a decentralized clearing house that stores all financial transactions andwhere tokens represent the ownership of real assets as smart contracts3 The authorsfurther show that the return profiles of the debt and equity token imply an underly-ing present value of the asset which would create arbitrage opportunities for traders
3In their paper Lippiatt and Oved (2018) select Ethereum smart contracts
Chapter 5 Statistical Properties of Cryptocurrency Order Books 49
Token AReplicates Debt
Token BReplicates Equity
Total Asset Capitalization
Cash Flow Distribution at Sale Event
A Interest
A Principal
A Excess B Excess
B Principal
FIGURE 52 Waterfall Model for Tokenized Real Assets Own representation basedon Lippiatt and Oved (2018)
Chapter 5 Statistical Properties of Cryptocurrency Order Books 50
if the tokens are not priced correctly Lippiatt and Oved (2018) claim that the tok-enization of real assets can reduce illiquidity The most obvious increase in liquidityis due to the simplification of trading and pricing of a real asset which creates a sec-ondary market for those alternative investments In accordance with Glosten andHarris (1988) the authors decompose the bid-ask spread and show that the asym-metric information component is reduced as a consequence of the transparency ofsmart contracts Smart contracts allow to see the supply and holdings of all marketparticipants leading to more educated investment decisions and thus reducing thelikelihood of asymmetric information (Lippiatt and Oved 2018) Further clearingcosts equal almost zero through the peer-to-peer transfer on the Ethereum networkMoreover a potential investor is not tied to either provide equity or debt but cancreate his individual riskreturn profile by blending debt and equity tokens of thesame asset making an investment attractive for a broader range of investors Whiletokenized real estate transactions are still often conducted in fiat currency in the endndash mainly due to a lack of investor acceptance ndash the real estate transaction discussedabove demonstrates how the transaction process for alternative assets is changingand how Ethereum differs from Bitcoin and Bitcoin Cash in itrsquos potential useDue to the volatile nature of the cryptocurrency market it is hard to derive a mean-ingful statement about the relevance of the above mentioned cryptocurrencies How-ever we find that the market share of Bitcoin Ethereum and Bitcoin Cash combinedremains somewhat stable over time and accounts for roughly 80 of the total cryp-tocurrency market capitalization as of September 2019 Bitcoin alone has a total mar-ket share of approximately 71 (Ethereum c 7 Bitcoin Cash c 2) Includingthese three major currencies in our analysis we are convinced to capture a represen-tative picture of the cryptocurrency market4
531 Data Acquisition Process
We use real-time market data updates for orders and trades provided by a websocketfeed which can be used to reconstruct a real-time order book While the websocketfeed is publicly available connections to it are rate-limited By creating a script weare able to store all updates received by the websocket feed locally Observing thebuilt-in Sequence Number we can assure that we do not miss any updates and infact can recreate the full order book at any point in time during our period of obser-vation Table 51 provides an example excerpt of the data that we are able to recordUpdates to the order book happen when the status of an order changes Type de-fines the event that occurs and is split into the four distinct categories receivedopen match or done Whenever a new order arrives it is first received andthen open in the LOB The order which is identified by its unique Order ID willremain in the order book until either the entity who created the order cancels it or
4Based on a market capitalization of BTC 1889 bn USD ETH 192 bn USD BCH 53 bn USD asof 04 September 2019
Chapter 5 Statistical Properties of Cryptocurrency Order Books 51
TABLE 51 Example of events occuring in a Cryptocurrency Order Book
Notes This table shows an excerpt of the development of the BitcoinEuro order book on 18th April2018 Note that all events shown in this table occur in just 0043 seconds If a new limit order is re-ceived it will remain open in the order book until it either gets canceled by the trader or matchedagainst another order Figures have been slightly simplified for readability eg the true Order ID iscomposed of 32 digits to guarantee uniqueness Sequence increases by 1 whenever a new event oc-curs and proves that no updates in the order book were missed We further have information on thebest bid and ask during each event
Type Order Side Price Size Time Sequence Order Remaining ReasonType ID Size
open sell 662877 215918320 0 0001received limit buy 652861 0001 215918323 1
open buy 652861 215918323 2 0001done sell 663627 215918336 3 0001 canceleddone sell 664527 215918356 4 0001 canceleddone sell 663777 215918360 5 0001 canceled
received limit buy 650761 0001 215918362 6open buy 650761 215918362 7 0001
received limit buy 651211 0001 215918363 8
the order is matched against another order The Order Type defines whether anorder has been issued as a market order (market) or a limit order (limit) Sideindicates from which side an order has been issued ie if someone wants to sell(ask side) or buy (bid side) some amount of the respective cryptocurrency Sizedefines the volume of the orderWe track the evolution of the order book for every currency pair combination whichcan be traded at the time of observation allowing us to obtain an immense amountof trading data (see Table 52) We are able to store and analyze more than 60 giga-byte worth of trading data for eleven different currency pairsFurther a Rest API is available which we use to compute periodic order book snap-shots While the Rest API is more bulky and slow it allows us to gather preprocesseddata which would be unfeasible to reconstruct from the websocket feed in a timelymanner We use the Rest API to capture a snapshot of the order book every ten min-utes We use this data to examine statistical properties of the average shape of theorder book
532 Descriptive Statistics
Table 52 gives an sample overview and some meta data about the gathered trad-ing data Our observation period spans from April 2018 to August 2019 (includinggaps) Due to hardware and software restrictions (eg forced reboots operatingsystem updates server connection losses and data storage limitations) we are notable to gather 100 of the daily order flow and therefore split the observation intomultiple sessions spread across the day
Chapter 5 Statistical Properties of Cryptocurrency Order Books 52
Table 53 provides a breakdown of order types and cancellation rates We find thatacross all currency pairs and order types the cancellation rate is remarkably highwith the BTCUSD sell side (9836) being the lowest overall cancellation rateMoreover we find that market orders are rare ie orders are almost exclusivelyissued as limit ordersHigh cancellation rates as depicted in Table 53 indicate the existence of high-frequencytrading (HFT) While a strict definition of HFT does not exist we find that the recentdefinition of high-frequency algorithmic trading as a subset of algorithmic tradingissued by the European Securities and Markets Authority (ESMA) under the rulesof MiFID II is a useful definition According to ESMA (European Commission 2014p36) HFT is mainly characterized by
bull an infrastructure intended to minimize network latencies (eg co-location)
bull the controlling of order initiation generation routing or execution by ma-chines without human interaction
bull high message intraday rates concerning orders quotes or cancellations
ESMA further states that HFT is characterized by a high daily portfolio turnovera high order-to-trade ratio and ending the trading day at or close to a flat position(European Commission 2014 p10)HFT is discussed controversial as the net economic benefit or loss remains unclearThe real economy benefits from HFT in many ways namely a higher liquidity anincreased trading and order book volume a reduction of the bid-ask spread and abetter price formation and execution coming along with an overall improved pricequality which ultimately decreases the cost of capital However enhanced HFTbears many risks eg an overload of the trading systems due to high order cancella-tion rates an increased price volatility and the potential for market manipulation Inaddition slower traders could stop trading suspecting that high-frequency tradersuse their informational advantage against slower traders to earn a profit off of themIn Germany HFT is regulated since 2013 by the high-frequency trading bill (Hochfre-quenzhandelsgesetz) which includes some major amendments in order to preventdangers and the misuse of HFT Notably in addition to improved system controlrisk control and transparency guidelines an order-to-trade ratio has been stipulatedwhich aims to limit the amount of updates a trading system is allowed to send to-wards an exchange The definition and measurement of the order-to-trade ratio hasto be provided in the stock exchange regulations (sect26a BoumlrsG)HFT regulation has also been discussed on an European level and responsibilities offirms engaging in HFT have also been defined in MiFID II to ensure market qualitynotably to store records of their trading systems for a minimum of five years and theimplementation of measures to prevent market distortionUnfortunately we can not infer the share of market participants engaging in HFT
Chapter 5 Statistical Properties of Cryptocurrency Order Books 53
andor algorithmic trading from our data as this number is even tough to measureat regulated stock exchanges5
Similar to Biais Hillion and Spatt (1995) we compute the conditional probabilitiesof LOB events Our results are presented in Table 54 The variation of conditionalprobabilities in each column of Table 54 indicates that order book events are notstatistically independent from previous order book events
54 Limit Order Book Characteristics in Cryptocurrency Mar-kets
541 The Shape of the Average Order Book
Incoming limit orders are stored in the LOB until they are either executed or can-celed The sum of the demanded or supplied quantity of the traded assets currentlyavailable in the LOB at a given price level defines the order book volume at thatprice level and represents the current queue size The sum over all price levels isnow referred to as the depth of the order book As incoming orders are determinedby market participants so is the volume of the order book at a given price levelIn Figure 52 the sum of the squares at each price level represents the volume of theorder book at that price levelOur analysis of LOBs provides insights into the microstructure of cryptocurrencymarkets and serves two major purposes First studying market behavior at cryp-tocurrency exchanges is crucial in order to understand whether cryptocurrenciescan extend the bandwidth of options for corporate finance by providing an alter-native way of raising capital Second identifying market behavior at cryptocur-rency exchanges analogous to those of security exchanges would indicate that majormarket characteristics are determined by intrinsic (economic) human behavior andare less determined by asset specific characteristics Thus studying cryptocurrencyLOBs is important for the fields of market microstructure but also yields valuableempirical insights for corporate and behavioral finance theory In this section westudy whether the behavior of market participants creates similar patterns in cryp-tocurrency limit order books as have been found in stock marketsExaminig stock markets Potters and Bouchaud (2003) argue that the shape of theaverage order book is not clear a priori While most of the incoming orders arrivein proximity to the current bid or ask price an order placed close to the currentprice has a larger probability to be executed and disappearing from the order bookBouchaud Meacutezard and Potters (2002) find the time-averaged shape of the limit or-der book to be characterized by a maximum distant to the current bid-ask spread
5Nonbinding estimations for the amount of HFT as a share of total trading activity mostly lie closeto the 50 range for stock exchanges Ultimately one may not abandon the fact that all algorithmictrading strategies act according to rules made by humans pursuing the goal to generate profit thusthey do not act independent and are just an extension of the human capabilities
Chapter 5 Statistical Properties of Cryptocurrency Order Books 54
TABLE 52 Sample Overview and Meta Statistics of Trading Data
Notes We recorded data from 18th of April 2018 until 31th of August 2019 of live order book updatesDuring this time frame we aggregate 60 GB of raw data tracking every event in the order book ofthe respective currency pair The deviation from Recorded (Days) and Full Data (Days) stems fromexogenous events eg forced operating system updates or a reboot of the server used for data storageWe exclude days where we are missing data due to those events Trading is allowed nonstop OneDay refers to a full 24h cycle rather than a traditional trading day which is dependent on the openinghours of the respective exchange BTC refers to Bitcoin BCH refers to Bitcoin Cash ETH refers toEthereum USD refers to US Dollar There are no records of currency pairs including BCH in April2018 as we were not able to receive live order book updates in this time frame We observe someinterrupts tracking the order flow since mid-2019 Upon closer inspection these interrupts do notfollow a pattern and appear to be unsystematic We tackle this data issue by controlling for monthlyfixed effects in our empirical analysis thus this occurrence does not raise any concerns regarding theunbiasedness of our results
Currency Pair Recorded Full Data Avg Filesize Avg No of Avg Length of of OF(Days) (Days) (MBDay) Sessions one Session Recorded
(per Day) (min) (per Day)
April 2018BTCUSD 12 11 262 24 19 31ETHUSD 12 11 239 27 18 33BCHUSD ndash ndash ndash ndash ndash ndash
May 2018BTCUSD 31 30 252 23 37 57ETHUSD 31 30 251 28 14 27BCHUSD 11 10 233 34 3 8
June 2018BTCUSD 22 20 233 24 26 40ETHUSD 24 23 244 25 19 31BCHUSD 23 22 237 23 45 41
July 2018BTCUSD ndash ndash ndash ndash ndash ndashETHUSD 6 5 105 20 7 9BCHUSD 13 13 80 25 2 3
February 2019BTCUSD 17 15 261 25 19 32ETHUSD 16 13 242 17 57 66BCHUSD 17 15 207 4 310 83
June 2019BTCUSD 7 7 72 13 4 3ETHUSD 7 6 71 13 10 9BCHUSD 8 8 71 13 8 7
July 2019BTCUSD 11 11 70 13 6 5ETHUSD 11 11 69 12 12 10BCHUSD 11 11 68 12 20 17
August 2019BTCUSD 17 17 69 12 8 7ETHUSD 17 17 68 10 19 14BCHUSD 17 17 68 9 36 25
Chapter 5 Statistical Properties of Cryptocurrency Order Books 55
TABLE 53 Breakdown of Order Types and Cancellation Rates
Notes This table provides conditional probabilities for specific events in the order book If a neworder arrives in the order book it is labeled as received and be either a market or a limit orderEach order remains in the order book until it is done which can be either due to a cancellationor because the order was matched (filled) The interpretation is shown on the following exam-ple The value 9958 in the first row and column (BTCUSD Panel A Sell Side Received Limit)is the probability of an incoming sell order in the BTCUSD order book to be a limit order ieP(Order Type = limit|Currency Pair = BTCUSD Type = received Side = sell) = 9958 DataSource Cryptocurrency limit order books obtained from April to August 2019 (see Table 52 for a dataoverview)
BTCUSD ETHUSD BCHUSD Mean
Panel A Sell Side
Received
Limit 9958 9976 9991 9975
Market 042 024 009 025
Done
Canceled 9836 9844 9881 9854
Filled 164 156 119 146
Panel B Buy Side
Received
Limit 9967 9979 9968 9971
