Journal of Banking and Finance 124 (2021) 106041

Contents lists available at ScienceDirect

Journal of Banking and Finance

journal homepage: www.elsevier.com/locate/jbf

How to measure the liquidity of cryptocurrency markets?

Alexander Brauneis a, Roland Mestel b,∗, Ryan Riordan c, Erik Theissen b,d
a
University of Klagenfurt, Department of Finance & Accounting, Universitaetsstrasse 65-67, Klagenfurt A-9020, Austria
b
University of Graz, Institute of Banking and Finance, Universitaetsstrasse 15/F2, Graz A-8010, Austria
c
Smith School of Business, Queen’s University, Goodes Hall, 143 Union St West, Kingston, ON K7L 3N6, Canada
d
University of Mannheim, Finance Area, L9 1-2, Mannheim D-68161, Germany

a r t i c l e i n f o a b s t r a c t

Article history: This paper investigates the efficacy of low-frequency transactions-based liquidity measures to describe ac-
Received 22 July 2020 tual (high-frequency) liquidity. We show that the Corwin and Schultz (2012) and Abdi and Ranaldo (2017)
Accepted 27 December 2020
estimators outperform other measures in describing time-series variations, irrespective of the observa-
Available online 4 January 2021
tion frequency, trading venue, high-frequency liquidity benchmark, and cryptocurrency. Both measures
JEL classification: perform well during high and low return, volatility and volume periods. The Kyle and Obizhaeva (2016)
G12 estimator and the Amihud (2002) illiquidity ratio outperform when estimating liquidity levels. These two
G14 estimators also reliably identify liquidity differences between trading venues. Overall, the results suggest
that there is not yet a universally best measure but there are reasonably good low-frequency measures.
Keywords:
Cryptocurrencies © 2021 The Authors. Published by Elsevier B.V.
Liquidity This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/)
Capital markets

1. Introduction was of similar magnitude to global bitcoin trading, with $1.03 tril-
lion USD for the NYSE during December 2019.3
Bitcoin and other cryptocurrencies are now firmly entrenched The growing importance of bitcoin for payments and invest-
in the financial system. Bitcoin is becoming a widely accepted form ments is dependent on an efficient transfer of bitcoin for other
of online payment and more than 35 million bitcoin wallets are in currencies on cryptocurrency exchanges. The number of exchanges
existence. Trading in bitcoin exceeded $930 billion USD in January has exploded, making it difficult for investors to select an ex-
2020. Bitcoin also forms a growing part of investment portfolios change for trading and hedging. While trading has become rela-
and serves as the underlying for futures contracts,1 recently ex- tively frequent in cryptocurrencies the liquidity of these markets
ceeding $1 billion in open interest.2 Originally designed as a de- is difficult to determine. Cryptocurrency markets lack a regulated
centralized digital cash system using cryptographic hash functions data feed like the consolidated tape for U.S. equities. The lack of
to secure transactions, it is poised to overtake national fiat curren- a consolidated feed, coupled with the high number of exchanges
cies and other financial assets in terms of global importance. and jurisdictions makes it difficult to calculate high-frequency bid-
Bitcoin and other cryptocurrencies are traded on numerous ask spreads thereby hampering the comparison of liquidity across
trading platforms around the globe. Bitcoin can be traded 24 h per cryptocurrency exchanges. The bid-ask spread is an important met-
day, and seven days per week for US dollars, the Euro, Japanese ric when assessing an exchange in that it represents the costs of
Yen, as well as numerous other fiat and (crypto)currencies. immediately buying or selling a security. Bid-ask spreads are usu-
Monthly dollar trading volume on the New York Stock Exchange ally calculated using high-frequency intraday data that are both ex-
pensive to purchase and time-consuming to process. We compare

∗
Corresponding author.
E-mail addresses: alexander.brauneis@aau.at (A. Brauneis), roland.mestel@uni-
graz.at (R. Mestel), ryan.riordan@queensu.ca (R. Riordan), theissen@uni- 3
Data for bitcoin volume come from https://coinmarketcap.com/currencies/
mannheim.de (E. Theissen). bitcoin/historical-data/ and NYSE volume data are from https://focus.
URL: http://www.aau.at/fin (A. Brauneis), https://banken-finanzierung.uni- world-exchanges.org/issue/january-2020/market-statistics. Both sites were ac-
graz.at (R. Mestel), http://ryanriordan.ca/ (R. Riordan), https://www.bwl.uni- cessed on February 15, 2020. We note, though, that some cryptocurrency exchanges
mannheim.de/theissen/ (E. Theissen) appear to be overstating their reported volume (e.g. Hougan et al., 2019) and
1
See: https://www.cmegroup.com/trading/bitcoin-futures.html that ranking websites like coinmarketcap.com may be tempted to report in-
2
https://www.coindesk.com/bitcoin- futures- pass- 1b- in- open- interest- on- bitmex- flated trading volume due to their revenue model which is largely dependent on
for- first- time- since- march- crash crypto-exchanges (Alexander and Dakos, 2019).

https://doi.org/10.1016/j.jbankfin.2020.106041
0378-4266/© 2021 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/)
A. Brauneis, R. Mestel, R. Riordan et al. Journal of Banking and Finance 124 (2021) 106041

high-frequency measures of liquidity with easy to compute low- simple volume-based measures (the transaction frequency and the
frequency measures. USD trading volume). Our goal is to evaluate these measures’ abil-
Characterising liquidity across exchanges is important for in- ity to describe time-series variation in liquidity as well as level dif-
vestors, traders, and hedging strategies that use cryptocurrencies ferences across exchanges. This will allow us to recommend a spe-
(Hu et al., 2019) that can be negatively affected by the costs of cific liquidity proxy using easy-to-access transactions data and an
illiquidity. Additionally, cryptocurrency prices are not integrated easy-to-compute proxy. Using such a proxy offers enormous sav-
across exchanges (Makarov and Schoar, 2020) and the decision to ings compared to high-frequency order book measures.
trade on an exchange is binding as orders cannot easily be re- Our paper augments the literature on low-frequency
routed to exchanges that are more liquid or offering better prices. transactions-based liquidity measures by extending the analy-
With little information about individual exchanges traders may sis to cryptocurrencies, an important and emerging asset class.
have to rely on transactions data such as the daily high, low and We focus on results for bitcoin, the largest cryptocurrency. Results
closing prices to evaluate market quality. for ethereum, which are qualitatively similar in most respects,
We study the accuracy of liquidity measures derived from are discussed in Section 3.7, the corresponding tables and figures
transactions data. To this end we estimate low-frequency measures are in the appendix. We use the cost of a roundtrip trade in
derived from aggregate transactions data4 (prices and volumes) addition to quoted spreads, effective spreads and price impacts as
and compare them to high-frequency measures of transaction costs benchmark measures. The round-trip measure provides estimates
and price impact calculated from order book data. Our objective is of the execution costs for large trades and is thus important for
to identify the transactions-based measure that best describes ac- evaluating some trading strategies (Hu et al., 2019) and factor
tual liquidity on a cryptocurrency exchange. models (Liu et al., 2019).
Data on best bid and ask prices and order books are hard to The results suggest that the proxies that use high, low and clos-
obtain and process.5 As such, few papers use full order book data ing prices, the Corwin and Schultz (2012) and Abdi and Ranaldo
to study the liquidity of cryptocurrency markets (Brauneis et al., (2017)) estimators, best capture the time-series variation in cryp-
2019; Dyhrberg et al., 2018; Hautsch et al., 2018; Makarov and tocurrency liquidity. These measures work for all data frequencies,
Schoar, 2020 and Marshall et al., 2019). A number of low-frequency exchanges (Bitfinex, Bitstamp, Coinbase Pro), benchmark measures
measures have been developed and used to assess bond, commod- (quoted spread, effective spread, price impact, cost of a roundtrip
ity, foreign exchange and equity market liquidity (e.g. Fong et al., trade) and for both bitcoin and ethereum.9 Average time-series
2017; Goyenko et al., 2009; Karnaukh et al., 2015; Johann, Theissen, correlations describe the average relationship between benchmark
2020; Marshall et al., 2012; Schestag et al., 2016). Cryptocurrency measures and proxies but do not capture the relationship for ex-
markets have characteristics that differ from traditional markets,6 treme liquidity events that may be important for investment and
suggesting that liquidity formation on cryptocurrency exchanges hedging strategies. We use quantile dependence plots to under-
may differ from those of other asset markets. stand how well transactions-based liquidity measures capture the
We use a novel and comprehensive set of continuous transac- time-series properties of the benchmark measures across the dis-
tions data and order book snapshots comprising the 50 best bids tribution. This is an important extension since the relative perfor-
and asks for two major cryptocurrencies (bitcoin and ethereum) mance of liquidity proxies might be different depending on the
and three large exchanges (Bitfinex, Bitstamp and Coinbase Pro) liquidity regime. We also perform several sample splits and find
over a two-year period. First, we use these data to construct high- similar performance rankings of our liquidity proxies for high and
frequency measures of transaction costs and price impact. These low volume and volatility periods. Given the extreme volatility as-
measures serve as our liquidity benchmarks. In a second step, sociated with bitcoin and cryptocurrency markets more generally,
we use transactions data (prices and volumes) and calculate var- identifying liquidity proxies that perform well in different volatility
ious liquidity proxies at lower frequencies (1 h, 1 day, and 15 and volume scenarios is an important contribution.
days,7 respectively). Data to compute the measures are collected The popular Amihud (2002) illiquidity ratio does not capture
at the 1-minute, 1-hour and 1-day frequency.8 Individual low- the time-series variability of liquidity in the cryptocurrency mar-
frequency measures have been used to describe liquidity in cryp- kets.10 The poor performance is driven by the relationship be-
tocurrency markets (e.g. Brauneis and Mestel, 2018; Dimpfl, 2017; tween volume and liquidity that is assumed to be negative in
Fink and Johann, 2014; Shi, 2018) but the relative benefits of each Amihud (2002) and is positive in cryptocurrency markets. The
is not well understood. The most commonly used of these low- positive relation between bid-ask spreads and volume is at odds
frequency measures are the Roll (1984) serial covariance estima- with most theoretical predictions but has recently also been docu-
tor and the Amihud (2002) illiquidity ratio. We extend the analysis mented by Bogousslavsky and Collin-Dufresne (2020) for large US
to include low-frequency measures based on high and low prices stocks. 11
(Abdi and Ranaldo, 2017; Corwin and Schultz, 2012), the volatility- Similar to Hasbrouck and Seppi (2001), we construct a compos-
to-volume measure proposed by Kyle and Obizhaeva (2016) and ite estimator, the first principal component of the low-frequency
proxies, and find that it does not generally improve on the perfor-
mance of the best individual proxies.
4
Below, we use the term transactions-based measure synonymously with the
The measures that best describe the level of the bench-
terms low-frequency measure and liquidity proxy. mark measures are the Kyle and Obizhaeva (2016) and
5
It can be downloaded in real time from the REST APIs of each cryptocurrency Amihud (2002) estimators. In this application the Corwin and
exchanges or it can be purchased from vendors such as Kaiko.
6
e.g. markets are highly fragmented and weakly regulated; they are open 365
days a year and 24 h each day, they allow direct market access for all traders; trad-
9
ing platforms allow a direct transfer of fiat currency from and to bank accounts or Consistent with our results, Karnaukh et al. (2015) report the Corwin and
credit cards, and transactions are cleared and settled by exchanges directly; margin Schultz (2012) measure to have the highest correlation with high-frequency bid-ask
trading and short-selling is uncommon. spreads in FX markets.
7 10
Providing results at the monthly frequency is infeasible because high-frequency Conceptually, the Amihud (2002) illiquidity ratio is a proxy for the price impact,
data for cryptocurrencies are limited to 24 months. not for the spread. We find that it also performs poorly in tracking the time-series
8 variation of price impacts, a component of the effective spread.
In contrast to CRSP for equity markets that provides daily prices, cryptocurrency
11
prices are available at higher than daily frequency. For example, the site cryptodata- Amihud and Noh (forthcoming) present evidence of occasions where the illiq-
download.com provides free data for many currency pairs and trading venues at the uidity ratio and the inverse of volume move in opposite directions, implying that
hourly frequency. volume increases while liquidity, as measured by the illiquidity ratio, decreases.