Market 033 021 032 029
Done
Canceled 9922 9936 9877 9912
Filled 078 064 123 088
Chapter 5 Statistical Properties of Cryptocurrency Order Books 56
TAB
LE
54
Con
diti
onal
Prob
abili
ties
ofth
eO
rder
Type
inTi
me
tgiv
enth
eO
rder
Type
intminus
1
Not
eTh
eta
ble
show
sth
eco
ndit
iona
lpro
babi
litie
sof
spec
ific
orde
rty
pes
follo
win
gea
chot
her
inth
eor
der
book
Eac
hco
lum
nre
pres
ents
the
prob
abili
tyof
the
resp
ecti
veor
der
type
give
nth
eor
der
type
ofth
epr
evio
usor
der
equa
lsth
eor
der
type
ofth
ero
wT
heta
ble
can
bere
adas
follo
ws
The
cond
itio
nalp
roba
bilit
yth
ata
Smal
lBuy
Ord
eris
follo
wed
bya
Larg
eBu
yO
rder
is
Pro
b(L
arge
Buy t|S
mal
lBuy
tminus1)
=0
22
Th
esu
mof
each
row
equa
ls10
0
Not
eth
atth
esu
mof
each
colu
mn
does
not
give
the
unco
ndit
iona
lpro
babi
lity
ofth
ere
spec
tive
even
tSi
mila
rto
Biai
sH
illio
nan
dSp
att(
1995
)w
ede
fine
mar
keto
rder
sas
larg
eif
thei
rvo
lum
eex
ceed
sth
eex
isti
ngvo
lum
eat
the
best
bid
oras
kpr
ice
and
the
pric
eal
low
sth
eor
der
tocl
imb
thro
ugh
the
oppo
site
side
ofth
eor
der
book
Asm
allm
arke
tord
erdo
esno
tpos
sess
atle
asto
neof
thos
ech
arac
teri
stic
s
Tim
etminus
1Ti
me
tLa
rge
Buy
Smal
lBu
y
Buy
Ord
erin
side
Spre
ad
Buy
Ord
erat
Spre
ad
Buy
Ord
erbe
low
Spre
ad
Can
cele
dBu
yO
rder
Larg
eSe
llSm
all
Sell
Sell
Ord
erin
side
Spre
ad
Sell
Ord
erat
Spre
ad
Sell
Ord
erab
ove
Spre
ad
Can
cele
dSe
llO
rder
Larg
eBu
y1
13
258
12
86
213
14
43
354
50
06
010
2
13
045
15
11
135
8Sm
allB
uy0
22
258
4
09
224
21
76
332
40
05
029
0
72
169
12
60
205
2Bu
yO
rder
insi
deSp
read
005
0
11
676
5
68
174
550
58
001
0
04
235
0
47
720
9
29
Buy
Ord
erat
Spre
ad0
06
030
5
31
636
17
63
387
50
01
014
0
94
136
12
63
165
1Bu
yO
rder
belo
wSp
read
003
0
23
143
1
06
256
139
75
003
0
14
089
0
90
151
414
80
Can
cele
dBu
yO
rder
003
0
16
384
1
86
385
027
53
002
0
09
095
0
79
981
16
41
Larg
eSe
ll0
07
019
2
59
035
18
67
135
43
54
835
12
91
160
8
67
295
1Sm
allS
ell
004
0
48
070
2
10
182
824
51
073
5
53
827
2
20
119
425
23
Sell
Ord
erin
side
Spre
ad0
02
009
3
92
079
13
93
162
90
07
014
5
95
409
12
70
420
0Se
llO
rder
atSp
read
001
0
25
110
1
60
184
023
37
005
0
26
303
5
99
146
631
30
Sell
Ord
erab
ove
Spre
ad0
04
024
1
12
118
24
18
220
40
03
016
0
93
117
24
87
240
3C
ance
led
Sell
Ord
er0
02
015
1
27
104
13
67
256
80
02
011
2
98
181
26
72
265
4
Chapter 5 Statistical Properties of Cryptocurrency Order Books 57
This finding is surprising as the probability of an incoming order to be placed is high-est at the best price The authors find this shape across different stocks listed at theParis Bourse In addition they find that the shape of the time-average order book isroughly symmetric between the bid and the ask side6 Potters and Bouchaud (2003)analyze the average shape of the order book of two exchange traded funds that trackthe NASDAQ and the SampP500 performance respectively confirming a maximum ofthe queue size lying away from the bid-ask spread for one of the ETFs For the otherETF however the queue has a maximum at the current bid-ask spread7
In order to buy a cryptocurrency (eg Bitcoin Bitcoin Cash or Ethereum) a buyerneeds to find someone to trade the respective currency for another currency eg USDollar Bringing together buyers and sellers of cryptocurrencies is the purpose ofcryptocurrency exchanges which operate a LOB following the same set of rules asstock exchanges in particular the price-time priority and the first-come first-servedprinciple In this section we restrict our work and empirical analysis to the LOBof three major cryptocurrencies Bitcoin (BTC) Bitcoin Cash (BCH) and Ethereum(ETH) all of which can be bought and sold for US Dollar at a cryptocurrency ex-change We exclusively consider US Dollar order books to retain a common denom-inator allowing to compare our empirical results across cryptocurrencies
542 The Bitcoin Order Book
Below we analyze characteristics of the BitcoinUS Dollar limit order book Ex-isting literature mainly focuses on the empirical shape of the order book of stocksWhile Biais Hillion and Spatt (1995) and Naeligs and Skjeltorp (2006) examine theslope of the aggregated order book and potential connections to volume and volatil-ity Bouchaud Meacutezard and Potters (2002) show that a snapshot of the order bookcan deviate substantially from its average shapeThe slope of the order book is derived by computing the gradient for the additionalsupplied or demanded volume for deeper levels in the order book Economicallythe slope of the order book is the elasticity partqpartp describing how quantity (q) pro-vided in the order book changes as a function of the price (p) (Naeligs and Skjeltorp2006 p 415)First we plot the empirical order book volume and the aggregated volume of the
BTCUSD LOB as a function of the absolute distance measured in ticks (∆) towardsthe best price8 The results are shown on a daily-average basis in Figure 53 andFigure 54 for the ask and bid side of the LOB respectively It can be seen that theavailable volume is highest directly at the bid-ask spread Both sides of the order
6These findings have been empirically confirmed by Zovko and Farmer (2002) and Mike andFarmer (2008)
7The authors argue that the observed deviation may be due to the order book of this ETF not beingthe dominant player at NASDAQ as Island ECN is just one of many trading platforms of NASDAQcovering only 20 of the total trading volume of this specific ETF (Potters and Bouchaud 2003 p136)
8One tick corresponds to one US Dollar cent in the BTCUSD LOB
Chapter 5 Statistical Properties of Cryptocurrency Order Books 58
Day 1Day 11
Day 21Day 31Day 41Day 51
05
10152025
0 20 40 60 80 100 120 140 160 180 200 220 240 260 280 300 320 340 Obersvation DayAsk Volu
me (in BTC)
ΔDay 1
Day 11
Day 21
Day 31
Day 41
Day 51
05
10152025303540
0 20 40 60 80 100 120 140 160 180 200 220 240 260 280 300 320 340
Observa
tion Day
Aggrega
ted Ask
Volume
(in BTC
)
Δ
Notes The figure shows both the average volume (left) and the average aggregated volume (right)dependent on the distance ∆ to the best ask price for the first 350 ticks on a daily basis across thesample period We derive the data by computing the average volume at each tick per day based onsnapshots of the order book taken every 10 minutes Accordingly each single data point in the abovechart represents the average value of 144 observations Most of the volume is available directly at thespread however a pattern of slightly higher volume manifests surrounding definite numbers (100 200and 300 ticks away from the best ask price) emerge
FIGURE 53 Ask Volume and Aggregated Ask Volume of the BTCUSD Limit Or-der Book relative to the Distance (∆) towards the best Ask Price
book seem to behave almost symmetrical however the aggregated volume at thebid side increases more steadily in ∆ Looking at the shape of the aggregated or-der book in Figure 53 and Figure 54 we observe substantial variation of the shapeacross observation days raising the question if there is information hidden in theslope of the order bookFor both the bid and ask side a higher average volume at multiples of 100 ticksaway from the best price can be observed This effect can be found consistentlyacross the observation period and is shown in more detail in Figure 55 A highaverage volume across the whole sample period can be observed at round figuresespecially for ∆s that are a multiple of 100 Preferences for round figures have beendocumented in the financial literature before notably Corwin (2003) finds that theunderpricing in seasoned equity offers is significantly related to underwriter pric-ing conventions such as price rounding and pricing relative to the bid quote Fur-ther links between numeric fluency and human preferences have been documentedby Kettle and Haumlubl (2010) and we are confident that the observed pattern in LOBvolumes can be linked to human preferences as well9
Next we examine the slope of the BTCUSD LOB The slope of the order bookrepresents the elasticity of the market supply and demand of the respective cryp-tocurrency We test how the aggregated order book volume increases in ∆ by com-paring the empirical fit of three alternative sets of models The first model assumes alinear relationship between aggregated order book volume and ∆ the second model
9We test the statistical significance of this finding in Section 545
Chapter 5 Statistical Properties of Cryptocurrency Order Books 59
Day 1Day 11
Day 21Day 31
Day 41Day 51
05
10152025
0 20 40 60 80 100 120 140 160 180 200 220 240 260 280 300 320 340 Observation D
ay
Bid Volume (
in BTC)
ΔDay 1
Day 11Day 21
Day 31Day 41
Day 51
0510152025303540
0 20 40 60 80 100 120 140 160 180 200 220 240 260 280 300 320 340
Observatio
n Day
Aggregated
Bid Volum
e (in BTC)
Δ
Notes The figure shows both the average volume (left) and the average aggregated volume (right)dependent on the distance ∆ to the best bid price for the first 350 ticks on a daily basis across thesample period We derive the data by computing the average volume at each tick per day based onsnapshots of the order book taken every 10 minutes Accordingly each single data point in the abovechart represents the average value of 144 observations Most of the volume is available directly at thespread however a pattern of slightly higher volume surrounding definite numbers (100 200 and 300ticks away from the best ask price) emerge
FIGURE 54 Bid Volume and Aggregated Bid Volume of the BTCUSD Limit OrderBook relative to the Distance (∆) towards the best Bid Price
assumes a logarithmic relation while the third model assumes a square root rela-tion10
1 Aggregated Average LOB VolumeBTCUSD(∆) = β0 + β1 times ∆
2 Aggregated Average LOB VolumeBTCUSD(∆) = β0 + β1 times ln(∆)
3 Aggregated Average LOB VolumeBTCUSD(∆) = β0 + β1 timesradic
∆
The empirical fit is presented in Table 55 We find that the logarithmic model per-forms worst for both the bid and ask side of the BTCUSD LOB Surprisingly ndash withan R2 of 9885 and 9888 ndash the square root specification beats the linear modelfor both the ask and bid side of the order book by 623 and 114 percentage pointsrespectively Considering only the most relevant part of the aggregated order book(∆ le 100) we find the square root model to still fit better than the linear modelhowever the difference in explanatory power between the two models diminishesConsidering only the first 100 ticks away from the current best price the squareroot specification yields an R2 which is 145 (ask side) and 086 (bid side) percentagepoints higher than in the linear modelOur results suggest that the slope of the BTCUSD LOB decreases when the distancetowards the best price increases This finding is valid for both sides of the orderbook Close to the best price however a steady slope appears to be a reasonable ap-proximation for the slope of the order book Our results also indicate that assuming
10Based on our graphical analysis (see Figure 53 and Figure 54) we do not test for other functionalforms However we test whether the functional form is concave or convex in Section 59 Note thatthe aggregated average LOB volume increases in ∆ by definition
Chapter 5 Statistical Properties of Cryptocurrency Order Books 60
a logarithmic relationship to describe the link between the aggregated average orderbook volume and ∆ is not feasible
Notes The left (right) figure shows the aggregated bid (ask) volume relative to the best bid (ask)price across the observation period for the BTCUSD limit order book for the first 2000 ticks Volumedirectly at the spread is omitted due to scaling Volume peaks occur at round numbers at both sides ofthe BTCUSD limit order book
FIGURE 55 Average Ask and Bid Volume of the BTCUSD Limit Order Book rel-ative to the Distance Delta (∆) towards the Best Bid or Ask Price
Chapter 5 Statistical Properties of Cryptocurrency Order Books 61
TABLE 55 Slope of the Aggregated Order Book for BTCUSD
Notes The table shows the slope of the average order BitcoinUS Dollar LOB across our sample pe-riod The slope is derived from a linear regression for different model specifications Model specifi-cations (11)-(32) consider the slope of the supply side Respectively Model specifications (13)-(34)show the empirical results for the demand side Each observation has a high level of confidence asit denotes the average aggregated volume of the order book ∆ ticks away from the best price Wemeasure the volume every ten minutes continuously across our sample period
Ask Side
Sample ALL ∆ le 100
Model (11) (21) (31) (12) (22) (32)
∆ 002lowastlowastlowast 002lowastlowastlowast
Std Err 0 0
ln(∆) 26022lowastlowastlowast 062lowastlowastlowast
Std Err 041 002radic
∆ 518lowastlowastlowast 026lowastlowastlowast
Std Err 0 0
Constant yes yes yes yes yes yes
N 50413 50412 50413 101 100 101
R2 9262 8892 9885 9598 8653 9743
Bid Side
Sample ALL ∆ le 100
Model (13) (23) (33) (14) (24) (34)
∆ 003lowastlowastlowast 002lowastlowastlowast
Std Err 0 0
ln(∆) 33207lowastlowastlowast 054lowastlowastlowast
Std Err 068 002radic
∆ 686lowastlowastlowast 023lowastlowastlowast
Std Err 0 0
Constant yes yes yes yes yes yes
N 50413 50412 50413 101 100 101
R2 9774 8263 9888 9700 8667 9786
p lt 010 p lt 005 p lt 001
Chapter 5 Statistical Properties of Cryptocurrency Order Books 62
543 The Bitcoin Cash Order Book
Figure 56 and Figure 57 show the average ask and bid volume and the aggregatedaverage ask and bid volume per observation day of the BCHUSD limit order bookWe find that the shape of the order book is rather symmetrical between the bid andask side Compared to the BTCUSD aggregated order book however more volumeseems to be located away from the bid-ask spread resulting in a steeper slope of theorder book Similar to the previous section we compute three different models toexplain the slope of the order book as a function of ∆ We use the method of ordinaryleast squares (OLS) to compute the fit of three different models imposing a linearlogarithmic or square-root relation between the increase in ∆ and the aggregatedaverage order book volume
1 Aggregated Average LOB VolumeBCHUSD(∆) = β0 + β1 times ∆
2 Aggregated Average LOB VolumeBCHUSD(∆) = β0 + β1 times ln(∆)
3 Aggregated Average LOB VolumeBCHUSD(∆) = β0 + β1 timesradic
∆
Our results are presented in Table 56 In line with our previous findings for theBTCUSD order book we find the logarithmic model to possess the worst explana-tory power Comparing the linear to the square-root model we find that the impliedrelation depends on the number of ticks taken into account While both models donot differ much in their explanatory power the square-root relation appears to betterexplain the increase in order book volume for both the bid and ask side if all levelsof ∆ are taken into account Considering only the first 100 ticks closest to the bid-askspread the linear model outperforms the square-root model by three (ask side) andfive (bid side) percentage points respectivelyMotivated by the volume pattern which emerged in the BTCUSD order book wecompute the average order book volume across observation days dependent on ∆(Figure 58) Again we find large peaks at round figures for ∆ especially at mul-tiples of 100 at both sides of the BCHUSD order book Surprisingly even largerpeaks emerge at 500 1000 and 1500 ticks away from the best price It is importantto note that these peaks represent average values and are unlikely to be outliers asthey can be observed consistently across our observation period We further findthat the average shape of the order book appears to have a maximum away fromthe bid-ask spread which also has been observed in stock markets by BouchaudMeacutezard and Potters (2002) In the BCHUSD LOB the average volume generallyincreases in ∆ between 0 and approximately 250 ticks before it slowly declines from250 ticks onward Bouchaud Meacutezard and Potters (2002) propose an analytical ap-proximation to compute the average order book concluding that [T]he shape of theaverage order book therefore reflects the competition between a power-law flow oflimit orders with a finite lifetime and the price dynamics that removes the orders
Chapter 5 Statistical Properties of Cryptocurrency Order Books 63
Day 1Day 4Day 7Day 10Day 13Day 16Day 19Day 220
4
8
12
16
20
24
28
32
0 20 40 60 80 100 120 140 160 180 200 220 240 260 280 300 320 340 Observation DayAsk Volume (
in BCH)
Δ Day 1Day 5Day 9Day 13Day 17Day 210
50
100
150
200
250
300
350
400
0 20 40 60 80 100 120 140 160 180 200 220 240 260 280 300 320 340 Observation Day
Aggregated Ask Vo
lume (in BCH)
ΔNotes The figure shows both the average volume (left) and the average aggregated volume (right)dependent on the distance ∆ to the best ask price for the first 350 ticks on a daily basis across thesample period We derive the data by computing the average volume at each tick per day based onsnapshots of the order book taken every 10 minutes Accordingly each single data point in the abovechart represents the average value of 144 observations Most of the volume is available directly at thespread however a pattern of slightly higher volume manifests surrounding definite numbers (100 200and 300 ticks away from the best ask price) emerge
FIGURE 56 Ask Volume and Aggregated Ask Volume of the BCHUSD LimitOrder Book relative to the Distance (∆) towards the best Ask Price
Day 1Day 4Day 7Day 10Day 13Day 16Day 190
2
4
6
8
10
12
14
16
18
20
22
24
0 20 40 60 80 100 120 140 160 180 200 220 240 260 280 300 320 340 Observation DayBid Volum
e (in BCH)
ΔDay 1Day 3Day 5Day 7Day 9Day 11Day 13Day 15Day 17Day 19Day 21
0
50
100
150
200
250
300
0 20 40 60 80 100 120 140 160 180 200 220 240 260 280 300 320 340 Observatio
n Day
Aggregated
Bid Volum
e (in BCH)
ΔNotes The figure shows both the average volume (left) and the average aggregated volume (right)dependent on the distance ∆ to the best bid price for the first 350 ticks on a daily basis across thesample period We derive the data by computing the average volume at each tick per day based onsnapshots of the order book taken every 10 minutes Accordingly each single data point in the abovechart represents the average value of 144 observations Most of the volume is available directly at thespread however a pattern of slightly higher volume manifests surrounding definite numbers (100 200and 300 ticks away from the best ask price) emerge
FIGURE 57 Bid Volume and Aggregated Bid Volume of the BCHUSD Limit OrderBook relative to the Distance (∆) towards the best Bid Price
Chapter 5 Statistical Properties of Cryptocurrency Order Books 64
close to the current price These effects lead to a universal shape which will presum-ably hold for many different markets[] (Bouchaud Meacutezard and Potters 2002p11) The authors further indicate that preliminary results show that the same be-havior can be observed in futures marketsWe provide empirical evidence that a hump away from the current mid point canalso be observed in cryptocurrency markets However Bouchaud Meacutezard and Pot-ters (2002) do not report volume peaks at round figures making a significance testfor our findings inevitable To analyze this phenomenon we perform statistical testsacross the cryptocurrency order books for Bitcoin Bitcoin Cash and Ethereum inSection 545
Notes The left (right) figure shows the aggregated bid (ask) volume relative to the best bid (ask) priceacross the observation period for the BCHUSD limit order book for the first 2000 ticks Volumedirectly at the spread is omitted due to scaling Volume peaks occur at round numbers at both sides ofthe BCHUSD limit order book
FIGURE 58 Average Ask and Bid Volume of the BCHUSD Limit Order Bookrelative to the Distance Delta (∆) towards the Best Bid or Ask Price
Chapter 5 Statistical Properties of Cryptocurrency Order Books 65
TABLE 56 Slope of the Aggregated Order Book for BCHUSD
Notes The table shows the slope of the average order Bitcoin CashUS Dollar LOB across our sampleperiod The slope is derived from a linear regression for different model specifications Model speci-fications (11)-(32) consider the slope of the supply side Respectively Model specifications (13)-(34)show the empirical results for the demand side Each observation has a high level of confidence asit denotes the average aggregated volume of the order book ∆ ticks away from the best price Wemeasure the volume every ten minutes continuously across our sample period
Ask Side
Sample ALL ∆ le 100
Model (11) (21) (31) (12) (22) (32)
∆ 018lowastlowastlowast 028lowastlowastlowast
Std Err 0 0
ln(∆) 77522 789lowastlowastlowast
Std Err 224 002radic
∆ 2828lowastlowastlowast 334lowastlowastlowast
Std Err 0 0
Constant yes yes yes yes yes yes
N 15178 15177 15178 101 100 101
R2 9574 8871 9992 9933 8198 9633
Bid Side
Sample ALL ∆ le 100
Model (13) (23) (33) (14) (24) (34)
∆ 023lowastlowastlowast 018lowastlowastlowast
Std Err 0 0
ln(∆) 92255lowastlowastlowast 512lowastlowastlowast
Std Err 321 027radic
∆ 3441lowastlowastlowast 217lowastlowastlowast
Std Err 002 005
Constant yes yes yes yes yes yes
N 15178 15177 15178 101 100 101
R2 9822 8450 9947 9943 7864 9436
p lt 010 p lt 005 p lt 001
Chapter 5 Statistical Properties of Cryptocurrency Order Books 66
544 The Ethereum Order Book
The internal cryptocurrency of Ethereum (ETH) can be traded on cryptocurrencyexchanges against fiat money or other cryptocurrencies In this section we focusspecifically on the ETHUSD limit order book to obtain comparability with the Bit-coin and Bitcoin Cash order book ie one tick (∆) resembles a buy or sell price oneUS Dollar cent lower or higher than the current best bid or ask price in the ETHUSDLOBFigure 59 and Figure 510 show the volume and the aggregated ask volume depen-dent on the distance towards the best price for the bid and ask side respectively Wefind both sides of the order book to behave almost symmetrical Interestingly vol-
Day 1Day 6
Day 11Day 16Day 21Day 26Day 31Day 36Day 41Day 46Day 51
020406080
100120140160180
0 20 40 60 80 100 120 140 160 180 200 220 240 260 280 300 320 340 Observati
on Day
Ask Volum
e (in ETH
)
Δ Day 1Day 4Day 7Day 10Day 13Day 16Day 19Day 22Day 25Day 28Day 31Day 34Day 37Day 40Day 43Day 46Day 49Day 52
05001000150020002500
0 20 40 60 80 100 120 140 160 180 200 220 240 260 280 300 320 340 Observation DayAggrega
ted Ask Volume
(in ETH)
ΔNotes The figure shows both the average volume (left) and the average aggregated volume (right)dependent on the distance ∆ to the best ask price for the first 350 ticks on a daily basis across thesample period We derive the data by computing the average volume at each tick per day based onsnapshots of the order book taken every 10 minutes Accordingly each single data point in the abovechart represents the average value of 144 observations Most of the volume is available directly at thespread however a pattern of slightly higher volume manifests surrounding definite numbers (100 200and 300 ticks away from the best ask price) emerge
FIGURE 59 Ask Volume and Aggregated Ask Volume of the ETHUSD Limit Or-der Book relative to the Distance (∆) towards the best Ask Price
ume peaks at round figures seem to be omnipresent in the ETHUSD order booksupporting our hypothesis that this observation follows a pattern and is also observ-able across different assets To get a better picture of the phenomenon we computethe average volume at each ∆ across the observation period (Figure 511) For bothsides of the order book we find the available volume to be of magnitudes higherwhen ∆ is a multiple of 100 than at ∆s surrounding those numbers We also find thatndash analogous to the Bitcoin Cash order bookndash the average shape of the ETHUSDorder book appears to have a maximum away from the best bid or ask price oncemore supporting the applicability of the model proposed by Bouchaud Meacutezard andPotters (2002) in a different marketWe further find that the volume peaks seem to mimic the average shape of the orderbook ie only considering the volume at multiples of 100 we also find a maximumaway from the best ask or bid price This is especially observable for the ask side of
Chapter 5 Statistical Properties of Cryptocurrency Order Books 67
Day 1Day 6Day 11Day 16Day 21Day 26Day 31Day 36Day 41Day 46Day 51
020406080100120140
0 20 40 60 80 100 120 140 160 180 200 220 240 260 280 300 320 340 Observation DayBid Volu
me (in ETH)
Δ Day 1Day 4Day 7Day 10Day 13Day 16Day 19Day 22Day 25Day 28Day 31Day 34Day 37Day 40Day 43Day 46Day 49Day 52
05001000150020002500
0 20 40 60 80 100 120 140 160 180 200 220 240 260 280 300 320 340 Observation DayAggregate
d Bid Volume (in ETH
)Δ
Notes The figure shows both the average volume (left) and the average aggregated volume (right)dependent on the distance ∆ to the best bid price for the first 350 ticks on a daily basis across thesample period We derive the data by computing the average volume at each tick per day based onsnapshots of the order book taken every 10 minutes Accordingly each single data point in the abovechart represents the average value of 144 observations Most of the volume is available directly at thespread however a pattern of slightly higher volume manifests surrounding definite numbers (100 200and 300 ticks away from the best ask price) emerge
FIGURE 510 Bid Volume and Aggregated Bid Volume of the ETHUSD LimitOrder Book relative to the Distance (∆) towards the best Bid Price
Notes The left (right) figure shows the aggregated bid (ask) volume relative to the best bid (ask)price across the observation period for the ETHUSD limit order book for the first 2000 ticks Volumedirectly at the spread is omitted due to scaling Volume peaks occur at round numbers at both sides ofthe ETHUSD limit order book
FIGURE 511 Average Ask and Bid Volume of the ETHUSD Limit Order Bookrelative to the Distance Delta (∆) towards the Best Bid or Ask Price
Chapter 5 Statistical Properties of Cryptocurrency Order Books 68
the ETHUSD order book (Figure 511) However we find it difficult to explain thisobservation A behavioral explanation could be that there exists a group of lazyinvestors that does not care about incremental tick sizes and only considers a lessgranular price grid but still places orders based on a trade off between executionprobability and time priority We discuss and test this hypothesis in Section 545Finally we take a look at the slope of the ETHUSD order book Analogous to theanalysis of the BTCUSD and the BCHUSD LOB we compare the fit of three mod-els supposing three different functional forms of the relation between the averageaggregate order book volume and ∆
1 Aggregated Average LOB VolumeETHUSD(∆) = β0 + β1 times ∆
2 Aggregated Average LOB VolumeETHUSD(∆) = β0 + β1 times ln(∆)
3 Aggregated Average LOB VolumeETHUSD(∆) = β0 + β1 timesradic
∆
Similar to the Bitcoin and Bitcoin Cash LOB the logarithmic model yields the worstfit while the linear model has the highest explanatory power across all samples Themediocre performance of both the logarithmic and square-root model hints that theslope of the ETHUSD LOB might be better explained by a convex function11
545 The Lazy Investor Hypothesis
Based on the discovery of unusual high order book volume at specific ∆s away fromthe best price which we find consistently across all observed cryptocurrency LOBsduring the graphical analysis we formulate the following null hypothesis
The Lazy Investor Hypothesis There exists a group of cryptocurrency investorswhich disregard the full granularity of the price grid leading to a higher averagelimit order book volume at ∆ ticks away from the best price if ∆ is a multiple of 100
To our knowledge this anomaly has not yet been examined in the literature beforeWe test the statistical significance by constructing a binary variable I∆ which equalsone if the distance towards the best price is a multiple of 100 and zero otherwiseWe compute the coefficient of I∆ using OLS based on Equation 51
Limit Order Book Volume∆ = β0 + β1∆ + β2 I∆ + ε where (51)
I∆ =
1 i f ∆ = 100 mod(0)0 else
To take a linear decline of the average order book volume into account we include∆ as a control variable in our regression model If there exist lazy investors our
11We allow for a convex or concave shape and directly compare the Ethereum Bitcoin and BitcoinCash LOB in Section 546
Chapter 5 Statistical Properties of Cryptocurrency Order Books 69
TABLE 57 Slope of the Aggregated Order Book for ETHUSD
Notes The table shows the slope of the average order EthereumUS Dollar LOB across our sampleperiod The slope is derived from a linear regression for different model specifications Model speci-fications (11)-(32) consider the slope of the supply side Respectively Model specifications (13)-(34)show the empirical results for the demand side Each observation has a high level of confidence asit denotes the average aggregated volume of the order book ∆ ticks away from the best price Wemeasure the volume every ten minutes continuously across our sample period
Ask Side
Sample ALL ∆ le 100
Model (11) (21) (31) (12) (22) (32)
∆ 117lowastlowastlowast 207lowastlowastlowast
Std Err 0 0
ln(∆) 1090141 5621lowastlowastlowast
Std Err 2786 328radic
∆ 27040lowastlowastlowast 2410lowastlowastlowast
Std Err 018 071
Constant yes yes yes yes yes yes
N 35395 35394 35395 101 100 101
R2 9861 8122 9844 9928 7496 9214
Bid Side
Sample ALL ∆ le 100
Model (13) (23) (33) (14) (24) (34)
∆ 220lowastlowastlowast 154lowastlowastlowast
Std Err 0 002
ln(∆) 1981112lowastlowastlowast 4124lowastlowastlowast
Std Err 5726 260radic
∆ 50071lowastlowastlowast 1782lowastlowastlowast
Std Err 046 0
Constant yes yes yes yes yes yes
N 35395 35394 35395 101 100 101
R2 9981 7718 9712 9849 7191 8995
p lt 010 p lt 005 p lt 001
Chapter 5 Statistical Properties of Cryptocurrency Order Books 70
estimated coefficient for β2 should deviate significantly from zero The results foreach cryptocurrency are presented in Table 58 We find that the coefficient for ourconstructed variable is indeed highly significant for both sides of the BTCUSDthe BCHUSD and the ETHUSD LOB indicating that the LOB volume at specificlevels relative to the best price appears to be dependent on whether the level is amultitude of 100 ie a round figure
Key Finding 1 Based on our analysis of the limit order books of BTC ETH andBCH our findings support the lazy investor hypothesis The empirical results arehard to bring in line with the concept of rational agents yet the pattern could beexplained by a human preference for round figures dominating the desire to achievethe best possible price
The observed volume peaks also imply a market stabilizing purpose related to pricemovements If the price moves into the direction of a volume peak the peak acts asa dam that curtails heavy price movements eg the price can not drop lower than100 ticks of the current price without the volume at 100 ticks away from the currentprice being matched against incoming orders first12 Based on our findings futureattempts to explain the shape of the average order book should try to take this effectinto account
TABLE 58 Volume peaks in Cryptocurrency Limit Order Books
Notes This table shows the regression results for the average volume in the Bitcoin Bitcoin Cash andEthereum limit order book I∆ represents a dummy variable which equals one if the distance towardsthe best price (∆) can be divided by 100 without remainder
BTCUSD BCHUSD ETHUSD
Ask Side Bid Side Ask Side Bid Side Ask Side Bid Side
I∆ 028 035 158 151 811 901
Std Err 0 0 004 003 008 136
∆ yes yes yes yes yes yes
Constant yes yes yes yes yes yes
N 50412 50412 15177 15177 35394 35394
R2 3105 600 1205 1606 2358 279
p lt 010 p lt 005 p lt 001