2
A. Brauneis, R. Mestel, R. Riordan et al. Journal of Banking and Finance 124 (2021) 106041

Schultz (2012) and Abdi and Ranaldo (2017) estimators perform Table 1
Number of transactions and order book snapshots for three exchanges and
poorly. We find that the values obtained for these two estimators
the two cryptocurrencies bitcoin (BTC) and ethereum (ETH), both traded
and for the Roll (1984) estimator are negatively related to the against the USD. The sample period is 12/16/2017 to 12/16/2019.
data frequency, a finding that has been documented previously for
Number of transactions Number of order books
the Roll estimator (Roll, 1984; Harris, 1990) but has, to the best
of our knowledge, not been documented for the high-low spread BTC ETH BTC ETH
estimators. Bitfinex 37,148,069 23,820,982 7,271,422 7,336,639
An important application of liquidity proxies is to select an ex- Bitstamp 15,310,565 5,207,030 6,913,021 6,611,326
ecution venue among a number of alternatives. We use the low- Coinbase Pro 38,241,727 24,420,844 8,186,287 8,186,287
frequency estimators to rank trading venues according to their liq-
uidity. We find that the Amihud (2002) illiquidity ratio and the A potential problem associated with transactions data from
Kyle and Obizhaeva (2016) estimator best replicate the ’true’ rank- cryptocurrency exchanges are fake data. A widely cited report by
ing when compared to the ranking generated using high-frequency Hougan et al. (2019) argues that up to 95% of exchange-reported
order book measures.12 trading volume in bitcoin might not represent economically mean-
Our findings are useful for researchers, investors, traders, trad- ingful transactions or might even be plain fake. Collecting unique
ing venue operators and regulators to understand liquidity lev- high-frequency trade and order book data for bitcoin the authors
els and dynamics on cryptocurrency exchanges with relatively subject 83 cryptocurrency exchanges to several tests to identify ex-
easy to acquire and process aggregate price and volume data. changes that are likely to overstate trading volume. Only 10 ex-
Investors seeking the most liquid exchanges are best advised changes passed all the tests and are characterized as “real vol-
to use the Amihud (2002) illiquidity ratio or the Kyle and ume” exchanges. The three trading venues that we consider in the
Obizhaeva (2016) estimator. These two measures also provide good present study all belong to the latter group.
approximations of the level of liquidity. They are the measures From each trading venue we download data for two currency
of choice for market participants attempting to estimate execu- pairs, bitcoin versus US dollar (BTCUSD) and ethereum against
tion costs to evaluate trading strategies. In contrast, traders seek- US dollar (ETHUSD). The data set includes the price and the cor-
ing to time the liquidity of cryptocurrency markets and enter or responding dollar trading volume for each transaction, a UNIX
exit when markets are liquid should use the Abdi and Ranaldo time stamp, a unique exchange-specific ID and a trade indica-
(2017) and Corwin and Schultz (2012) estimator as they best cap- tor which indicates whether a transaction was buyer-initiated or
ture the time-series variability of the quoted and effective bid-ask seller-initiated. Table 1 lists the total number of transactions and
spread. Regulators and trading venue operators can learn from our order book snapshots for both currencies and all three markets. A
paper about how exchanges compare across time and use aggre- total of 90.7 (53.4) million transactions were executed for bitcoin
gate measures to study the impact on liquidity of regulatory or (ethereum) during the investigation period, most of them on Coin-
market changes. Researchers can use our results to guide their base Pro while Bitstamp reports least transactions.
choice of liquidity measures in empirical studies on cryptocurrency We observe several time intervals with gaps in the data. These
markets. Overall, our results suggest that the measure used should may be due to actually missing trading activity, technical problems
depend on the question being asked, as there is not (yet) a univer- (failure of the internet connection, no response from the server
sally best measure. etc.), or exchange-specific trading halts (e.g. due to maintenance,
The remainder of the paper is organized as follows. updates or hacker attacks). We identify between 6,329 (Coinbase
In Section 2 we describe our data and methodology, Pro - BTC) and 187,254 (Bitstamp - ETH) intervals without transac-
Section 3 presents the results, and Section 4 concludes. tion data exceeding 60 s (1 min), between 2,641 and 5,920 inter-
vals exceeding 600 s (10 min) and between 1,573 and 2,199 inter-
2. Data and methodology vals exceeding 1,800 s (30 min). 13 .
Table 2 provides descriptive statistics. Trading activity is
2.1. Data markedly higher for BTCUSD than for ETHUSD on all three ex-
changes. The differences are more pronounced for the USD trading
We compile a high-frequency data set that covers the two-year volume than for the number of transactions, implying that the av-
period from 12/16/2017 0 0:0 0 UTC to 12/16/2019 0 0:0 0 UTC, a to- erage trade size is smaller for the currency pair ETHUSD. With re-
tal of 730 trading days (17,520 h). Over this period we used Matlab spect to the number of transactions Coinbase Pro (Bitstamp) is the
to continuously access the public and freely accessible REST APIs most active (least active) exchange for both currencies. However,
of three large trading venues, Bitfinex, Bitstamp and Coinbase Pro average daily USD volume is highest on Bitfinex and lowest on
(formerly known as GDAX). These are among the largest cryptocur- Bitstamp. Concerning average USD trade size Bitstamp (Coinbase
rency spot trading platforms. All three venues operate an electronic Pro) has the highest (lowest) level for the pair BTCUSD, while for
central limit order book with orders being matched based on price ETHUSD Bitfinex (Coinbase Pro) shows the highest (lowest) level.
and time priority. The standard deviation of ETHUSD returns is larger than that of
The REST APIs provide live information on transactions and the BTCUSD returns on all three venues. Across venues, price returns
current state of the order book. All public endpoints at each of are most volatile on Bitstamp and least volatile on Coinbase Pro.14
these exchanges use GET requests for different types of informa- Besides transactions data we retrieve order book data from the
tion. We request records on ‘Trades’ / ‘Transactions’ and the ‘Order three trading platforms. Specifically, we collect the 50 best bid
book’. Depending on the venue, request parameters vary. For in- and best ask prices with corresponding volumes, resulting in a
stance, Bitstamp only provides the full order book (with usually total of 14.6 million (13.5 million, 16.4 million) order book snap-
thousands of entries) whereas order book requests at Bitfinex and
Coinbase Pro may be limited to the 50 best price levels on each 13
For Bitstamp and Coinbase Pro we have a 19 days lack of data in September
side of the market. 2018
14
We note that the differences in the standard deviation of returns may reflect
liquidity differences because the return standard deviation is affected by bid-ask
12
The Corwin and Schultz (2012) estimator does very well for bitcoin but does bounce. In fact, as shown in Table 3 below, bid-ask spreads are largest on Bitstamp,
poorly for ethereum. a result that has also been confirmed by Brauneis et al. (2019).

3
A. Brauneis, R. Mestel, R. Riordan et al. Journal of Banking and Finance 124 (2021) 106041

Table 2
Descriptive statistics for transactions data and the quote midpoint for the pairs BTCUSD and ETHUSD. Number of transactions and dollar volume refer to daily
averages, the standard deviation of price returns σ (r ) as well as the standard deviation of the quote midpoint returns σ (MQ ) refer to returns normalized to
60 s. The sample period is 12/16/2017 to 12/16/2019.

BTCUSD ETHUSD

# TX [10 0 0] dollar vol [mio USD] σ (r ) [bp] σ (MQ ) [bp] # TX [10 0 0] dollar vol [mio USD] σ (r ) [bp] σ (MQ ) [bp]
Bitfinex 51.15 149.87 13.44 13.51 32.74 57.46 16.88 15.87
Bitstamp 21.67 66.93 15.19 12.28 7.36 11.68 26.44 14.53
Coinbase Pro 54.05 87.73 12.78 13.23 34.61 40.17 15.22 15.27

Table 3
Descriptive data for benchmark liquidity measures for the pair BTCUSD used in the empirical analysis in the empirical analysis. Values are
based on our complete record of all transactions and order book snapshots and represent averages at a daily resolution. The table reports
descriptive statistics for the quoted spread (QS), the effective spread (ES), the price impact (PI) and the percentage cost of a roundtrip trade
(CRT). Q1 (Q3) denotes the first (third) quartile. The unit of measurement is basis points. The sample period is 12/16/2017 to 12/16/2019.

exchange mean std. dev. Q1 median Q3 num daily obs.

Bitfinex QS 0.721 0.691 0.293 0.456 0.986 695

ES 1.040 1.075 0.459 0.743 1.353
PI 0.394 0.330 0.159 0.303 0.527
CRT 3.833 1.927 2.351 3.649 4.677
Bitstamp QS 6.553 3.270 3.888 6.377 8.163 663
ES 6.992 3.332 4.451 6.675 8.456
PI 0.492 0.415 0.214 0.363 0.644
CRT 13.23 5.765 9.601 11.88 15.70
Coinbase Pro QS 0.636 3.965 0.061 0.194 0.652 668
ES 1.173 1.352 0.420 0.816 1.456
PI 0.395 0.491 0.120 0.252 0.485
CRT 3.407 4.202 2.272 2.945 3.911

shots for Bitfinex (Bitstamp, Coinbase Pro) for the two cryptocur- used to calculate the transactions-based measures (to be described
rencies under investigation (see Table 1). As for the transactions below) at the hourly, daily and 15-day frequencies, respectively.
data we observe a considerable number of intervals without or- Because the three trading venues under investigation are located
der book snapshots. There are between 13,038 (Coinbase Pro) and in different time zones we follow coinmarketcap.com and define a
84,461 (Bitfinex) intervals without data exceeding 60 s. The num- trading day as lasting from 0 0:0 0 UTC to 23:59 UTC. For an interval
bers of intervals without order book snapshots exceeding 600 s to be included in the analysis we require that data are available for
and 1,800 s are roughly equal across the three exchanges and at least 80% of the subintervals. Thus, when we aggregate minute-
amount to approximately 3,300 and 2,200, respectively. The stan- by-minute (hour-by-hour, daily) data to the hourly (daily, 15-day)
dard deviations of quote midpoint returns are similar across trad- frequency we require at least 48 min (19 h, 12 days) with valid
ing venues and are generally higher for ETHUSD than for BTCUSD data. The final data set roughly comprises 12,500 hourly intervals,
(see Table 2). 670 daily intervals and 46 15-day intervals, respectively.
Besides the transactions-based measures we need to calculate
2.2. Measures of liquidity benchmark measures. To this end we use the complete record of
all transactions and all order book snapshots and calculate average
The purpose of our paper is to assess and compare the accuracy quoted and effective spreads, price impacts and cost of a roundtrip
of transactions-based measures of liquidity. In doing so we take the trade (to be defined below) at the hourly, daily and 15-day fre-
perspective of a researcher who has access to data on open, high, quency.
low and close prices and on the number of transactions and the In the sequel we first describe the high-frequency measures
dollar trading volume. which we use as benchmark measures and then the transactions-
For our analysis we need to specify (a) the frequency at which based measures that we wish to evaluate.
these data are available (measured by the length of the subintervals 2.2.1. High-frequency benchmark measures
i in the sequel) and (b) the frequency at which the transactions- • Percentage Quoted Spread (QS)
based measures are calculated (measured by the length of the in- The percentage quoted spread is the difference between the
tervals t). Unlike for other financial markets (e.g. stock markets), best ask price P a and the best bid price P b of each order book
price and volume data for cryptocurrencies are easily available for snapshot, divided by the quote midpoint MQ = (P b + P a )/2 and
higher than daily frequencies. We therefore choose three distinct averaged over all observations in the interval
setups. Nt Pja −Pjb
QSt = N1 j=1 MQ
.
t j
• Data are available at the 1-minute frequency and are used to
The subscript j denotes the jth order book snapshot in interval
estimate transactions-based liquidity measures at the hourly
t and Nt is the total number of order book snapshots in interval
frequency.
t.
• Data are available at the 1-hour frequency and are used to es-
• Percentage Effective Spread (ES)
timate liquidity measures at the daily frequency.
To estimate the effective bid-ask spread we combine order book
• Data are available at the daily level and are used to calculate
snapshots with the first transaction that occurs after the snap-
liquidity measures at a 15-day frequency.
shot.15 The price of this transaction is denoted P + . The average
To construct the data set we use our record of all transactions
and extract the open, high, low and close price as well as the num- 15
When there is more than one transaction between two order book snapshots
ber of transactions and the dollar volume at the respective fre- we only use the first of these transactions. We further require the transaction to
quencies of one minute, one hour and one day. These data are then happen within 60 sec after the order book record.