12Note that this is a hypothetical case It is likely that some market participants will cancel their limitorder before the price reaches the volume peak Our finding would also induce that the probability ofprice movements of increasing magnitude does not decrease steadily
Chapter 5 Statistical Properties of Cryptocurrency Order Books 71
546 Convexity of the Order Book Slope
Motivated by our finding that the functional form of the demand and supply curvesslightly vary across cryptocurrencies (see Table 55 56 and 57) we test whethercryptocurrency order books have a concave or convex demand and supply curve13
Table 55 56 and 57 equally show that we can rule out a logarithmic relation Thuswe test for convexity by allowing the exponent to vary in a polynomial model whichcan be regarded as an extension to the proposed square-root relation model
Aggregated LOB Volume∆ = LOB Volume at the spread + ∆β (52)
By applying the natural logartihm to Formula 52 we can estimate β using the
Notes The upper left picture shows the theoretical aggregated order book volume implied by Equa-tion 52 dependent on the β-parameter The following pictures show the actual aggregated order bookvolume of the BTCUSD order book derived from our data Each data point at the respective tick (∆Delta) is calculated as the average order book volume at that tick derived from 10 minute snapshotsacross 94 observation days ie each data point is an average of roughly 13500 observations Con-sidering all tick levels (upper right picture) the total aggregated order book volume of the BTCUSDcurrency pair appears to be concave implying a β-parameter of less than one
FIGURE 512 ConvexityConcavity of the Aggregated BTCUSD Limit OrderBook
13The aggregated average order book volume of the ask (bid) side gives the available total volumeavailable at this price This means that the aggregated average volume can be interpreted as the de-mand (supply) curve of the respective cryptocurrency
Chapter 5 Statistical Properties of Cryptocurrency Order Books 72
method of OLS Consequently β lt 1 (β gt 1) indicates a concave (convex) demandor supply curve The results are presented in Table 59 and show that Bitcoin andBitcoin Cash demand and supply curves are concave while the Ethereum supplyand demand curve can be characterized as slightly convex The direction of our re-sults do not change when we restrict our sample to the first 100 ticks away from thebest price However it is worthwhile to note that the measured β-coefficients areclose to one justifying a linear approximation in empirical analysis without losingmuch of the explanatory power
Key Finding 2 We conclude that it is reasonable to approximate the slope of theorder book linearly
55 Informativeness of Order Book Characteristics
In the previous sections we mainly examine static limit order book characteristicsIn this section we are focusing on dynamic relations of the LOB and investigatea potential link between the slope of the order book and cryptocurrency returnsWe also hypothesize non-mechanic connections between the order book slope andtrading activityNaeligs and Skjeltorp (2006) find a negative link between the order book slope andvolume volatility and the correlation between volume and volatility in Norwegianstock markets and show that the slope can be regarded as a proxy for disagreementamong investors We follow the approach of Naeligs and Skjeltorp (2006) in that wealso investigate three groups of models relating to the order book slope (SLOPEit)
1 Price changeit = f (SLOPEit )
2 Number of tradesit = f (SLOPEit )
3 Correlation(Price changeit Number of tradesit) = f (SLOPEit )
The first set of models examines the relation between the price change of a cryptocur-rency and the slope of the order book In an efficient market price changes occurwhen new information about the value of the underlying assets emerge Observinga link between price changes and the slope of the order book indirectly suggests thatthe volume of the order book ie the aggregated supply and demand possibly con-tains information about the value of the underlying assetThe second set of models considers the relation between the number of trades andthe slope of the order book Each trader has to choose between buying directly orplacing a limit order We are interested in whether traders consider the existing vol-ume in the order book when placing a new order A link between the slope of theorder book and trading activity could shed light on the dilemma a trader faces whendeciding at what price level he should place a new order
Chapter 5 Statistical Properties of Cryptocurrency Order Books 73
TABLE 59 ConcavityConvexity of Aggregated Cryptocurrency Limit OrderBooks
Notes We test the concavity (convexity) of the average aggregated volume of of the Bit-coin Bitcoin Cash and Ethereum LOB across our sample period Our initial equation equalsAggregated LOB Volume∆ = LOB Volume at the spread + ∆β We estimate β using OLS and trans-forming the initial equation ln(Aggregated LOB Volume∆ minus Volume at the spread) = β lowast ln(∆) Conse-quently β lt 1 (β gt 1) indicates a concave (convex) demand or supply curve
Ask Side
Sample ALL ∆ le 100
Currency (USD) BTC BCH ETH BTC BCH ETH
β-Coeff 064 086 104 017 074 118
Std Err 0 0 0 001 0 001
Constant no no no no no no
N 50412 15177 14 35394 100 100 100
R2 9985 9994 9994 8853 9982 9950
p-Value (H0 β gt 1) lt001 lt001 gt099 lt001 lt001 gt099
Bid Side
Sample ALL ∆ le 100
Currency (USD) BTC BCH ETH BTC BCH ETH
β-Coeff 065 087 108 013 059 109
Std Err 0 0 0 001 001 001
Constant no no no no no no
N 50412 15177 35394 100 100 100
R2 9971 9996 9999 7166 9926 9950
p-Value (H0 β gt 1) lt001 lt001 gt099 lt001 lt001 gt099
p lt 010 p lt 005 p lt 001
Chapter 5 Statistical Properties of Cryptocurrency Order Books 74
The third set of models examines the link between the correlation between pricechange and the number of trades (volume-volatility relation) and the slope of theorder book The volume-volatility relationship is well documented for different fi-nancial markets including stocks currencies oil futures and other derivatives (Fos-ter 1995 Fung and Patterson 1999 Sarwar 2003 and Naeligs and Skjeltorp 2006) Therelation can be explained assuming that new information is not incorporated inprices directly but over time However the relation could also be explained in thecase that traders do not agree on the impact of new information14 Studying the im-pact of the order book slope on this relation could yield insight into what causes thevolume-volatility relation
551 Measuring the Slope of the Limit Order Book
We define the average aggregated LOB volume up to a price level of ∆ ticks awayfrom the best price for the ask side of the order book as
Avg aggr LOB Volume∆asktj =∆
sumi=1
(1N
N
sumn=1
VOLinasktj
)(53)
VOLinasktj gives the ask side LOB volume of currency j at a price level Ask price + iduring the n-th order book snapshot at day t As we capture a snapshot of the LOBevery ten minutes N amounts to 144 snapshots per dayIn order to create our slope variable we estimate the following regression modelusing OLS
ln(Avg aggr LOB Volume∆asktj) = β0 + βasktj times ln(Ask Price∆) (54)
βasktj is the estimated elasticity partqpartp of the aggregated LOB volume with respect to
price and resembles our slope measure SLOPEasktj A higher value of βasktj resem-bles a steeper order book We compute the β-parameter for both sides of the LOBfor each currency and each day The resulting slope measure can be represented inmatrix notation
SLOPEask =
βaskt=1BTC βaskt=1BCH βaskt=1ETH
βaskt=2BTC βaskt=2BCH βaskt=2ETH
(55)
We repeat the above steps analogous for the bid side to compute the average slope ofthe LOB We receive our final slope measure by computing the average daily slopeof the LOB
14This argumentation closely resembles the second interpretation of the widely used illiquidity mea-sure proposed by Amihud (2002)
Chapter 5 Statistical Properties of Cryptocurrency Order Books 75
SLOPE =SLOPEask + SLOPEbid
2(56)
We compute SLOPE considering two different sets of data SLOPE100 is calculatedusing only the first 100 ticks away from the best bid or ask price thus representingthe slope of the inner order book whereas SLOPE10000 includes price levels of upto 10000 ticks ndash ie up to 100 USD ndash away from the best bid or ask price capturinginformation potentially hidden in the depths of the order book (the deeper slope)
552 Order Book Slope and Volatility
In efficient markets price jumps occur through the arrival of new information thatimpacts the value of an asset If all investors agree on the impact of the new informa-tion price adjustments happen without trading Investors cancel existing limit or-ders and place new limit orders around the new equilibrium price level consideringthe updated information state However if there is no consensus among investorsabout the price impact of the new information trading takes place until all investorsthat do not agree on the price impact sold or bought their assets thereby moving theprice to itrsquos new equilibrium where consensus is reached againWe capture daily price jumps by computing the daily volatility of each cryptocur-rency according to the following formula
Volatilityit =
∣∣∣∣ LOB Mid pricet
LOB Mid pricetminus1minus 1∣∣∣∣ (57)
In general volatility is expected to be higher in less liquid assets as the supply anddemand side of the LOB is not thick enough to fulfill large market orders withoutexecuting limit orders at deeper price levels moving the mid price in consequenceSpontaneous demand or supply of large quantities lead to larger price changes inilliquid assets ie we would expect the SLOPE of illiquid assets to be more gentlea priori imposing a positive relationship between SLOPE and liquidity When in-vestigating the link between the order book slope and volatility we also control forilliquidity as illiquidity and volatility are likely correlatedWe include trading activity in our regression by deriving the number of trades perday based on Equation 58 We include a scaling factor in Equation 58 to make thenumber of trades comparable across days and cryptocurrencies (see Table 52)
N o f Tradesit = Recorded Tradesit times1440
Recorded Minutesit(58)
We also compute the average trade size per day for each cryptocurrency based onEquation 59
Trade Sizeit =Trading Volumeit
N o f Tradesit(59)
Chapter 5 Statistical Properties of Cryptocurrency Order Books 76
We further compute the average bid-ask spread (Spreadit) from our data and in-clude the market capitalization in itrsquos logarithmic form (ln(MCAP)it) The variablesln(MCAP)it and Spreadit are closely tied and proxy for liquidity We estimate thefollowing linear model
Volatilityit = β0 + β1SLOPE100it + β2SLOPE10000it + β3N o f Tradesit
+ β4Trade Sizeit + β5ln(MCAP)it + β6Spreadit + ci + uit (510)
Using fixed-effects regressions we get rid of time-constant unobserved effects ciTable 510 shows the results for six different model specifications We find that theslope of the inner order book has a significant positive effect on volatility in Model 1however this effect diminishes when controlling for the number and size of trades
Key Finding 3 We conclude that the slope of the inner LOB can not explain re-turn variation
Models 4-6 consider the LOB volume of the first 10000 ticks and reveal a positivelink between the deeper order book slope and return variation The effect remainssignificant at the 10 level across all three model specifications This finding indi-cates that a steeper order book slope can be associated with higher volatility Thisfinding seems to be contradictory at first as a steeper slope should prevent largeprice jumps as the LOB holds enough volume to serve large market orders withoutshifting prices too much However as we control for the trading activity and liq-uidity the volatility increase due to a steeper slope is likely caused by informationwhere investors agree upon the price impact and consequently adjust their activelimit orders
Key Finding 4 Our findings indicate that a steeper slope can be associated withmore non-trading volume ie investors adjust their limit orders based on new un-ambiguous information more actively when there is more LOB volume close to thespread
This finding could be explained by a lower execution probability of a limit orderceteris paribus when the slope is steep A steep slope indicates that a lot of the LOBvolume is centered around the spread reducing the execution probability of limitorders deeper in the LOB Traders observe this reduced execution probability andadjust their limit orders accordingly Through this channel limit orders deeper inthe LOB are still relevant for the price formation process even though they do notaffect the price mechanically
Chapter 5 Statistical Properties of Cryptocurrency Order Books 77
We further find that more trades and larger average trade size lead to higher volatil-ity The effect remains significant when controlling for ln(MCAP) and Spread Ar-guably trading activity is often interpreted as a sign of liquidity and we would ex-pect a decrease in volatility when trading activity is high However we controlfor illiquidity by including ln(MCAP) and Spread Our results hint a very nervouscryptocurrency market environment where many hectic trades are executed in ashort period of time leading to huge price jumps
TABLE 510 Volatility and the Slope of the Order Book
Notes This table shows the fixed effects regression results for Equation 510 Volatility is measured asthe daily absolute return of a cryptocurrency Coefficients of N of trades SLOPE100 and SLOPE10000are multiplied by 105 for better readability The coefficient of Trade size ln(MCAP) and spread is multi-plied by 103
Model (1) (2) (3) (4) (5) (6)
SLOPE100 047 036 040
(265) (157) (148)
SLOPE10000 952 943 950
(172) (171) (166)
N of trades 004 004 005 005
(386) (361) (520) (410)
Trade size 073 071 071 073
(404) (382) (405) (390)
ln(MCAP) -287 273
(minus031) (035)
Spread -163 -319
(minus010) (minus02)
Month fixed effects yes yes yes yes yes yes
N times T (currency-days) 231 186 186 231 186 186
R2 1120 2920 2920 960 2930 2940
p lt 010 p lt 005 p lt 001 t-statistics in parentheses
553 Order Book Slope and Trading Activity
In this section we examine the link between the order book slope and trading activ-ity measured by the number of trades per day We compute four different models
Chapter 5 Statistical Properties of Cryptocurrency Order Books 78
based on Equation 511 Our results are presented in Table 511
N o f Tradesit = β0 + β1SLOPE100it + β2SLOPE10000it + β3Trade Sizeit
+ β4ln(MCAP)it + β5Spreadit + ci + uit (511)
Key Finding 5 We find that the order book slope is significantly positive relatedto trading activity Interestingly this relation changes sign and becomes negativewhen considering the slope measure computed from the first 10000 ticks instead ofrestricting the data used for computing the slope to the first 100 ticks
With regard to causality we argue that it is reasonable to assume that an investorinspects the order book before placing an order and not afterwards assuming hisgoal is to make an educated trading decision If this is the case it is plausible thatthe slope of the order book affects trading activity and not vice versa Naeligs andSkjeltorp (2006) find the same astounding result in Norwegian stock markets andconclude that the order book slope contains different information based on the depthof the order book used to compute the slope Our result confirms their findings andshows that a sign change can be observed in cryptocurrency markets as well evenby using an alternative approach to calculate the slopeWe further find that the average trade size does not seem to influence trading ac-tivity and that the number of trades is higher when the spread increases A narrowbid-ask spread generally reduces trading costs which should amplify trading per seHowever upon closer inspection we find that the spread in our data equals the mini-mum tick size of 001 USD in 4202 of all observations whereas the average spreadis 050 USD indicating that when the spread deviates from the minimum tick sizethe deviation is quite large Such a change in the spread could be interpreted bymarket participants as a trading signal which would explain our empirical results