4
A. Brauneis, R. Mestel, R. Riordan et al. Journal of Banking and Finance 124 (2021) 106041

percentage effective spread in interval t is then calculated as price impact on these two exchanges amounts to approximately
Nt+ 2·|Pj+ −MQ j | 35% of the effective spread, implying that the suppliers of liquidity
ESt = N1+ j=1 MQ
t j earn a small positive realized spread on average. Quoted spreads,
Nt+ is the number of order book snapshots that are followed by effective spreads and the cost of a roundtrip trade are much larger
a transaction before the next order book snapshot is recorded. and more volatile on Bitstamp than on the other two exchanges.19
• Percentage Price Impact (P I) The price impact, on the other hand, is only slightly larger on Bit-
To estimate the price impact we use data sequences consist- stamp than on the other two venues. Thus, suppliers of liquidity
ing of an order book snapshot, the first transaction after the on Bitstamp appear to be earning significant realized spreads.
snapshot and the subsequent order book snapshot. The per- We also calculated correlations between our benchmark liquid-
centage price impact is then calculated as the signed percent- ity measures (not tabulated). Using the daily data set for the pair
age change in the quote midpoint from the pre-transaction or- BTCUSD (results for other frequencies as well as for ETHUSD are
der book snapshot to the post-transaction snapshot,16 averaged similar and available upon request) we find the highest correlation
over all data sequences (as defined above) in the interval between QS and CRT (exchange average: 0.89). Note that QS is a
Nt −1 + special case of CRT, measuring the cost of a roundtrip for trades
1  (MQ j+1 − MQ j )
P It = + Q+ · not exceeding the quoted inside depth. The high correlation be-
Nt − 1 j MQ j tween QS and CRT is therefore not surprising and in line with sim-
j=1
ilar results from stock markets (Irvine et al., 20 0 0). We take it as
where Q+
j
denotes the trade indicator (+1 for a buyer-initiated evidence that QS is a good indicator for market liquidity not only
trade and −1 for a seller-initiated trade) of the transaction oc- at but also beyond the inside spread. PI has the lowest correla-
curring after the order book snapshot j Conrad, Wahal (2020).17 tions with QS (exchange average: 0.49) and CRT (exchange aver-
• Percentage cost of a roundtrip trade (CRT (Y )) age: 0.55). This confirms that PI captures a different dimension of
To assess the liquidity for larger trades we use the order book market liquidity than the spread measures.
data to calculate the weighted average prices at which a buy Fig. 1 shows the evolution of the hourly benchmark measures
and a sell order of a given size Y would execute. The weighted for the pair BTCUSD over time. Overall the patterns reveal strong
average price for executing a transaction of size Y USD given similarities in the liquidity measures, both within and between the
K
k=1 k k
A ·V three exchanges. Starting from higher levels in December 2017, liq-
the current state of the order book is defined as  K sub-
V
K k=1 k uidity measures continuously decrease until the end of Novem-
ject to k=1 Ak · Vk = Y where Ak and Vk are the price and vol- ber 2018 where they increase sharply. Throughout 2019 liquidity
ume of the kth order, respectively. Note that the K th order may is lower and more volatile than during most of 2018.
be subject to partial execution, depending on the outstanding Corresponding to Table 3, Table 9 in the appendix provides de-
dollar volume required to entirely fill the transaction volume Y . scriptive statistics for the benchmark liquidity measures for the
We set Y equal to the 99% quantile of the corresponding (aggre- currency pair ETHUSD at the daily frequency (again, results for the
gate) trade size distribution. For the currency pair BTCUSD this other frequencies are essentially identical and are available upon
value is approximately equal to USD 32,100, while for ETHUSD request). As for BTCUSD we find the levels of our four benchmark
Y roughly corresponds to USD 17,400. measures to be very similar on Bitfinex and Coinbase Pro. Aver-
To estimate the cost of a roundtrip trade of size Y, CRT (Y ), we age quoted spreads are about twice as high as those for the pair
calculate the weighted average prices for a market buy order BTCUSD. Average effective spreads for ETHUSD are below 2.2 bps
and a market sell order of size Y and then express the differ- on Bitfinex and Coinbase Pro, but again are higher than those for
ence between the two prices as a fraction of their midpoint. Fi- BTCUSD. Average price impacts on both trading venues amount to
nally, we calculate an equally-weighted average across all order roughly 30% of effective spreads , again implying that the suppliers
book snapshot in interval t. of liquidity earn a small realized spread on average.
The CRT (Y ) measure is conceptually similar to the quoted bid- As for BTCUSD, Bitstamp is substantially less liquid for ETHUSD
ask spread. However, while the quoted spread measures the than the other two exchanges.20 The average quoted spread (effec-
transaction costs of a small trade (defined as a trade the size tive spread) amounts to 13.17 bps (13.45 bps). Again the average
of which does not exceed the quoted depth), the CRT (Y ) mea-
sure estimates the execution costs of a trade of size Y .
quoted spread (effective spread) of 1.05 (0.31) basis points. USDCAD is the least
Table 3 shows descriptive statistics for the benchmark measures liquid of the analyzed pairs with respect to the quoted spread (8.27 basis points)
for BTCUSD obtained from daily data (results at the hourly and 15- while AUDUSD has the highest effective spread (1.38 basis points.
day frequency are virtually identical and available upon request). 19
Because of the higher execution costs on Bitstamp traders may want to avoid
Overall, the percentage trading costs in the cryptocurrency markets Bitstamp. However, there are several reasons why we may still observe significant
trading activity on Bitstamp. First, most transactions are small. The median trade
are very low. Average quoted and effective spreads on Bitfinex and
size on Bitstamp is 354 USD (the corresponding values for Bitfinex and Coinbase
Coinbase Pro are below 1.2 bps while the cost of a roundtrip trade Pro are 500 and 140 USD, respectively). Assuming a quoted spread of 6.6 bps (the
of size USD 32,100 on these two venues are below 4 bps.18 The median quoted spread on Bitstamp), the execution costs of a median-sized trade on
Bitstamp amount to 0.14 USD (354 USD multiplied by the half-spread), an amount
which traders may deem negligible. Second, there are frictions beyond the bid-ask
16
From the numbers in Table 1 it follows that we observe an order book snapshot spread. For example, trading venues differ in the ways how traders can transfer
every nine seconds on average. Thus, the horizon over which we calculate the price and withdraw fiat money to and from their accounts. These differences can result
impact is slightly less than nine seconds on average. This is in line with Conrad and in cost and speed differences between the exchanges. Third, not all traders are free
Wahal (2020) who recommend to use a horizon of no more than 15 s for liquid to choose where to trade. For example, Bitfinex did not accept US residents as cus-
stocks. tomers during our sample period. Fourth, traders may prefer to trade on a venue
17
As before, when there is more than one transaction between two order book in or close to their home country, e.g. because they are more familiar with the leg-
snapshots we only use the first of these transactions. Also, we discard order book islative regime.
observations more than 60 sec apart. We lose one observation in each interval be- 20
Evidently, ETHSUD is a rather infrequently traded pair on Bitstamp (roughly 5
cause the last order book snapshot in an interval is discarded as it is not followed million transactions over our investigation period, compared to roughly 24 million
by another snapshot in the same interval. on Bitfinex and Coinbase Pro) which is why we only have 4410 observations of
18
By way of comparison: Mancini et al. (2013) report liquidity for the 9 most hourly data on Bitstamp that match the 80% data availability criterion (compared
traded exchange rates on the EBS platform over the period January 2007 to De- to more than 11,920 one hour intervals that match this criterion on Bitstamp for
cember 2009. They find EURUSD to be the most liquid rate with a mean relative the pair BTCUSD.

5
A. Brauneis, R. Mestel, R. Riordan et al. Journal of Banking and Finance 124 (2021) 106041

Fig. 1. This figure plots a 72 h moving average of the benchmark liquidity measures for the pair BTCUSD at an hourly resolution. Quoted spread (gray solid line), effective
spread (light gray dashed line) and price impact (black solid line) are shown on the left axis, the percentage cost of a roundtrip trade (black dashed line) is shown on the
right axis. The unit of measurement is basis points. The sample period is 12/16/2017 to 12/16/2019. Note that for Bitstamp and Coinbase Pro we have a 19 days lack of data
in September 2018.

Table 4
Descriptive data for proxy liquidity measures for the pair BTCUSD used in the empirical analysis at a
daily resolution. The table reports descriptive statistics for the number of transactions (TX), the dollar
volume ($ Vol, million USD), the Amihud measure (Amihud, values ∗ 1e6), Roll’s returns based mea-
sure (Roll_r, basispoints), Roll’s price based measure (Roll_p), the Kyle and Obizhaeva measure (Kyle,
values ∗ 1e3), the Corwin and Schultz measure (CS, basispoints) and the Abdi and Ranaldo measure
(AR, basispoints). The sample period is 12/16/2017 to 12/16/2019.

exchange mean std. dev. Q1 median Q3 no. daily obs

Bitfinex TX 53,175 45,859 22,622 37,453 67,338 695

$ Vol 155.1 186.9 38.01 83.25 190.3
Amihud 0.015 0.186 0.001 0.002 0.004
Roll_r 37.01 45.46 0 25.55 54.57
Roll_p 30.71 46.28 0 17.03 40.27
Kyle 0.077 0.024 0.060 0.073 0.090
CS 21.63 19.49 9.158 15.67 27.45
AR 23.60 21.23 10.62 17.02 28.81
Bitstamp TX 22,873 18,369 9,705 17,697 30,353 663
$ Vol 70.44 68.12 24.98 46.50 92.63
Amihud 0.038 0.347 0.002 0.003 0.006
Roll_r 37.03 45.66 0 25.69 52.68
Roll_p 31.14 48.65 0 16.15 40.18
Kyle 0.093 0.029 0.073 0.088 0.109
CS 24.25 18.12 12.14 19.20 29.73
AR 24.06 20.36 10.92 17.89 28.28
Coinbase Pro TX 56,928 33,687 35,507 46,277 67,235 668
$ Vol 91.78 96.97 30.50 61.93 115.1
Amihud 0.006 0.059 0.001 0.002 0.003
Roll_r 35.58 46.34 0 22.77 51.40
Roll_p 29.82 48.65 0 13.73 38.51
Kyle 0.085 0.026 0.067 0.081 0.098
CS 19.44 16.91 8.527 14.64 24.10
AR 22.74 20.02 10.13 16.66 26.74

price impact is not much larger on Bitstamp than on Bitfinex and gregated to one liquidity estimate for each interval t. We use the
Coinbase Pro, implying substantial realized spreads to be earned following transactions-based measures.
by liquidity suppliers on Bitstamp.

2.2.2. Transactions-based proxy measures • Number of transactions (T X)

As noted previously, all transactions-based liquidity measures For each interval t we calculate the unweighted average of the

are calculated from data on open, high, low and closing prices as number of transactions in the subintervals i, T Xt = 1I i T Xt,i ,
well as the number of transactions and the dollar trading volume where I denotes the number of subintervals in interval t.
for each subinterval i. The data for the subintervals are then ag- • Dollar Volume ($V ol)