554 Order Book Slope and the Volume-Volatility Relation
Next we examine the interplay between the volume-volatility relation and the slopeof the order book The volume-volatility relation is a well known phenomenonobserved across many markets describing the empirical observation of high pricevolatility coupled with high trading volume which has been confirmed by a varietyof studies (see Karpoff 1987)In our data we find a positive correlation between volatility and daily trading vol-ume of 1919 and a positive correlation between volatility and the number of tradesof 2644 indicating that the volume-volatility relation can be observed in cryp-tocurrency markets as well We are interested in the cause of this relationship anddifferent theoretical models have been proposed mainly focusing on market effi-ciency the informedness of traders and speculative trading (Glosten and Harris
Chapter 5 Statistical Properties of Cryptocurrency Order Books 79
TABLE 511 Trading Activity and the Slope of the Order Book
Notes This table shows the fixed effects regression results for Equation 511 The number of trades (Nof trades) is the dependent variable
Model (1) (2) (3) (4)
SLOPE100 1203 416
(873) (223)
SLOPE10000 6792 -6736
(143) (minus171)
Trade size -4395 -12390
(minus034) (minus096)
ln(MCAP) 2416250 3267030
(403) (677)
Spread 453150 553640
(441) (545)
Month fixed effects yes yes yes yes
N times T (currency-days) 186 186 186 186
R2 5520 6340 3650 6290
p lt 010 p lt 005 p lt 001 t-statistics in parentheses
Chapter 5 Statistical Properties of Cryptocurrency Order Books 80
1988 Foster and Viswanathan 1995) We test whether the order book slope ie theavailability of potential trading volume can help to explain this relation Naeligs andSkjeltorp (2006) find that the slope of the order book is significantly negatively re-lated to the volume-volatility relation concluding that a stronger volume-volatilityrelation is associated with a flat slope of the order book However the authors do notdirectly interpret their results Similar to Naeligs and Skjeltorp (2006) we investigatethis relationship by computing the daily correlation coefficient Corr(N o f tradesit|Rit|) measured over a month between the number of trades and the absolute re-turn15 Naeligs and Skjeltorp (2006) argue that the number of trades is the crucialcomponent of trading volume Regressing Corr(N o f tradesit |Rit|) on our slopemeasure using a fixed-effects regression model we only find weak support for thenegative relationship between the volume-volatility relation and the slope of theorder book We compute the volume-volatility relation alternatively by directly in-corporating daily trading volume Corr(Trading Volumeit |Rit|) and estimate thefollowing regression model
Corr(Trading Volumeit |Rit|) = β0 + β1SLOPE100it + β2SLOPE10000it
+ β3Trade Sizeit + β4ln(MCAP)it + β5Spreadit + ci + uit (512)
Key Finding 6 Our results indicate that there is a significant amount of informationin the order book and the slope of the order book should be considered in theoreticalmodels trying to explain the cause of the volume-volatility relation We find that thevolume-volatility relation seems to be stronger when the slope of the order book issteeper which contradicts the results of Naeligs and Skjeltorp (2006)
The effect appears to be robust across model specifications and can be observed forboth slope measure specifications and increases in magnitude when controlling formarket capitalization Further the explanatory power in all models presented inTable 512 is approximately 60 and does not vary much between models It is note-worthy that the first and the fourth model already explain 6080 and 5900 of thevariation in the volume-volatility relation
555 A Note on Causality
While it is generally difficult to derive causality in economics it is especially chal-lenging in a market microstructure setting However some conjectures with respectto causality can be derived economically To further increase the robustness of ourresults we perform additional causality tests on our slope measure trading activityand volatility for different sub periods Granger causality test results are depicted inTable 513 and Table 514 We find that the null can be rejected at the 5 level for all
15We do not adjust for day-of-week effects as proposed by Naeligs and Skjeltorp 2006 as cryptocurren-cies can be traded at any time
Chapter 5 Statistical Properties of Cryptocurrency Order Books 81
TABLE 512 The Volume-Volatility relation and the Slope of the Order Book
Notes This table shows the fixed effects regression results for Equation 512 The dependent variableCorr(Trading Volumeit |Rit|) is the daily correlation coefficient measured over a month between USDtrading volume and the absolute return Coefficients of SLOPE100 and SLOPE10000 are multiplied by104 for better readability The coefficient of Spread and Trade size has been multiplied by 102
Model (1) (2) (3) (4) (5) (6)
SLOPE100 037 037 044
(368) (298) (357)
SLOPE10000 560 597 561
(190) (181) (190)
Spread 071 084 161 136
(089) (098) (210) (162)
Trade size minus003 011
(minus028) (minus108)
ln(MCAP) minus002 007
(minus043) (214)
Month fixed effects yes yes yes yes yes yes
N times T (currency-days) 231 186 231 231 186 231
R2 6080 6150 6070 5900 6030 5910
p lt 010 p lt 005 p lt 001 t-statistics in parentheses
tests performed on the granger-causal relationship between number of trades andthe slope of the order book (Table 513) indicating a bidirectional granger-causal re-lationship between the two variables16 Examining the causal relationship betweenvolatility and the slope of the order book (Table 514) reveals a less unambiguouspicture We can not reject bidirectional Granger causality between volatility andslope changes for Bitcoin at the 10 level and for Bitcoin Cash at the 1 level whilewe find evidence for unidirectional causality from volatility to slope changes forEthereum
16For most of the tests shown in Table 513 the null hypothesis can even be rejected at a 1 signifi-cance level
Chapter 5 Statistical Properties of Cryptocurrency Order Books 82
TABLE 513 Linear Granger Causality Test Results on Trades
Notes This table shows Granger causality test results between the number of trades and order bookslope changes Lag lengths are set with the Akaike (1974) information criterion adjusted by the numberof observations considering the first five lagged days Sig denotes the marginal significance level ofthe computed χ2-statistic used to test the zero restrictions implied by the null hypothesis of Grangernoncausality Slope indicates the number of ticks considered for estimating the order book slope mea-sure
H0 No of trades do not H0 Slope changes do notcause slope changes cause No of trades
Lags N Slope χ2 Sign χ2 Sign
Panel A Bitcoin (April 2018 ndash August 2019)
5 18 100 2077 000 2872 0005 18 10000 11883 000 2496 000
Panel B Bitcoin Cash (May 2018 ndash August 2019)
4 13 100 19617 000 1133 0024 13 10000 1315 001 4619 000
Panel C Ethereum (April 2018 ndash August 2019)
5 15 100 6630 000 1594 0015 15 10000 8031 000 1567 001
Chapter 5 Statistical Properties of Cryptocurrency Order Books 83
TABLE 514 Linear Granger Causality Test Results on Volatility
Notes This table shows Granger causality test results between volatility (|Rit|) and order book slopechanges Lag lengths are set with the Akaike (1974) information criterion adjusted by the number ofobservations considering the first five lagged days Sign denotes the marginal significance level of thecomputed χ2-statistic used to test the zero restrictions implied by the null hypothesis of Granger non-causality Slope indicates the number of ticks considered for estimating the order book slope measure
H0 Volatility does not H0 Slope changes do notcause slope changes cause volatility
Lags N Slope χ2 Sign χ2 Sign
Panel A Bitcoin (April 2018 ndash August 2019)
5 22 100 1027 007 4196 0002 51 10000 1232 000 662 004
Panel B Bitcoin Cash (May 2018 ndash August 2019)
5 14 100 2550 000 23472 0005 14 10000 2254 000 1765 000
Panel C Ethereum (April 2018 ndash August 2019)
1 68 100 854 000 001 0751 68 10000 249 012 023 063
Chapter 5 Statistical Properties of Cryptocurrency Order Books 84
56 Conclusion
This study examines the statistical properties of cryptocurrency limit order booksOur descriptive analysis reveals that the secondary market for cryptocurrencies isshaped by high cancellation rates and a preference for limit orders over market or-ders These findings are valid for both the buy and sell side Our subsequent empir-ical analysis reveals six key findingsBuilding upon our in-depth analysis of the LOBs of three major cryptocurrencieswe find evidence supporting our hypothesis that a group of lazy investors dis-regard the full granularity of the price grid which is reflected in volume peaks atcertain price levels distant to the best price (Key Finding 1) Testing the empiri-cal fit of different explanatory models for the slope of the order book we find thata linear approximation of the slope of the order book is reasonable without losingmuch explanatory power (Key Finding 2) Employing three different models wegain empirical evidence on the interplay between the slope of the order book pricechanges and trading activity We compute two slope measures (the inner slope andthe deeper slope) different in the range of data used for estimation The empiricalanalysis further reveals that the inner slope can not explain return variation whilethe deeper slope seems to contain information about cryptocurrency returns (KeyFinding 3) This finding indicates that limit orders in the depths of the order bookndash even though having a low probability to be executed and not being mechanicallylinked to price changes ndash are still relevant for the price formation process This find-ing suggests that traders incorporate the whole state of the order book when buyingor selling cryptocurrencies (Key Finding 4) This explanation is also practically rea-sonable as the state of the limit order book is visible to traders at any given timeIn addition we find that the inner slope of the order book has a significant positiveeffect on trading activity However this relation changes sign when considering thedeeper slope of the order book (Key Finding 5) This phenomenon has also beenobserved in stock markets by Naeligs and Skjeltorp (2006) Our results point into thesame direction and show that this anomaly can be observed in cryptocurrency mar-kets as well Moreover we find a positive relationship between trading volume andvolatility confirming the volume-volatility to be prevalent in cryptocurrency mar-kets as well Surprisingly we find the relation to be weaker when the slope of theorder book is steeper (Key Finding 6) This finding is significant across six differ-ent model specifications and contradicts the empirical results presented by Naeligs andSkjeltorp (2006)We further find that the spread equals the minimum tick size of 001 USD in 4202of the time in our data raising questions about the perception of the endogeneity ofthe spread by traders in todayrsquos markets This issue emerges from the finite gran-ularity of the price grid and has also been discussed by Biais Hillion and Spatt(1995)
85
Chapter 6
Heuristics in Cryptocurrency LimitOrder Placement
Exchanges and trading platforms allow us to analyze the behavior of interactingmarket participants From a scientific point of view the main advantage of financialmarkets is that all market participants trade the same asset with the aim to max-imize profits incentivizing rational behavior This unique setting allows to studynot only economic and financial theories but also theories of human behavior Inthis study we show that a power-law used to describe the distribution of limit or-der prices in stock markets can be extended to cryptocurrency markets despite thevery different market frameworks We hypothesize that cryptocurrency traders fallback to heuristics when placing limit orders provide a straightforward model ex-tension that accounts for this behavior and show that our model fits the empiricaldata better than the vanilla power-law model proposed in the literature
61 Introduction
In this study we examine the probability of incoming orders in a limit order mar-ket Today most of security trading is arranged via electronic order matching by ex-changes operating a limit order book (LOB) The flow of incoming limit orders yieldsinsights into market dynamics at the fine-granular level and therefore receives par-ticular interest from academia Zovko and Farmer (2002) and Bouchaud Meacutezardand Potters (2002) find a striking behavioral pattern while observing the placementof limit orders in stock markets Bouchaud Meacutezard and Potters (2002) state thatthe distribution of incoming limit order prices depends on the distance towards thebest available price and can be described by a universal power-law Further studiesprovide empirical evidence that supports the validity of the proposed power-law forvarying stocks and time frames (see Potters and Bouchaud 2003 Mike and Farmer2008 and Cont Stoikov and Talreja 2010)We extend the current state of scientific knowledge by showing that the univer-sal validity of the power-law can be generally extended to cryptocurrency mar-kets Accompanied by the increasing popularity and adoption of cryptocurrencies
Chapter 6 Heuristics in Cryptocurrency Limit Order Placement 86
secondary markets for cryptocurrencies emerge Those trading platforms operateLOBs in the same way as traditional stock exchanges and similar trading rules ap-ply1 However there exist some considerable differences between traditional ex-changes and cryptocurrency trading platforms as well Cryptocurrency trading plat-forms typically operate globally and without a break while traditional (national) ex-changes only provide services during localized business hours Further traded vol-ume in stock markets is of magnitudes higher than cryptocurrency trading volumeand important characteristics of market participants likely differ as well betweenthese market places eg investment horizon trading strategy or location Thosemarket conditions raise doubts on the unconditional transferability of the power-law to cryptocurrency marketsUsing cryptocurrency limit order flow data from a major cryptocurrency tradingplatform we find that order placement is increased when the relative distance to-wards the best price equals an (positive) integer We explain our finding by tradersusing a heuristic when placing limit orders Supposing that traders reduce the com-plexity of order placement by not considering the full granularity of the price gridwe propose a straightforward extension of the power-law relation and show that theextended model fits the empirical distribution of incoming limit order prices Theappeal of this extension lies in itrsquos simplicity which reflects the simplification madeby traders during limit order placement The existence of a substantial amount oftraders that rely on a simple heuristic when placing limit orders might indicate thatthe cryptocurrency market is still an emerging market where inefficiencies exist Toour knowledge no previous study that focuses on stock markets detects this place-ment behavior The remainder of this investigation is structured as follows Section62 provides the theoretical background of the proposed power-law in limit ordermarkets We motivate our hypothesis in Section 63 We derive our theoretical modeland provide empirical results in Section 64 Section 65 concludes
62 Statistics and Distribution of Incoming Limit Order Prices
In a limit order market traders submit a limit order to buy or sell a quantity ofan asset for a specific price Limit orders at the highest bid or lowest sell price arematched against incoming market orders while other limit orders remain in the LOBuntil they are either executed when the current price reaches their price level at alater point in time or canceled by the trader Similar to Bouchaud Meacutezard andPotters (2002) we denote ∆ as the absolute difference measured in ticks on the USDollar price grid between the current best price and the price of an incoming limitorder
∆ = |best available priceminus limit order price|1Notably most cryptocurrency trading platforms operate a LOW that follows a first-come first
serve principle and a price-time priority
Chapter 6 Heuristics in Cryptocurrency Limit Order Placement 87
Note that ∆ can be computed for bid and ask limit orders similarly An impatienttrader chooses ∆ to be close to zero thereby increasing the likelihood of quick orderexecution2 However the time advantage of placing a limit order close to the cur-rent bid or ask price runs in opposition to the risk of having the order executed at anunfavorable price Hence each trader faces a trade-off when placing a limit orderZovko and Farmer (2002) state that the choice of placing a limit order also dependson the individual goal of a trader and his trading strategy making order placement acomplex task Based on the assumption that traders differ in their expectations of fu-ture returns time horizon and risk aversion Chiarella Iori and Perelloacute (2009) showthat heterogeneous trading rules impact the limit order flow Hence the distributionof incoming limit order prices is not a priori clearBy analyzing stock LOBs from the Paris Bourse Bouchaud Meacutezard and Potters(2002) detect that the probability of a limit order arriving at ∆ can be described by auniversal power-law of the following form3
ρ(∆) sim 1∆1+micro
The advantage of power-law distributions lies in their simple representation andtheir prevalence in real world data Power-laws can be found across a wide rangeof complex economic relations that are influenced by many independent factors Infact the well-known pareto distribution (Pareto 1964) which is widely applied ineconomics is a power-lawUsing stock order flow data of three listed stocks Bouchaud Meacutezard and Potters(2002) estimate that micro = 06 While subsequent studies validate the power-law dis-tribution disagreement persists about the true value of micro
63 Limit Order Placement
We gather meta data by tracking the limit order flow via the application program-ming interface of a large cryptocurrency exchange Our data originates from fourcryptocurrencies traded against the US Dollar (USD) in April and June 2018 namelyBitcoin (BTC) Ethereum (ETH) Bitcoin Cash (BCH) and Litecoin (LTC)4 In totalwe gather 9 625 526 BTCUSD 34 905 938 ETHUSD 17 827 541 BCHUSD and33 467 649 LTCUSD limit order prices with a maximum of 500 ticks away fromthe current best price Supplemental information about the cryptocurrencies in oursample is given in Table 61 Figure 61 shows the empirical distribution of incoming
2∆ = 0 implies a limit order at the best bid price or ask price3Note that the proposed power-law is scale-invariant ie if ∆ is multiplied by a constant c we
would derive the following direct proportionality ρ(c∆) = 1(c∆)1+micro = cminus(1+micro)ρ(∆) prop ρ(∆) where prop
denotes direct proportionality4BTC ETH BCH and LTC combined account for roughly 80 of the total cryptocurrency market
capitalization
Chapter 6 Heuristics in Cryptocurrency Limit Order Placement 88
TABLE 61 Descriptive Statistics
Notes This table contains some supplemental information about the four cryptocurrency pairsBTCUSD ETHUSD BCHUSD and LTCUSD in our sample
BTCUSD ETHUSD BCHUSD LTCUSD
Initial price (USD) 818601 52435 126254 14092Final price (USD) 618003 41399 66551 7949Tick size (USD) 001 001 001 001Avg daily transaction volume (thsd) 931 BTC 10721 ETH 2104 BCH 20225 LTCAvg daily transaction volume (USD mn) 7763 6723 2090 1998Total limit orders1 (mn) 963 3491 1783 3347
limit ask orders (mn) 516 1955 973 1799 limit bid orders (mn) 446 1535 810 1548
1 Numbers only consider limit orders with ∆ le 500 that were recorded from April to June 2018 (with gaps)
limit order prices dependent on the distance towards the best price (∆) in logarith-mized form up to a maximum distance of ∆max = 500 for BTC ETH BCH and LTCThe depicted values are aggregated across the bid and ask sideFrom visual inspection the power-law seems to fit the distribution of limit orderprices quite well considering incoming BTCUSD and LTCUSD limit orders How-ever the distribution of incoming limit orders for ETHUSD reveals a deviatingpicture and shows that while most orders are placed close to the best price a localmaximum exists at about 100 to 200 ticks away from the current best price Thislocal maximum can be observed in the BCHUSD order placement as well and iseven more pronounced5 Nevertheless the observed cryptocurrencies consistentlyshow that the probability of order placement diminishes when moving away fromthe best price as implied by the power-law leading to our first hypothesis
Hypothesis A (H0A) The distribution of incoming limit order prices in cryptocur-rency markets follows a power-law like the order flow in stock markets
Analyzing Figure 61 reveals that the distribution of incoming limit orders exhibitspeaks which can be observed across all four currencies to a varying degree Theyare most eminent in the limit order flow of the BTCUSD currency pair We sup-pose that these peaks do not occur at random but follow a simple rule originating ina tradersrsquo heuristics used to reduce the complexity of the order placement decisionprocess which leads to our second hypothesis
Hypothesis B (H0B) The probability of an incoming limit order is increased whenthe distance towards the best price (∆) divided by 100 is a positive integer
5While it is not the focus of this study we suppose that the humps are related to the liquidity of therespective asset
Chapter 6 Heuristics in Cryptocurrency Limit Order Placement 89
Note that an incoming limit order at ∆ = 100 refers to a limit order placed at exactly100 USD away from the best price To illustrate our hypothesis we include verticallines in Figure 61 highlighting the associated values for ∆ = [100 200 500] Wefind the respective vertical lines to be located exactly at the peaks of the distributionacross all observed cryptocurrencies
64 Empirical Results
To empirically fit the power-law proposed by Bouchaud Meacutezard and Potters (2002)and in order to empirically test itrsquos universal character by applying it in cryptocur-rency markets (H0A) we estimate the value of the parameter micro We do so by trans-forming the proposed power-law equation by taking the natural logarithm at bothsides
P(∆) =c
∆1+micro
rArr ln (P(∆))︸ ︷︷ ︸Y
= ln(c)︸ ︷︷ ︸β0
+ (minus1minus micro)︸ ︷︷ ︸β1
middot ln(∆)︸ ︷︷ ︸X1
P(∆) denotes the frequency of a limit buy or limit sell order arriving ∆ ticks awayfrom the current best price As indicated by the brackets the transformed equationcan be substituted to derive an equation that is linear in its parameters Using themethod of ordinary least squares this property allows us to compute an estimatorfor micro by estimating the value of β1 as micro = minusβ1 minus 1Motivated by our graphical analysis of order placement behavior (Figure 61) andH0B we extend this model by including a factor λ and a function DN(∆) that de-pends on the value of ∆ and has the value 1 if ∆
100 isin N and zero otherwise Wepropose the following expression describing the distribution of incoming limit or-der prices
P(∆) =c middot eλmiddotDN(∆)
∆1+micro
rArr ln (P(∆))︸ ︷︷ ︸Y
= ln(c)︸ ︷︷ ︸β0
+ (minus1minus micro)︸ ︷︷ ︸β1
middot ln(∆)︸ ︷︷ ︸X1
+ λ︸︷︷︸β2
middotDN(∆)︸ ︷︷ ︸X2
As shown above the extended model can be transformed by applying the logarithmat both sides as well allowing us to estimate its parameters analog to the vanillapower-law model We measure DN(∆) by creating a dummy variable indicatingwhether the distance of an incoming limit order price divided by 100 is a naturalnumber ie whether the price of an incoming limit order i is 1 2 5 USD awayfrom the best price ε denotes the error term Our final empirical model is shown in
Chapter 6 Heuristics in Cryptocurrency Limit Order Placement 90
Formula 61
ln(P(∆))i = β0 + β1ln(∆)i + β2DN(∆)i + εi where (61)
DN(∆) =
1 i f ∆
100 isinN
0 else
Our empirical results are shown in Table 62 We find β1 and micro respectively to be
TABLE 62 Fitted Power Law Results
Notes This table provides regression results for Equation 61 Note that we estimate micro by computingmicro = minus(β1 + 1) To capture potential measurement errors we allow a narrow interval of [Nminus 005 N+005] for which the dummy variable DN = 1 Data was recorded from April to June 2018 (with gaps)The exponent of the denominator of the power-law is denoted as 1 + micro ie a negative value gt minus1 formicro is not surprising
Currency(USD) BTC BTC ETH ETH BCH BCH LTC LTC
micro -066 -065 -026 -025 -078 -077 085 085(-2670) (-2895) (-3142) (-3182) (-688) (-706) (-5577) (-5600)
DN(∆) 031 023 022 025(796) (304) (203) (236)
Constant yes yes yes yes yes yes yes yesAdj R2 5880 6338 6640 6694 849 906 8617 8630N 500 500 500 500 500 500 500 500
p lt 010 p lt 005 p lt 001 t-statistics in parenthesis
statistically significant at the 1-level across all cryptocurrencies and model spec-ifications indicating that the postulated universal nature of the power-law can beextended to cryptocurrency markets as well thereby strongly boosting itrsquos externalvalidity
Interim Result A Therefore we can not reject H0A
The explanatory power of the ordinary power-law of up to 8617 (LTCUSD) forcryptocurrencies indicates that the power-law may represent a fundamental charac-teristic of competitive limit order markets and should hence be considered an asset-independent feature in the fields of market microstructure As cryptocurrency mar-kets did not yet exist when the power-law distribution was first discovered ourfindings provide empirical evidence that the power-law distribution is an inherentproperty of competitive marketsMoreover our results show that the estimated parameter micro varies between minus078 to085 which is considerably below the estimated value in stock markets ranging from06 to 15 (see Bouchaud Meacutezard and Potters 2002 and Zovko and Farmer 2002)
Chapter 6 Heuristics in Cryptocurrency Limit Order Placement 91
A low value for micro implies that more limit orders are placed away from the currentprice in cryptocurrency markets This finding suggests that cryptocurrency tradersexpect larger USD price jumps in cryptocurrency markets which would be in linewith the general perception of cryptocurrencies being more volatile than stocksWe further find that the coefficient λ of our constructed variable DN(∆) is statisti-cally significant and of the same magnitude across all four cryptocurrencies
Interim Result B Hence we can not reject H0B
The results show that the probability of a specific limit order price is higher when thelimit order price is exactly 100 200 500 USD away from the current best priceWe suggest that this occurrence is of behavioral nature and traders prefer roundednumbers when setting the price level of limit orders ie they follow heuristics whenplacing limit orders Obviously traders seem to base their decision on the relativedistance towards the best price rather than the best price itself Eg our results sug-gest that a trader does not contemplate about buying at 50 USD but rather aboutbuying at 5 USD below the current price6 Including DN in our regression modelalso increases the explanatory power by 013 to 458 percentage points
Key Finding We argue that while the simple form of the power-law is appealing inexplaining the limit order price distribution the existence of behavioral preferencesof traders makes it necessary to account for specific values of the relative distancetowards the best price in future theoretical models
To our knowledge the existence of peaks in order price frequencies has not beenobserved by studies focusing on highly efficient stock marketsFigure 62 compares the fit of the universal power-law proposed by Bouchaud Meacutez-ard and Potters (2002) with the fit of our extended model exemplary for incomingBTCUSD limit orders Figure 62 demonstrates that solely using the power-law todescribe the distribution of limit order prices leads to a slight underestimation ofP(∆) ndash especially close to the best price ndash which can be reduced by accounting forpeaks
65 Conclusion
In this study we examine the limit order placement of four major cryptocurrenciesand compute the relative distance of incoming limit order prices towards the bestprice at arrival We show that limit order placement in cryptocurrency markets can
6Preferences of financial agents for rounded numbers have been documented in the literature be-fore eg Corwin (2003) finds that the issue yield in seasoned equity offers is related to underwriterpricing conventions such as price rounding and pricing relative to the bid quote Further links betweennumeric fluency and human preferences have been documented by Kettle and Haumlubl (2010)
Chapter 6 Heuristics in Cryptocurrency Limit Order Placement 92
be described by a power-law that was first discovered in the stock market literatureFurther we find empirical evidence that traders prefer integers when consideringhow far from the best price they place their limit order This preference for integersis likely the result of applied heuristics during limit order placement Traders resortto heuristics when dealing with the complexity of limit order placement and reducethe complexity by disregarding the full granularity of the price gridWe suggest a straightforward extension of the power-law approach to describe thedistribution of limit order prices and show that accounting for the observed behav-ioral regularity yields a better empirical fit than the plain power-law approach sug-gested in the literature Given the statistical significance of our finding we suggestthat future models trying to explain limit order placement should take this behav-ioral regularity into accountFurther studies are necessary to better understand the factors that influence the limitorder placement process in cryptocurrency markets over time Our data gives hintsto the fact that the observed phenomenon of an increased probability for incominglimit orders at certain numbers might be scale-invariant in that more incoming limitorders are placed at 01 USD than at 011 USD (05 USD than at 049 USD) away fromthe best price However our extended model needs to be tested in other markets andtime frames We would expect that it can show an improved explanatory power es-pecially in illiquid and emerging markets Finally the occurrence of humps in thelimit order price distribution in our data indicates a potential link between limit or-der prices and liquidity This complex relation needs further research and is crucialto fully understand and describe order placement behavior
Chapter 6 Heuristics in Cryptocurrency Limit Order Placement 93
0 1 2 3 4 5 6 7
log(∆)
8
9
10
11
12
13
14
15
16
log(P(∆
))
Fitted Power Law micro=-066
Left to right log(100) log(200) log(500)
BTCUSD
0 1 2 3 4 5 6 7
log(∆)
8
9
10
11
12
13
log(P(∆
))
Fitted Power Law micro=-078
Left to right log(100) log(200) log(500)
BCHUSD
0 1 2 3 4 5 6 7
log(∆)
9
10
11
12
13
14
15
log(P(∆
))