6
A. Brauneis, R. Mestel, R. Riordan et al. Journal of Banking and Finance 124 (2021) 106041

Our second transactions-based measure is the unweighted and low price data from two adjacent subintervals i and i + 1.
average of the reported dollar transaction volume $V olt = It is defined
 as
1 
I i $V olt,i in all subintervals i belonging to interval t. ARi = max{4(ci − p̄i )(ci − p̄i+1 ), 0}
• The Amihud (2002) illiquidity ratio (Amihud) The ARt estimator for interval t is the average of the ARt,i mea-
Amihud (2002) illiquidity ratio for each subinterval is the ab- sures for all adjacent subintervals i in t,
1 I−1
solute return (measured from the opening price to the clos- ARt = I−1 i=1 ARt,i .
ing price of the subinterval) divided by the dollar trading vol-
 |C /O −1| Table 4 reports descriptive statistics for our proxy liquidity
ume in the subinterval, Amihudt = 1I i t,i$Volt,i , where Ot,i measures at the daily frequency (results at the hourly and 15-day
t,i
and Ct,i denote the opening and closing price in subinterval i frequency are again available upon request). When interpreting the
in t, respectively. The illiquidity ratio for interval t is the un- numbers it should be kept in mind that several measures (Abdi
weighted average of the ratios for the subintervals i in t. We and Ranaldo, 2017; Corwin and Schultz, 2012; Kyle and Obizhaeva,
note that conceptually the illiquidity ratio is a measure of price 2016; Roll, 1984) estimate the effective bid-ask spread while the
impact. However, in empirical applications it is routinely used Amihud (2002) illiquidity ratio is a measure of price impact and
as a proxy for liquidity at large. the two volume metrics measure the number of trades and the
• The Roll (1984) serial covariance estimator (Rol l ) dollar trading volume, respectively. However, even among those
The Roll (1984) estimator is based on the serial covariance of measures that estimate the effective spread there are large dif-
successive price changes. For each interval t we obtain one ferences. The Roll (1984) estimator delivers the largest and the
spread estimate from the closing prices of all subintervals i in Kyle and Obizhaeva (2016) estimator delivers the smallest spread
t. If the serial covariance is positive we set the estimator to 0. estimates. We will compare the mean values shown in Table 4 to
We calculate two versions of Roll’s measure, one based on price the effective spread calculated from high-frequency quote data in
changes and one based on returns. The formal expression for Section 3.5 below.
the return-based estimator is For the currency pair ETHUSD descriptive statistics for our
 proxy liquidity measures are reported in Table 10 in the appendix.
Ct,i Ct,i−1 As for our benchmark measures the results indicate that the pair
Rollt = 2 · − min(cov[ , ], 0 )
Ct,i−1 Ct,i−2 ETHUSD is less liquid than BTCUSD: volume-based proxy measures
show lower values, while price-based measures are higher.21 We
where  is the first difference operator. In the results section, will discuss these results in more detail in Section 3.7.
Roll_p (Roll_r) refers to the price- (return-)based version, re-
spectively. 3. Results
• The Kyle and Obizhaeva (2016) estimator (Kyle)
Kyle and Obizhaeva (2016) derive an illiquidity index based on We present the results in six steps. We first report time-
the ratio of volatility to dollar volume of an asset within a given series correlations between the transactions-based proxies and the
interval. It is defined as benchmark measures. Correlations are a global measure of lin-
 1 / 3 ear dependence. To analyze whether the dependence structure is
σt,i2 (r ) different in the tails of the distribution we analyze, in step 2,
Kylet = 
$V olt,i quantile dependencies based on the empirical distribution func-
i
tions. In a third step we aggregate the transactions-based measures
into a composite measure, the first principal component of the
where the volatility estimator σt,i2 (r ) is the mean of the squared
transactions-based measures, and analyze whether it has higher
returns of all subintervals i in interval t.
time-series correlation with the benchmark measures than the
• The Corwin and Schultz (2012) estimator (CS).
best of the individual transactions-based measures. In order to in-
The CS estimator is calculated from the high and low prices of
vestigate whether the performance of the transactions-based mea-
two adjacent subintervals i, i + 1. It is defined as
exp(α )−1 ) sures depends on the specific market environment we then, in step
CSi,i+1 = 2(1+ exp(α )

2 
2 4, split our sample along several dimensions. Specifically, we esti-
√ √ 
α = 2β −√ β −
3−2 2
γ√
, β = ln L i
3−2 2
H H
+ ln L i+1 , γ =
i+1
mate time-series correlations between the proxies and the bench-

2
i mark measures for the first and the second half of our sample pe-
ln
Hi,i+1 riod, for high and low return periods, high and low volatility pe-
Li,i+1
riods and high and low volume periods. Next we report the mean
Hi and Li denote the high and low prices, respectively, in subin- absolute errors and root mean squared errors of the transactions-
terval i, while Hi,i+1 and Li,i+1 refer to the high and low price, based measures. Finally, to capture the cross-sectional dimension,
respectively, of two adjacent subintervals i and i + 1. We follow we analyze how frequently the liquidity ranking across the ex-
Corwin and Schultz (2012) and set negative values of the proxy changes produced by the transactions-based measures is equal to
to zero. The CSt estimator for period t is the unweighted aver- the ranking produced by the benchmark measures.
age of all CS estimators for adjacent subintervals in t. In Sections 3.1–3.6 we present results for the currency pair BTC-
Corwin and Schultz (2012) propose a method to adjust their USD in detail; qualitative results for the pair ETHUSD are similar in
estimator for the overnight trading halt. We do not need to most respects and are summarized in Section 3.7.
implement this modification because cryptocurrency exchanges
operate 24 h a day and seven days a week. There are thus no 3.1. Time-series correlations
regular trading halts.
• The Abdi and Ranaldo (2017) estimator (AR) An accurate transactions-based measure should capture the
Abdi and Ranaldo (2017) propose an estimator based on the time-series variation in liquidity and should thus be positively cor-
natural logarithms of high, low and closing prices in subin-
terval i, denoted hi = ln(Hi ), li = ln(Li ) and ci = ln(Ci ), respec- 121
We note that the price-based Roll measure delivers an estimate of the dollar
tively. Further, denote by p̄i = (hi + li )/2 the midpoint between spread, not of the percentage spread. The numerical values are lower for ETHUSD
the high and the low log prices in subinterval i. We use the than for BTCUSD because the dollar price of ethereum is only a fraction of the bit-
‘two-day corrected’ version of the estimator which uses high coin price.

7
A. Brauneis, R. Mestel, R. Riordan et al. Journal of Banking and Finance 124 (2021) 106041

Fig. 2. This figure shows time series correlations between benchmark measures of liquidity and low-frequency proxy measures for the pair BTCUSD. Values represent simple
averages across the three exchanges Bitfinex, Bitstamp, and Coinbase Pro. Liquidity measures are calculated on a daily basis over the sample period 12/16/2017 to 12/16/2019.

related with the benchmark measures. We therefore estimate time- tive.23 The poor performance of the Amihud (2002) illiquidity ratio
series correlations between the low-frequency measures and the deserves discussion because this ratio is widely used as a measure
high-frequency measures. We do so separately for three exchanges of liquidity in empirical microstructure research. We argue that the
(Bitfinex, Bitstamp and Coinbase Pro) and three time frames (1- lack of correlation between the illiquidity ratio and the benchmark
minute data (1-hour-data, daily data) aggregated to the hourly measures is caused by the strong and positive relation between
(daily, 15-daily) frequency). As it turns out, the results for the three liquidity and trading activity discussed above. The illiquidity ra-
exchanges are very similar. We therefore report averages across tio is based on the presumption that, in a less liquid market, a
the trading venues.22 In the description of the results we empha- given dollar trading volume will have a larger impact on prices
size the findings for the daily data (i.e. hourly data aggregated and will thus result in a larger price change. Put differently, for
to the daily frequency). We believe that most researchers using a given price change higher volume points to a more liquid market
transactions-based liquidity proxies will do so to obtain daily es- and should thus be associated with lower execution costs accord-
timates, and the hourly raw data required to calculate these daily ing to the inherent logic of the measure. However, in the markets
estimates are easily and freely available, e.g. from cryptodatadown- under investigation volume is positively related to execution costs,
load.com. a relation that runs counter the logic of the illiquidity ratio. We
The results are presented in Figs. 2–4. Focusing on the re- wish to reemphasize that the finding of a positive relation between
sults for the daily intervals (Fig. 2) we find that the Abdi and trading activity and execution costs, even though at odds with the
Ranaldo (2017) and Corwin and Schultz (2012) estimators perform predictions of standard theory, is not confined to the cryptocur-
best. The Corwin and Schultz (2012) estimator exhibits the high- rency markets under consideration here. Bogousslavsky and Collin-
est time series correlation and the Abdi and Ranaldo (2017) esti- Dufresne (2020) have recently documented a similar finding for
mator the second-highest correlation with the quoted spread, the large US stocks.
effective spread and the cost of a roundtrip trade, with correla- The results for the two alternative time frames (one-minute
tions ranging from 0.54 to 0.75. The correlation with the price im- data aggregated to the hourly frequency (Fig. 3) and daily data
pact is markedly higher, at 0.86 for both the Corwin and Schultz aggregated to the 15-day frequency (Fig. 4) are similar. The
(2012) and Abdi and Ranaldo (2017) estimator. Corwin and Schultz (2012) and the Abdi and Ranaldo (2017) es-
Interestingly, the number of transactions and the dollar trading timators yield higher correlations with the benchmark measures
volume are highly correlated with the four benchmark measures, than the other transactions-based estimators. The strong and pos-
and particularly so with the price impact (with correlation coef- itive correlations documented above for the daily data frequency
ficients of 0.82 and 0.80 for the number of transactions and the between the transaction frequency and dollar trading volume on
dollar volume, respectively. What is most surprising is the sign of the one hand and the benchmark measures on the other hand per-
the coefficients. Both measures are positively related to the bench- sist at the other data frequencies. In fact, at the 15-daily frequency
mark measures, implying that higher trading activity is associated the volume-based proxies perform better than the other proxies
with higher execution costs. for three out of four benchmark measures.
The other transactions-based proxies achieve much lower corre- The performance of the Kyle and Obizhaeva (2016) estima-
lations. The measure that performs worst in our horse race is the tor is better at higher data frequencies while the Roll (1984) es-
Amihud (2002) illiquidity ratio. It is virtually uncorrelated with the timator appears to perform better at lower frequencies. The
benchmark measures, and the sign of the correlation is even nega-
23
In their study on the foreign exchange market Karnaukh et al. (2015) find that
the Amihud (2002) illiquidity ratio performs reasonably well. When calculating the
illiquidity ratio the authors use the number of transactions as a proxy for the dollar
22
Results for individual exchanges are available upon request. trading volume because they do not have access to volume data.

8
A. Brauneis, R. Mestel, R. Riordan et al. Journal of Banking and Finance 124 (2021) 106041

Fig. 3. This figure shows time series correlations between benchmark measures of liquidity and low-frequency proxy measures for the pair BTCUSD. Values represent
simple averages across the three exchanges Bitfinex, Bitstamp, and Coinbase Pro. Liquidity measures are calculated on an hourly basis over the sample period 12/16/2017 to
12/16/2019.

Fig. 4. This figure shows time series correlations between benchmark measures of liquidity and low-frequency proxy measures for the pair BTCUSD. Values represent
simple averages across the three exchanges Bitfinex, Bitstamp, and Coinbase Pro. Liquidity measures are calculated on a 15-daily basis over the sample period 12/16/2017 to
12/16/2019.

Amihud (2002) illiquidity ratio continues to be the worst- highest data frequencies. The Corwin and Schultz (2012)) and the
performing measure. As explained above, the most likely reason Abdi, Ranaldo (2017) estimators perform best. Each of them sig-
is the positive relation between trading activity and spreads in the nificantly outperforms the second-ranking estimator in six cases.
cryptocurrency markets. At the 15-daily data frequency there is no compelling evidence in
So far we have documented differences in correlations across favor of significant differences between the two best-performing
the different transactions-based measures, but we do not know transactions-based measures.
whether the differences are significant. We therefore now perform
a formal test based on the Fisher r-to-z transformation. Specifi- 3.2. Quantile dependence
cally we test, separately for each of the four benchmark measures
and the three trading venues, whether the correlation between the The correlation between the time series of transactions-based
best-performing proxy and the benchmark measure is significantly proxies and the benchmark measures of liquidity provides a global
higher than the correlation between the second-best performing measure of dependence. However, it is conceivable that a proxy
proxy and the benchmark. The results are reported in Table 5. The measure that fits the benchmark well in times of high liquidity
evidence in favor of significant differences is limited to the two (i.e. in times of low bid-ask spreads) performs poorly in times of
low liquidity and vice versa. We therefore use quantile dependence

9
A. Brauneis, R. Mestel, R. Riordan et al. Journal of Banking and Finance 124 (2021) 106041

Table 5
This table reports the best performing (in terms of correlation) proxy liquidity measure for each of the four benchmark liquidity measures
(second best in parentheses) for the pair BTCUSD. ∗ ∗ /∗ denote statistical significance of the difference between the best and second best
measure’s correlation coefficient using Fisher r-to-z transformation at the 1%/5% level. For example, the entry CS (AR)∗ ∗ for hourly data
of Bitfinex in column QS indicates that CS features the highest correlation with the quoted spread, while AR is ranked second. These
correlations are statistically different at the 1% level. The sample period is 12/16/2017 to 12/16/2019.

QS ES PI CRT
∗∗ ∗∗ ∗∗
hourly Bitfinex CS (AR) AR (CS) AR (CS) Kyle (AR)
Bitstamp AR (CS)∗ ∗ AR (CS)∗ ∗ TX ($Vol)∗ ∗ AR (CS)∗ ∗
Coinbase Pro CS (AR)∗ ∗ CS (AR)∗ ∗ AR (CS)∗ ∗ CS (AR)
daily Bitfinex CS (AR) CS (AR) CS (TX) CS (AR)
Bitstamp CS (AR)∗ CS (AR)∗ CS (AR) CS (AR)∗
Coinbase Pro $Vol (CS) $Vol (TX)∗ ∗ AR (CS) CS (AR)
15-daily Bitfinex TX ($Vol) TX ($Vol) TX ($Vol) CS (AR)
Bitstamp $Vol (TX) $Vol (TX) $Vol (TX) TX ($Vol)
Coinbase Pro Roll_p ($Vol) $Vol (Roll_p) $Vol (TX) CS (AR)

Fig. 5. This figure plots the quantile dependence (averaged across the three trading venues) as a function of the quantile q in steps of 0.01 for the four benchmark liquidity
measures and each of the 8 proxy measures for the pair BTCUSD for daily data. The sample period is 12/16/2017 to 12/16/2019.

to analyze the dependence structure between the benchmark and ing scaled ranks, i.e. we transform the data into ranks and then
proxy measures in more detail. The quantile dependence of or- rescale these ranks onto the unit interval.
der q between two random variables η p and ηb is generally de- Intuitively, quantile dependence works as follows. For any q ≤
fined as the conditional probability that Fp (η p ) is smaller (greater) 0.5 consider the q · T smallest observations for the benchmark
than q given that Fb (ηb ) is smaller (greater) than q for q ≤ 0.5 measure, where T is the total number of observations. Then con-
(q > 0.5):24 sider the q · T smallest values for a transactions-based proxy and
determine the fraction of coinciding values. This fraction is the es-
P [Fp (η p ) ≤ q | Fb (ηb ) ≤ q], for q ∈ (0, 0.5] timate of the quantile dependence,  λqp,b .
λqp,b =
P [Fp (η p ) > q | Fb (ηb ) > q], for q ∈ (0.5, 1 ). We use data at the daily frequency25 to estimate the quantile
dependence separately for each trading venue and then calculate
In our application the subscripts b and p refer to the bench-
averages across venues. We present the results using quantile de-
mark measures and the transactions-based proxies, respectively. q
pendence plots which show the quantile dependence as a func-
denotes a quantile and F denotes the cumulative distribution func-
tion of q. Higher quantile dependence implies a closer relation be-
tion (CDF). An empirical estimate of λq is given by
p,b
tween the benchmark measures and the transactions-based prox-
⎧1  ies. The results for our four benchmark measures and eight proxies
⎨ T
t=1 1[Fp (η
p,t )≤q,F (η
 )≤q] , for q ∈ (0, 0.5], are shown in Fig. 5. The dependence between the benchmark and

λqp,b = Tq b b,t

⎩ 1 T proxy measures is generally stronger in the center of the distri-

1  , for q ∈ (0.5, 1 ).
T (1 − q ) t=1 [Fp (ηp,t )≤q,Fb (ηb,t )≤q] bution and weaker in the tails. The dependence in the tails ap-
 
Fˆj ηˆ j ; j ∈ {b, p} denotes the empirical distribution functions of
the benchmark and proxy measure, respectively. We estimate it us-
25
Results for the hourly frequency are qualitatively similar and are available upon
request. The number of observations at the 15-day frequency is too low to reliably
24
See e.g. Duan et al. (2019). estimate quantile dependence, particularly in the tails of the distributions.