Fitted Power Law micro=-026
Left to right log(100) log(200) log(500)
ETHUSD
Notes This figure shows the cumulative distribution of ∆ of incoming orders as a function of ∆ forBTCUSD BCHUSD ETHUSD and LTCUSD We take the logarithm at both sides ∆ le 500
FIGURE 61 Incoming Limit Orders and Fitted Power-Law
Chapter 6 Heuristics in Cryptocurrency Limit Order Placement 94
0 1 2 3 4 5 6 7
log(∆)
4
6
8
10
12
14
16
18
20
log(P(∆
))
Fitted Power Law micro=085
Left to right log(100) log(200) log(500)
LTCUSD
FIGURE 61 (continued)
1000
030
000
5000
070
000
0 100 200 300 400 500Delta
BTCUSD Fitted Power-LawFitted Power-Law (Extended Model)
P(D
elta
)
Notes We set an interval of [Nminus 001 N + 001] for which the dummy variable DN = 1 Note thedecreasing magnitude of the peaks in the extended model caused by the parameter λ which also fitsthe empirical data The predicted peaks decrease because λ is incorporated multiplicative rather thanadditive in the extended model We estimate λ as the β-coefficient of the dummy variable DN = 1
FIGURE 62 Fitted Power-Law and Extended Model for BTCUSD
95
Chapter 7
ReaktionenderKryptowaumlhrungsmaumlrkteaufdieCOVID-19-Pandemie
Chapter 7 has been published as a journal articleHaumlfner David Jianan He and Dirk Schiereck (2020) Reaktionen der Kryptowaumlh-rungsmaumlrkte auf die COVID-19-Pandemie In Corporate Finance 05-06 pp 141ndash144ISSN 2198-8889
96
Chapter 8
Concluding Remarks
This dissertation explores the behavior and interaction of market participants in dif-ferent market conditions and is focused on empirically analyzing the efficiency ofvarious financial market segments and the behavior of market participants undercompetition and their reaction to exogenous shocks and imperfect informationIn Chapter 2 we examine the development of corporate social responsibility andits compatibility with efficient investments By reviewing 74 leading articles pub-lished between 1984 and 2016 we provide a review on the scientific literature andhighlight the importance and effects of socially responsible and sustainable invest-ments from three different perspectives the investor level the company level andthe portfolio level We provide an in-depth analysis of the motivation behavior anddemographics of socially responsible investors We further show that the motivationof the management financial inducement and exogenous influence ndash eg publicityor consumer behavior ndash can impact the extent to which companies act responsiblePortfolio implications are focused on the financial effects ie risk and return of so-cially responsible investments We group the existing literature and empirical find-ings geographically as the regulatory framework and government support is likelyto shape the investment environment eg Henke (2016) finds that European fundshave higher responsibility scores We discuss the historic development of sociallyresponsible investments and gather empirical results considering the financial per-formance of socially responsible investment funds over time It remains an openquestion whether responsible investment funds perform better or worse than unre-stricted funds as evidence for both directions is provided in the literature Howeverwe do not find clear evidence that social responsible investments struggle which ndashconsidering the restricted investment universe ndash is quite remarkable However fur-ther empirical studies including more recent data are necessary to determine theimpact of social responsibility on financial performanceIn Chapter 3 we empirically analyze the impact of a regulatory change in Euro-pean financial markets on information asymmetry idiosyncratic risk and liquidityThe introduction of MiFID II in January 2018 included extensive changes for finan-cial markets with the superordinate aim to increase transparency The most drasticchange concerns the market for investment research provided by financial analysts
Chapter 8 Concluding Remarks 97
We discuss how the literature defines the role of analysts in financial markets andhighlight how investment research is conducted in the EU prior to MiFID II andcompare the regulatory framework with the US Using data of 1281 EU firms and1646 US firms we employ a difference-in-difference approach to determine the ef-fect of MiFID II on a firmrsquos analyst coverage liquidity stock trading volume thebid-ask spread and its idiosyncratic risk We empirically show that MiFID II had asignificant impact on European financial markets as the overall bid-ask spread andthe idiosyncratic risk increased We hypothesize that a financial analyst experiencesincreased competition since the implementation of MiFID II which should have averifiable effect on the bid-ask spread We provide empirical evidence that an in-crease in analyst coverage reduces the bid-ask spread of a stock emphasizing theinformational role of financial analysts By estimating the effect of analyst cover-age on the bid-ask spread (while controlling for liquidity) we find evidence thatfinancial analysts affect the level of asymmetric information of a stock This effect isamplified by MiFID IIIn Chapter 4 we empirically analyze the structure and performance of ICOs Weshow that ICOs share many similarities with IPOs and are currently generally per-formed by young and entrepreneurial companies due to their low direct costs Gath-ering data from 175 ICOs we empirically analyze the indirect costs of an ICO bycomputing the underpricing during ICOs We find that the underpricing phenome-non ndash defined as a positive return at the first trading day of a stock ndash can be observedduring an ICO as well and is even more apparent than during IPOs There existseveral explanatory approaches for the existence of IPO underpricing of which themajority is centered on information asymmetry between the issuing firm the under-writing bank and investors We discuss these hypotheses and their transferabilityonto the ICO setting Our results suggest that the high level of ICO underpricing isrelated to a higher degree of information asymmetry between the issuing firm andinvestors or between investors as information about ICOs is scarce oftentimes witha white paper as the sole source of information Therefore investors might demandunderpricing as compensation for participating in an ICO with little prior knowl-edge about the risk and return profile However we cannot rule out behavioralexplanatory approaches and further research is necessary to fully understand thecause for the high level of underpricing observed during ICOs We further find thatthe long-term performance of new cryptocurrencies generally remains below theperformance of established cryptocurrencies This anomaly has been documentedin the stock markets as well (see Ritter and Welch 2002)
Chapter 8 Concluding Remarks 98
In Chapter 5 we analyze statistical properties of cryptocurrency limit order booksFirst we motivate the importance of market microstructure as a prerequisite for ef-ficient capital allocation Next we discuss the scope of application for cryptocurren-cies and their disruptive potential We gather data from one of the largest cryptocur-rency exchanges and reconstruct the limit order books for three different cryptocur-rencies (Bitcoin Ethereum and Bitcoin Cash) We find two distinctive features of theaggregated limit order book volume known as the slope of the order book The slopevaries over time and can have linear concave or even convex properties We test theempirical fit for different shapes of the slope of the order book and find that a linearslope is generally a sufficient approximation and justifiable with respect to modelsimplicity We further stumble upon an anomaly in the distribution of limit orderbook volume We observe volume peaks in the limit order book at specific pricelevels relative to the best price This anomaly is detectable across all analyzed limitorder books We hypothesize that this pattern is created by a group of investorswhich do not consider the entire granularity of the price grid This leads to an accu-mulation of available volume at specific price levels We coin this the ldquolazy investorhypothesisrdquo and develop a straight-forward regression model to test the empiricalsignificance of this finding and find the respective estimated regression coefficientto be highly significant at both the bid and the ask side across all limit order booksFinally we test whether there is information stored in the order book slope Similarempirical work has been done by Naeligs and Skjeltorp (2006) who study stock limit or-der books We investigate three groups of models using the slope of the order bookto explain price changes trading volume and the correlation between price changesand trading volume We compute the slope on a daily basis averaged across the bidand ask side of the limit order book and show that the slope of the limit order bookcan help to explain variation in the dependent variables Our findings also indicatethat limit orders placed far away from the best price still appear to be relevant forthe price formation process This finding is interesting as there is no plausible me-chanical link explaining this relationshipIn Chapter 6 we focus on the limit order placement behavior of cryptocurrencytraders There have been some seemingly universal statistical laws discovered inthe stock market literature that describe the placement of limit orders Each traderhas to decide at which price level he or she issues a limit order Each trader faces atrade-off regarding the execution probability of a limit order time-priority and thethreat of having the limit order executed at an unfavorable price The decision on theprice level further depends on the individual trading strategy liquidity preferencesand other factors making limit order placement a complex process This makes theemergence of patterns in the arrival of limit orders even more surprising Zovko andFarmer (2002) and Bouchaud Meacutezard and Potters (2002) show that the probabilityof an incoming limit order can be described by a universal power-law in stock mar-kets We document that this power-law can be observed in cryptocurrency markets
Chapter 8 Concluding Remarks 99
as well Moreover we find empirical evidence that cryptocurrency traders seem toprefer certain price levels and more limit orders are issued at price levels exactly100 200 500 USD away from the best price We hypothesize that the probabil-ity of an incoming limit order is increased when the distance towards the best pricedivided by 100 is an integer indicating that traders apply heuristics when placinglimit orders We extend the power-law model to account for this occurrence andshow that our extended model better fits the empirical data than the plain modelIn Chapter 7 we analyze the impact of a radical change of the macroeconomic en-vironment on the market for cryptocurrencies The outbreak of the coronavirusin 20192020 had a sudden impact on all industries and caused unprecedencedchanges to the market environment across all asset classes around the globe Inthis chapter we focus on the impact on the cryptocurrency market and the poten-tial of cryptocurrencies as a save investment haven in times of a crisis We find thatcryptocurrencies with a high market capitalization seem to be affected by the overalldepreciation at the beginning of the outbreak and ndash in line with Corbet et al (2020)ndash their suitability as a save investment haven is therefore questioned Howeversmaller cryptocurrencies seem to be affected less by the virus outbreak and mightbe more appropriate for a crisis-resistant save haven portfolio However the fullextent of COVID-19 and its impact on cryptocurrency markets is not yet foreseeableand further research needs to be conducted to understand the ways in which a virusoutbreak affects this emerging asset classAfter completion of this thesis our findings concerning the six broader researchquestions derived in Chapter 1 can be summarized as follows
bull Research Question 1 Is the concept of market efficiency compatible with so-cial responsibility
Result Market efficiency is compatible with social responsibility mainly dueto non-monetary benefits (eg consumer reactions personal values or pub-licity) Further empirically evidence suggests that funds engaging in sociallyresponsible investments do not perform significantly worse than unrestrictedfunds
bull Research Question 2 How do participants in highly efficient stock marketsreact to an exogenous shock in market design (regulatory change)
Result The net positive effect of regulation on financial markets remains anopen question and side effects of a regulatory change are tough to predict Inour empirical analysis however we specifically focus on MiFID II and showthat this regulatory change led to an increase in the average bid-ask spreadand idiosyncratic risk across EU stock markets We also find that the effect of
Chapter 8 Concluding Remarks 100
analyst coverage on the bid-ask spread is enhanced since the implementationof MiFID II
bull Research Question 3 How do investors cope with information asymmetry
Result While investors do not have the capabilities to obtain all price-relevantinformation they do not seem to be deterred from information asymmetryThey rather seek compensation for engaging in uncertain and thereby riskyinvestments We show that ICOs ndash which are typically surrounded by a highlevel of information asymmetry between investors and the issuing company ndashexhibit high initial returns at the first trading day We compare our results tothe IPO process and find that ICO underpricing is higher than IPO underpric-ing supporting the hypothesis that the underpricing is caused by informationasymmetry and can be regarded as a compensation for investors
bull Research Question 4 How does the market microstructure shape market dy-namics
Result We group market dynamics into three distinct categories Price volatil-ity trading volume and the volume-volatility relation We try to connect themarket microstructure and the market development by computing daily aver-ages of certain microstructure features Analyzing the market for cryptocur-rencies we find that the slope of the order book ndash representing the aggregatedlimit order book volume ndash has an impact on all three of the categories Themarket microstructure seems to significantly impact the broader market devel-opment and should be considered in asset pricing models Traders can observethe state of the limit order book at any given point in time and seem to con-sider its shape during their investment decision
bull Research Question 5 How do traders behave when placing limit buy and sellorders in a competitive market setting
Result We find that limit order placement is a complex task and the solu-tion is determined by individual time and price preferences and the respectivetrading strategy We find empirical evidence suggesting that traders resort toheuristics when determining limit order prices By disregarding the granular-ity of the price grid traders reduce the complexity and find a solution moreeasily This behavioral regularity has not been documented in the literaturebefore and we provide an extended model that takes this finding into account
Chapter 8 Concluding Remarks 101
Our empirical results show that our theoretical model can explain the distri-bution of incoming limit order prices better than existing models
bull Research Question 6 How does the cryptocurrency market react to an exoge-nous shock (market crisis)