10
A. Brauneis, R. Mestel, R. Riordan et al. Journal of Banking and Finance 124 (2021) 106041

Table 6 3.4. Sample splits

This table reports mean correlations among the first principal component of
proxy liquidity measures and the four benchmark measures for the pair BTCUSD.
Column ‘expl. var’ shows the percentage of total variance explained by the first It may be the case that some of the transactions-based liq-
principal component. Rows denoted ’mean’ contain equally-weighted averages uidity measures perform better under specific circumstances, e.g.
across exchanges. The sample period is 12/16/2017 to 12/16/2019. earlier or later in the sample period or at times of high or
expl. var QS ES PI CRT low volatility. To shed light on this issue we split our sample
along several dimensions. The analysis is performed using the
hourly Bitfinex 56.34 0.761 0.707 0.773 0.576
hourly data aggregated to the daily level.27 We start by separately
Bitstamp 49.75 0.725 0.789 0.718 0.724
Coinbase Pro 57.15 0.582 0.815 0.780 0.630 considering the first and the second half of the sample period,
mean 54.41 0.689 0.770 0.757 0.643 with the resulting sub-samples covering 12/16/2017 0 0:0 0 UTC to
daily Bitfinex 56.38 0.818 0.741 0.859 0.657 12/15/2018 24:00 UTC and 12/16/2018 0 0:0 0 UTC to 12/16/2019
Bitstamp 57.87 0.702 0.760 0.885 0.767
0 0:0 0 UTC, respectively. Subsequently we split the sample into
Coinbase Pro 57.87 0.078 0.764 0.844 0.253
mean 57.37 0.533 0.755 0.863 0.559 terciles according to the signed return (measured by the average
15-daily Bitfinex 53.10 0.765 0.787 0.896 0.539 of the one-hour returns within a daily interval), return volatility
Bitstamp 54.09 0.702 0.724 0.875 0.723 (measured by the standard deviation of the hour-by-hour returns
Coinbase Pro 56.78 0.489 0.778 0.810 0.594 within a one-day interval) and the dollar volume. We then cal-
mean 54.65 0.652 0.763 0.860 0.619
culate separate time-series correlations for the first and the third
tercile.
For each sample split we present results for all four bench-
mark measures. We first calculate time-series correlations for each
pears to be asymmetric, it is higher for larger than for smaller
trading venue and then average the correlations across venues.
values.26
These average correlations are shown in Figs. 6–9. During the
With respect to the ranking of the transactions-based liquid-
second half of the sample period the correlations between the
ity measures the results from the quantile dependence analysis are
transactions-based proxies and the benchmark measures are lower
consistent with the results shown in Fig. 2 above. In particular, the
than those in the first half for most proxies (Fig. 6). This reduc-
Abdi and Ranaldo (2017) and Corwin and Schultz (2012) estimators
tion is much more pronounced for the quoted spread and the cost
perform very well over the entire distribution, i.e. for high as well
of a roundtrip trade than for the effective spread and the price
as low levels of liquidity. The two measures of trading activity, the
impact. These results allow the conclusion that the performance
number of transactions and the dollar volume, also perform well.
of the transactions-based proxies is sufficiently stable over time
As before, the Amihud (2002) illiquidity ratio performs poorly.
when they are used to track the time-series variation of the ef-
fective spread.
The time series correlations between the transactions-based
proxies and the benchmark measures do not differ much between
3.3. Composite estimator
high and low return periods (Fig. 7). The performance tends to be
better in low return periods.
The different transactions-based measures capture different as-
When we consider the sub-samples split by volatility and vol-
pects of liquidity. It is, therefore, conceivable that a combination of
ume (Figs. 8 and 9, respectively) we find that the time series cor-
these measures better captures the time-series variation of liquid-
relations are higher in high volatility and high volume periods for
ity. To test whether this is the case we construct a composite es-
all benchmark measures.
timator based on the eight low-frequency measures. We first stan-
What is most important, though, is that our previous results
dardize all variables by subtracting the mean and dividing by their
concerning the relative performance of the low-frequency liquid-
standard deviation. We then extract the first principal component
ity proxies still hold. The Abdi and Ranaldo (2017) and Corwin and
of the standardized data and estimate the time-series correlations
Schultz (2012) estimators have the best overall performance un-
between the first principal component and the benchmark mea-
der almost all conditions. The Kyle and Obizhaeva (2016) estimator
sures. Table 6 shows the results for each time frame and each of
performs rather well in the low volume periods but is unable to
the three exchanges.
track liquidity across high volatility periods.
The first principal component explains roughly 55% of the vari-
ation in the data. The time-series correlations are highest with the
3.5. Mean absolute errors and root mean squared errors
effective spread and the price impact, with average values (across
the three exchanges) ranging from 0.76 to 0.86. The time-series
The previous analyses have focused on the ability of the
correlations are lower (with values ranging from 0.53 to 0.69)
transactions-based proxies to capture the time-series variability
when one of the other benchmark measures is used.
of the benchmark liquidity measures. An alternative question is
Comparing the results in Table 6 to those in Figs. 2–4 re-
whether the proxy measures are able to accurately estimate the
veals that the best performing individual estimators achieve higher
level of the benchmark measures. Investors take the level of liq-
time-series correlations than the composite estimator for each
uidity into account in their trading strategies and portfolio al-
benchmark measure at the hourly and 15-daily frequencies. At the
locations. Further, because the magnitude of the execution costs
daily frequency the composite estimator performs virtually equally
determines whether a given price difference (e.g. for the same
well as the Corwin and Schultz (2012) and Abdi and Ranaldo
cryptocurrency at two different trading venues) can be prof-
(2017) estimators. We thus conclude that the benefit of calcu-
itably exploited, the level of liquidity is also related to market
lating all transactions-based proxies and aggregating them to a
efficiency.
composite estimator is limited, and is confined to specific data
We use as performance metrics the prediction error between
frequencies.
the liquidity benchmark and the liquidity proxy as measured

27
Results for the hourly frequency are qualitatively similar and are available upon
26
When interpreting the results note that a quantile dependence of 0.5 for q = 0.5 request. The number of observations at the 15-day frequency is too low to split the
is expected when the two distributions are independent. sample into terciles.

11
A. Brauneis, R. Mestel, R. Riordan et al. Journal of Banking and Finance 124 (2021) 106041

Fig. 6. This figure shows time series correlations between benchmark measures of liquidity and low-frequency proxy measures for the first and second half of the sample
for the pair BTCUSD. Values represent simple averages across the three exchanges Bitfinex, Bitstamp, and Coinbase Pro. Liquidity measures are calculated on a daily basis
over the sample period 12/16/2017 to 12/16/2019.

by the root mean squared error (RMSE) and the mean absolute age effective spread very well at the highest data frequency but
error (MAE). Table 7 reports the corresponding results for five very poorly at lower frequencies. The Roll (1984) measure performs
transactions-based measures (Amihud, Roll_r, Kyle, CS and AR) worst.
with respect to the two benchmark measures ES and P I. We ex- When the price impact is used as benchmark measure the
clude as proxies the trading frequency, Roll’s price based estimator Amihud (2002) illiquidity ratio and the Kyle and Obizhaeva
and the dollar volume from the analysis because these measures (2016) estimator yield the best results for all frequencies and both
are obviously unable to directly capture the percentage transaction metrics for the prediction errors. The Roll (1984) estimator and the
costs. two measures based on high and low prices, the Abdi and Ranaldo
When it comes to capturing the level of the effective spread (2017) and Corwin and Schultz (2012) estimators, are unable to
the Kyle and Obizhaeva (2016) estimator performs very well, par- capture the levels of the benchmark measures.
ticularly when data at lower frequencies are used. It has the low- One striking observation is that the levels of the Roll (1984),
est RMSE and MAE for two data frequencies (daily, and 15-daily). Abdi and Ranaldo (2017) and Corwin and Schultz (2012) estimators
Surprisingly, the Amihud (2002) illiquidity ratio comes close to appear to strongly depend on the data frequency. All three mea-
Kyle and Obizhaeva (2016). The Corwin and Schultz (2012) and sures deliver values which increase strongly as we move to lower
Abdi and Ranaldo (2017) proxies capture the level of the percent-

12
A. Brauneis, R. Mestel, R. Riordan et al. Journal of Banking and Finance 124 (2021) 106041

Fig. 7. This figure shows time series correlations between benchmark measures of liquidity and low-frequency proxy measures in the subset of high and low return intervals
for the pair BTCUSD. Values represent simple averages across the three exchanges Bitfinex, Bitstamp, and Coinbase Pro. Liquidity measures are calculated on a daily basis
over the sample period 12/16/2017 to 12/16/2019.

data frequencies. This phenomenon as such is not new. Already in 3.6. Cross-sectional analysis
his original paper, Roll (1984) obtained much larger spread esti-
mates from weekly than from daily data. Harris (1990) argued that One potential application of transactions-based liquidity mea-
the difference can partly be explained by a small sample bias in sures is to compare the liquidity of different trading venues. A
the estimator of the serial covariance.28 The results in Table 7 sug- good proxy measure should produce the same ranking of the
gest that the high-low spread estimators developed by Corwin and venues as the benchmark measures. Therefore, in order to eval-
Schultz (2012) and Abdi and Ranaldo (2017) are subject to a similar uate the low-frequency measures we simply analyze how fre-
bias. quently the liquidity ranking across trading venues produced by
the transactions-based measures is equal to the ranking produced
by the benchmark measures. We perform the analysis separately
28
Specifically, he showed that the expected value of the serial covariance esti- for each exchange pair (Bitfinex/Bitstamp, Bitfinex/Coinbase Pro,
mator is E (SCov ) = −s4 − −nσ where s is the spread, σ 2 is the variance of price
2 2

and Bitstamp/Coinbase Pro) and for each time frame. For each in-
changes and n is the number of observations. The bias in the serial covariance esti-
mator, − −nσ , increases with the square of the observation interval. Under ideal con-
2 terval (one hour, one day, 15 days) and each exchange pair we
ditions (i.e. i.i.d. returns and continuous trading seven days a week, as is the rule in record the corresponding liquidity ranking based on the bench-
cryptocurrency markets), the variance of weekly price changes is seven times the mark measures and based on the transactions-based proxies and
variance of daily price changes while the number of observations is one seventh. then simply count the fraction of identical rankings. By chance, this
Consequently, the bias in weekly data is 49 times the bias in daily data.

13
A. Brauneis, R. Mestel, R. Riordan et al. Journal of Banking and Finance 124 (2021) 106041

Fig. 8. This figure shows time series correlations between benchmark measures of liquidity and low-frequency proxy measures in the subset of high and low volatility
intervals for the pair BTCUSD. Values represent simple averages across the three exchanges Bitfinex, Bitstamp, and Coinbase Pro. Liquidity measures are calculated on a daily
basis over the sample period 12/16/2017 to 12/16/2019.

fraction should be 50%. Therefore, we test whether the actual frac- Second, the ability of some transactions-based measures to
tions are significantly larger than 50% using a simple binomial test. capture the cross-venue differences in liquidity depends on the
The results are presented in Table 8. data frequency. In particular, the higher the data frequency
Two general patterns emerge. First, the results for the quoted (and, correspondingly, the number of observations), the better
spread, the effective spread and the cost of a roundtrip trade are the performance of the Abdi and Ranaldo (2017) estimator. The
better than those for the price impact. Several of the transactions- Amihud (2002) illiquidity ratio, the Corwin and Schultz (2012) and
based measures (in particular the Amihud (2002) illiquidity ra- the Kyle and Obizhaeva (2016) estimators, in contrast, are more
tio, the Kyle and Obizhaeva (2016) measure, and the Abdi and consistent. They perform well at all data frequencies.29
Ranaldo (2017) and Corwin and Schultz (2012) estimators) repli- Overall, when considering all 36 comparisons (4 benchmark
cate the ranking of trading venues according to these three bench- measures, 3 trading venue pairs, 3 data frequencies) of each
mark measures well, with fractions of correct rankings ranging up
to 97.3%. For the price impact, on the other hand, the percentage of
matching rankings is lower and is often close to 50% even for the 29
Note that the Amihud (2002) illiquidity ratio does not capture the liquid-
best-performing estimators. The liquidity proxies thus do not pro- ity differences between Bitfinex and Coinbase well. However, as documented in
vide valuable information on the ranking of price impacts across Table 3 above, the liquidity differences between these two exchanges are small.
different trading venues. When liquidity differences are small, ranking venues according to their liquidity is
less important.