Result We find that smaller cryptocurrencies exhibit return patterns that sep-arate them from cryptocurrencies with a high market capitalization Tradingvolume seems to be a predictor for returns of small cryptocurrencies This rela-tion is enhanced since the exogenous shock caused by COVID-19 Further theinteraction term between return and trading volume has a significant negativeeffect on future returns of large cryptocurrencies While all cryptocurrenciessuffered from the initial macroeconomic shock caused by the COVID-19 pan-demic the market for smaller cryptocurrencies seems to be more resilient toexogenous shocks
In conclusion this dissertation provides an in-depth analysis of the complex inter-play between financial market agents in different market settings We find that infor-mation asymmetry plays a crucial role across markets time periods and geograph-ical regions and provide empirical evidence about the extent to which informationdisparity can affect asset prices and liquidity Throughout this dissertation we showthat investors act responsible despite no direct monetary incentive (Chapter 2) relyon financial analysts to obtain price-relevant information and are affected by regu-latory changes (Chapter 3) demand a premium when not having full informationduring an investment decision (Chapter 4) use the limit order book as a source ofinformation (Chapter 5) and resort to heuristics when solving complex tasks (Chap-ter 6) Chapter 7 finally showcases how sensitive the market equilibrium can beto a sudden change in the market setting and discusses the adjustments made byinvestors
102
Bibliography
Akaike Hirotugu (1974) ldquoA new look at the statistical model identificationrdquo InIEEE transactions on automatic control 196 pp 716ndash723
AMF (2018) MiFID II Impact of the new tick size regime URL httpswwwamf-franceorgen_USPublicationsLettres-et-cahiersRisques-et-tendances
ArchivesdocId=workspace5C3A5C2F5C2FSpacesStore5C2F4ee6cbf6-
c425-4537-ab74-ef249b9d316d (visited on 01152019)Amihud Yakov (2002) ldquoIlliquidity and stock returns cross-section and time-series
effectsrdquo In Journal of Financial Markets 51 pp 31ndash56Barber Brad Reuven Lehavy Maureen McNichols and Brett Trueman (2001) ldquoCan
investors profit from the prophets Security analyst recommendations and stockreturnsrdquo In The Journal of Finance 562 pp 531ndash563
Bessler Wolfgang and Matthias Stanzel (2007) ldquoQualitaumlt und Effizienz der Gewinn-prognosen von Analystenrdquo In Kredit und Kapital 401 pp 89ndash129
Biais Bruno Pierre Hillion and Chester Spatt (1995) ldquoAn empirical analysis of thelimit order book and the order flow in the Paris Bourserdquo In The Journal of Finance505 pp 1655ndash1689
Bouchaud Jean-Philippe Marc Meacutezard and Marc Potters (2002) ldquoStatistical proper-ties of stock order books empirical results and modelsrdquo In Quantitative Finance24 pp 251ndash256
Casey Jean-Pierre and Karel Lannoo (2009) The MiFID Revolution Cambridge Uni-versity Press
Chen Qi Itay Goldstein and Wei Jiang (2006) ldquoPrice informativeness and invest-ment sensitivity to stock pricerdquo In The Review of Financial Studies 203 pp 619ndash650
Chiarella Carl Giulia Iori and Josep Perelloacute (2009) ldquoThe impact of heterogeneoustrading rules on the limit order book and order flowsrdquo In Journal of EconomicDynamics and Control 333 pp 525ndash537
Chung Kee H Thomas H McInish Robert A Wood and Donald J Wyhowski(1995) ldquoProduction of information information asymmetry and the bid-askspread Empirical evidence from analystsrsquo forecastsrdquo In Journal of Banking amp Fi-nance 196 pp 1025ndash1046
Citigate Dewe Rogerson (2019) 11th Annual IR Survey URL httpscitigate5C- dewe5C- rogersoncomuk- companies- among- hardest- hit- decline-
Bibliography 103
analyst-coverage-according-citigate-dewe-rogersons-11th-annual-ir-
survey (visited on 10282019)Cont Rama (2001) ldquoEmpirical properties of asset returns stylized facts and statisti-
cal issuesrdquo In Quantitative Finance 12 pp 223ndash236Cont Rama Sasha Stoikov and Rishi Talreja (2010) ldquoA stochastic model for order
book dynamicsrdquo In Operations Research 583 pp 549ndash563Copeland Thomas E and Dan Galai (1983) ldquoInformation effects on the bid-ask
spreadrdquo In The Journal of Finance 385 pp 1457ndash1469Corbet Shaen Charles Larkin and Brian Lucey (2020) ldquoThe contagion effects of
the covid-19 pandemic Evidence from gold and cryptocurrenciesrdquo In FinanceResearch Letters 35 p 101554
Corwin Shane A (2003) ldquoThe determinants of underpricing for seasoned equityoffersrdquo In The Journal of Finance 585 pp 2249ndash2279
De Bondt Werner F M and Richard H Thaler (1990) ldquoDo security analysts overre-actrdquo In The American Economic Review pp 52ndash57
De Vries Alex (2018) ldquoBitcoinrsquos growing energy problemrdquo In Joule 25 pp 801ndash805Deutsche Boumlrse (2017) Die Auswirkungen von MiFID II auf die Verfuumlgbarkeit von Re-
search URL httpswwwdeutsche- boersecomresourceblob166524ab0a3455fe3aab17051126128607e577dataMarket-Trends-MiFID-II-Research
pdf (visited on 11042019)Deutscher Investor Relations Verband (2017) Moumlgliche Auswirkungen von MiFID II
fuumlr Emittenten URL httpswwwdirkorgdirk_webseitestaticuploads171006-MV_KS_Praesentation_Bommer_MiFIDpdf (visited on 11042019)
Durnev Artyom Randall Morck Bernard Yeung and Paul Zarowin (2003) ldquoDoesgreater firm-specific return variation mean more or less informed stock pricingrdquoIn Journal of Accounting Research 415 pp 797ndash836
Easley David Nicholas M Kiefer Maureen OrsquoHara and Joseph B Paperman (1996)ldquoLiquidity information and infrequently traded stocksrdquo In The Journal of Finance514 pp 1405ndash1436
Easley David Maureen OrsquoHara and Joseph Paperman (1998) ldquoFinancial analystsand information-based traderdquo In Journal of Financial Markets 12 pp 175ndash201
ESMA (2018) Amendment to Commission Delegated Regulation (EU) 2017588 (RTS 11)URL httpswwwesmaeuropaeupress-newsconsultationsamendment-commission- delegated- regulation- eu- 2017588- rts- 11TODO (visited on01152019)
European Commission (2014) ldquoDirective 201465EU of the European Parliamentand of the Council of 15 May 2014 on Market Financial Instruments and Amend-ing Directive 200292EC and Directive 201161EUrdquo In Official Journal of theEuropean Union 173 pp 349ndash496
Fama Eugene F (1970) ldquoEfficient capital markets A review of theory and empiricalworkrdquo In The Journal of Finance 252 pp 383ndash417
Bibliography 104
Fang Bingxu Ole-Kristian Hope Zhongwei Huang and Rucsandra Moldovan (2019)ldquoThe Effects of MiFID II on Sell-Side Analysts Buy-Side Analysts and FirmsrdquoURL httpspapersssrncomsol3paperscfmabstract_id=3422155(visited on 11042019)
Ferrarini Guido and Eddy Wymeersch (2006) Investor protection in Europe corporatelaw making the MiFID and beyond Oxford University Press
Foster Andrew J (1995) ldquoVolume-volatility relationships for crude oil futures mar-ketsrdquo In Journal of Futures Markets 158 pp 929ndash951
Foster F Douglas and S Viswanathan (1995) ldquoCan speculative trading explain thevolumendashvolatility relationrdquo In Journal of Business amp Economic Statistics 134pp 379ndash396
French Kenneth R and Richard Roll (1986) ldquoStock return variances The arrival ofinformation and the reaction of tradersrdquo In Journal of Financial Economics 171pp 5ndash26
Fung Hung-Gay and Gary A Patterson (1999) ldquoThe dynamic relationship of volatil-ity volume and market depth in currency futures marketsrdquo In Journal of Inter-national Financial Markets Institutions and Money 91 pp 33ndash59
Givoly Dan and Josef Lakonishok (1979) ldquoThe information content of financial ana-lystsrsquo forecasts of earnings Some evidence on semi-strong inefficiencyrdquo In Jour-nal of Accounting and Economics 13 pp 165ndash185
Gleason Cristi A and Charles M C Lee (2003) ldquoAnalyst forecast revisions and mar-ket price discoveryrdquo In The Accounting Review 781 pp 193ndash225
Glosten Lawrence R (1987) ldquoComponents of the bid-ask spread and the statisticalproperties of transaction pricesrdquo In The Journal of Finance 425 pp 1293ndash1307
Glosten Lawrence R and Lawrence E Harris (1988) ldquoEstimating the components ofthe bidask spreadrdquo In Journal of Financial Economics 211 pp 123ndash142
Glosten Lawrence R and Paul R Milgrom (1985) ldquoBid ask and transaction pricesin a specialist market with heterogeneously informed tradersrdquo In Journal of Fi-nancial Economics 141 pp 71ndash100
Grossman Sanford J and Joseph E Stiglitz (1980) ldquoOn the impossibility of informa-tionally efficient marketsrdquo In The American Economic Review 703 pp 393ndash408
Henke Hans-Martin (2016) ldquoThe effect of social screening on bond mutual fundperformancerdquo In Journal of Banking amp Finance 67 pp 69ndash84
Holste Bjoern and Christoph Gallus (2019) Sind Krypto-Waumlhrungsmaumlrkte Fair(AreCrypto-Currency Markets Fair) Working Paper
Johnsen Edward J and John Grady (2017) SEC provides relief to US firms attemptingto comply with EU MiFID IIrsquos research ldquounbundling provisions URL httpswwwdlapipercomenusinsightspublications201710sec-relief-to-us-
firms-attempting-to-comply (visited on 01102019)
Bibliography 105
Karpoff Jonathan M (1987) ldquoThe relation between price changes and trading vol-ume A surveyrdquo In Journal of Financial and Quantitative Analysis 221 pp 109ndash126
Kelly Bryan and Alexander Ljungqvist (2012) ldquoTesting asymmetric-information as-set pricing modelsrdquo In The Review of Financial Studies 255 pp 1366ndash1413
Kettle Keri and Gerald Haumlubl (2010) ldquoNumeric fluency and preferencerdquo In NA ndashNorth American Advances 37 pp 150ndash152
Kim Oliver and Robert E Verrecchia (1994) ldquoMarket liquidity and volume aroundearnings announcementsrdquo In Journal of Accounting and Economics 171-2 pp 41ndash67
Lippiatt Todd and Michael Oved (2018) The Two Token Waterfall URL httpsframeworkfactoraio (visited on 02072019)
Mantegna Rosario N and H Eugene Stanley (1999) Introduction to econophysics cor-relations and complexity in finance Cambridge University Press
Marshall Ben R Nhut H Nguyen and Nuttawat Visaltanachoti (2019) Bitcoin liq-uidity Working Paper Massey University and Auckland University of Technol-ogy
Mike Szabolcs and J Doyne Farmer (2008) ldquoAn empirical behavioral model of liq-uidity and volatilityrdquo In Journal of Economic Dynamics and Control 321 pp 200ndash234
Morck Randall Bernard Yeung and Wayne Yu (2000) ldquoThe information content ofstock markets why do emerging markets have synchronous stock price move-mentsrdquo In Journal of Financial Economics 581-2 pp 215ndash260
Naeligs Randi and Johannes A Skjeltorp (2006) ldquoOrder book characteristics and thevolumendashvolatility relation Empirical evidence from a limit order marketrdquo InJournal of Financial Markets 94 pp 408ndash432
Nakamoto Satoshi (2008) Bitcoin A peer-to-peer electronic cash system Working PaperNofer Michael Peter Gomber Oliver Hinz and Dirk Schiereck (2017) ldquoBlockchainrdquo
In Business amp Information Systems Engineering 593 pp 183ndash187OrsquoHara Maureen (1997) Market microstructure theory WileyOrsquoHara Maureen (2015) ldquoHigh frequency market microstructurerdquo In Journal of Fi-
nancial Economics 1162 pp 257ndash270Pareto Vilfredo (1964) Cours drsquoeacuteconomie politique Vol 1 Librairie DrozPiotroski Joseph D and Darren T Roulstone (2004) ldquoThe influence of analysts
institutional investors and insiders on the incorporation of market industryand firm-specific information into stock pricesrdquo In The Accounting Review 794pp 1119ndash1151
Potters Marc and Jean-Philippe Bouchaud (2003) ldquoMore statistical properties of or-der books and price impactrdquo In Physica A Statistical Mechanics and its Applica-tions 3241-2 pp 133ndash140
Bibliography 106
Preece Rhodri (2019) MiFID II One Year On URL httpswwwcfainstituteorg-mediadocumentssurveycfa-mifid-II-survey-reportashx (visitedon 10292019)
Preis Tobias Sebastian Golke Wolfgang Paul and Johannes J Schneider (2006) ldquoMulti-agent-based order book model of financial marketsrdquo In EPL (Europhysics Letters)753 pp 510ndash516
Ritter Jay R (1987) ldquoThe costs of going publicrdquo In Journal of Financial Economics192 pp 269ndash281
Ritter Jay R and Ivo Welch (2002) ldquoA review of IPO activity pricing and alloca-tionsrdquo In The Journal of Finance 574 pp 1795ndash1828
Roll Richard (1988) ldquoR2rdquo In The Journal of Finance 433 pp 541ndash566Sarwar Ghulam (2003) ldquoThe interrelation of price volatility and trading volume
of currency optionsrdquo In Journal of Futures Markets Futures Options and OtherDerivative Products 237 pp 681ndash700
SEC (2017) Response of the chief counselrsquos office division of investment management URLhttpswwwsecgovdivisionsinvestmentnoaction2017sifma-102617-
202ahtm (visited on 01142019)Stoll Hans R (1989) ldquoInferring the components of the bid-ask spread Theory and
empirical testsrdquo In The Journal of Finance 441 pp 115ndash134Thomsen Steen and Frederik Vinten (2014) ldquoDelistings and the costs of governance
a study of European stock exchanges 1996ndash2004rdquo In Journal of Management ampGovernance 183 pp 793ndash833
Wallmeier Martin (2005) ldquoAnalystsrsquo earnings forecasts for DAX100 firms during thestock market boom of the 1990srdquo In Financial Markets and Portfolio Management192 pp 131ndash151
Wolfson Rachel (2018) A first for Manhattan $30M Real Estate Property Tokenized WithBlockchain URL httpswwwforbescomsitesrachelwolfson20181003a- first- for- manhattan- 30m- real- estate- property- tokenized- with-
blockchain1910a2d48957 (visited on 02072019)Womack Kent L (1996) ldquoDo brokerage analystsrsquo recommendations have investment
valuerdquo In The Journal of Finance 511 pp 137ndash167Zhu PengCheng Vijay Jog and Isaac Otchere (2014) ldquoIdiosyncratic volatility and
mergers and acquisitions in emerging marketsrdquo In Emerging Markets Review 19pp 18ndash48
Zovko Ilija and J Doyne Farmer (2002) ldquoThe power of patience a behavioural reg-ularity in limit-order placementrdquo In Quantitative Finance 25 pp 387ndash392
- Abstract
- Acknowledgements
- Synopsis
-
- Motivation
- Thesis Structure
-
- What do we know about socially responsible investments
- The Role of Investment Research in view of MiFID II An Empirical Analysis of Information Asymmetry Idiosyncratic Risk and Liquidity
-
- Introduction
- Regulation of Investment Research
-
- Investment Research in the European Union
- Investment Research in the United States
-
- The Role of Investment Research in Capital Markets
- Variable Definition
-
- Illiquidity
- The Bid-Ask Spread
- Stock Price Informativeness and Idiosyncratic Risk
-
- Data
-
- Descriptive Statistics
-
- Empirical Analysis
-
- Difference-in-Differences Approach
- Multiple Regression on the Bid-Ask Spread
- Multiple Regression on Idiosyncratic Risk
- Summary Statistics on Medium-Term Effects on Research Coverage
-
- Conclusion
-
- Innovative Finanzierungen uumlber Initial Coin Offerings Struktur und bisherige Performance
- Statistical Properties of Cryptocurrency Order Books
-
- Introduction
- Structure of a Limit Order Book
- Data Collection and Description
-
- Data Acquisition Process
- Descriptive Statistics
-
- Limit Order Book Characteristics in Cryptocurrency Markets
-
- The Shape of the Average Order Book
- The Bitcoin Order Book
- The Bitcoin Cash Order Book
- The Ethereum Order Book
- The Lazy Investor Hypothesis
- Convexity of the Order Book Slope
-
- Informativeness of Order Book Characteristics
-
- Measuring the Slope of the Limit Order Book
- Order Book Slope and Volatility
- Order Book Slope and Trading Activity
- Order Book Slope and the Volume-Volatility Relation
- A Note on Causality
-
- Conclusion
-
- Heuristics in Cryptocurrency Limit Order Placement
-
- Introduction
- Statistics and Distribution of Incoming Limit Order Prices
- Limit Order Placement
- Empirical Results
- Conclusion
-
- Reaktionen der Kryptowaumlhrungsmaumlrkte auf die COVID-19-Pandemie
- Concluding Remarks
- Bibliography
-