14
A. Brauneis, R. Mestel, R. Riordan et al. Journal of Banking and Finance 124 (2021) 106041

Fig. 9. This figure shows time series correlations between benchmark measures of liquidity and low-frequency proxy measures in the subset of high and low volume
intervals for the pair BTCUSD. Values represent simple averages across the three exchanges Bitfinex, Bitstamp, and Coinbase Pro. Liquidity measures are calculated on a daily
basis over the sample period 12/16/2017 to 12/16/2019.

transactions-based measure to the benchmark measures, the venues. All tables and figures we are referring to are in the ap-
Corwin and Schultz (2012) estimator achieves a rate of cor- pendix.
rect rankings significantly more frequently than expected by pure When considering the time-series correlations between our
chance (i.e., a rate of correct rankings significantly above 50%) in low-frequency liquidity measures and the high-frequency bench-
26 cases. The corresponding numbers for the Amihud (2002) illiq- mark measures we mostly find correlation levels that are slightly
uidity ratio, the Kyle and Obizhaeva (2016) measure and the lower for ETHUSD than for BTCUSD, particularly at the lowest data
Abdi and Ranaldo (2017) estimator are 26, 25 and 18, respectively. frequency (see Figs. 10–12). The Corwin and Schultz (2012) and
Abdi and Ranaldo (2017) estimators yield the highest correlations
for QS, ES and P I at the hourly and the daily frequency, with of-
3.7. Results for ethereum
ten almost identical correlation levels achieved by these two es-
timators. When the cost of a roundtrip trade CRT (Y )30 is used
In this section we briefly discuss our results for the currency
pair ETHUSD which are in most respects qualitatively similar to
those for BTCUSD. As for BTCUSD, results for the three trading 30
We set the dollar trading volume Y to USD 17,400 which corresponds to the
venues are very similar. We therefore report averages across the 99% quantile of the aggregate trade size distribution for the currency pair ETHUSD.

15
A. Brauneis, R. Mestel, R. Riordan et al. Journal of Banking and Finance 124 (2021) 106041

Table 7 To summarize, our findings relating to time-series correlations

This table reports average root mean squared errors (RMSE) and average mean ab-
between proxy and benchmark measures for the pair ETHUSD are
solute errors (MAE) for the three exchanges for the pair BTCUSD. The panels in
the table refer to hourly, daily and 15-daily results respectively. Lines depict proxy similar to the results for BTCUSD.
measures whilst columns in the table refer to the benchmark liquidity measures ef- In a next step we investigate the ability of the low-frequency
fective spread (ES) and price impact (PI). Note that Amihud is conceptually a proxy measures to capture the level of the high-frequency benchmarks.
for the price impact, while the other low-frequency measures Roll_r, Kyle, CS and We use the same performance metrics as in Section 3.5 and ob-
AR estimate the spread. All values are multiplied by 10 0 0. The sample period is
tain results for the pair ETHUSD that are again qualitatively sim-
12/16/2017 to 12/16/2019.
ilar to those for BTCUSD (see Table 12), with one important ex-
RMSE MAE ception. The Amihud (2002) illiquidity ratio performs poorly at the
Hourly ES PI ES PI hourly frequency because there are several one-hour intervals with
Amihud 0.487 0.292 0.322 0.050 very little volume but considerable price changes. The Kyle and
Roll_r 0.585 0.738 0.340 0.452 Obizhaeva (2016) estimator performs best overall. The illiquidity
Kyle 0.346 0.050 0.267 0.036 ratio does well at lower data frequencies, particularly when the
CS 0.224 0.296 0.161 0.182 price impact is used as benchmark. As for the currency pair BT-
AR 0.231 0.436 0.149 0.319
CUSD we find the Roll (1984), the Abdi and Ranaldo (2017) and the
daily ES PI ES PI Corwin and Schultz (2012) estimators to perform poorly at lower
Amihud 0.367 0.059 0.306 0.042 data frequencies. Again, the performance of these proxy measures
Roll_r 5.607 5.816 3.533 3.639 gets worse the longer the time frame, probably because of the
Kyle 0.308 0.057 0.242 0.049 small sample bias mentioned on page 31.
CS 2.511 2.783 1.872 2.135
Finally, we analyze the ability of the low-frequency measures to
AR 2.801 3.063 2.041 2.304
replicate the ranking produced by the benchmark measures as in
15-daily ES PI ES PI Section 3.6 above. Results are displayed in Table 13. Two measures
Amihud 0.358 0.054 0.320 0.435 stand out, the Amihud (2002) illiquidity ratio and the Kyle and
Roll_r 31.23 31.44 21.71 21.81 Obizhaeva (2016) estimator. Both measures achieve a rate of cor-
Kyle 0.284 0.074 0.244 0.067 rect rankings significantly above 50% in 25 out of 36 cases. The
CS 15.46 15.74 13.49 13.77
AR 15.36 15.64 13.23 13.51
Corwin and Schultz (2012) estimator that performed well when
applied to BTCUSD, does poorly when applied to ETHUSD.

as benchmark measure the Kyle and Obizhaeva (2016) estimator 4. Conclusion

(which performs rather poorly for the other benchmark measures)
performs best. As for the pair BTCUSD the volume proxy measures In this paper we compare the performance of transactions-
are surprisingly highly positively correlated with the benchmark based liquidity measures to benchmark measures derived from
measures, particularly with the effective spread. high-frequency order book data. We use data for the two most
The quantile dependence plots for ETHUSD (see Figure 13) show actively traded cryptocurrencies, bitcoin and ethereum, and from
that, besides the Corwin and Schultz (2012) and the Abdi and three trading venues. We consider four benchmark measures, (a)
Ranaldo (2017) estimators, also the Kyle and Obizhaeva (2016) es- the quoted and (b) the effective spread, (c) the price impact, and
timator performs well. For two of the benchmark measures, and (d) the cost of a roundtrip trade, and we consider the performance
only for low values of q, the Amihud (2002) illiquidity ratio also of the transactions-based measures across three dimensions, (i)
performs well. When we construct a composite estimator from the their ability to capture the time-series variation in liquidity, (ii)
eight proxy measures by means of a principal component analysis their ability to capture the level of liquidity, and (iii) their ability
we find the time-series correlations between the composite esti- to capture cross-exchange differences in liquidity.
mator and our benchmark measures to be almost universally lower We find that no estimator performs well across all dimensions.
than the correlations between the benchmarks and the Corwin and The Corwin and Schultz (2012) and Abdi and Ranaldo (2017) es-
Schultz (2012) and the Abdi and Ranaldo (2017) estimators, respec- timators best capture the time series variation in liquidity. This is
tively (see Table 11). true overall, at different quantiles of the distribution, in the first
We separately calculate the time-series correlations between and the second half of the sample period, and in sub-samples of
the proxy liquidity measures and the benchmark measures for the high and low return, high and low volatility, and high and low vol-
first and the second half of the sample period, for high and low re- ume periods. The measures that perform best in the cross-sectional
turn, high and low volatility, and high and low volume periods for analysis are the Amihud (2002) illiquidity ratio and the Kyle and
the daily time frame. Results are reported in Figs. 14–17. The time- Obizhaeva (2016) estimator because they do well at all data fre-
series correlations in the first and the second half of the sample quencies and for both currency pairs. When estimating the level of
periods are roughly similar, especially when considering ES, PI and the benchmark measures again the Amihud (2002) illiquidity ratio
CRT. Thus, and in contrast to the results for BTCUSD, we do not and the Kyle and Obizhaeva (2016) estimator perform best while
find markedly lower correlations in the second half of the sample the Corwin and Schultz (2012) and Abdi and Ranaldo (2017) esti-
period. mators do poorly in this respect.
Correlations tend to be higher in high volume and high volatil- Overall our results suggest that investors should use
ity periods for all our benchmark measures while high and low re- the Amihud (2002) illiquidity ratio or the Kyle and
turn intervals yield roughly similar correlations. Most importantly Obizhaeva (2016) estimator to identify the most liquid exchange.
we find that the rankings of the proxy measures remain largely un- The same recommendation holds for investors when estimating
changed in the sub-samples that result from our sample splits. The the level of execution costs in order to incorporate these into their
Corwin and Schultz (2012) and Abdi and Ranaldo (2017) estimators trading strategies. On the other hand, researchers looking for a
have the highest correlations with almost all benchmark measures measure that captures the time-series variation of liquidity, or
under almost all conditions. When the cost of a roundtrip trade is investors hoping to time the liquidity of cryptocurrency markets
used as the benchmark the Kyle and Obizhaeva (2016) estimator and enter or exit when markets are liquid are best served by the
in some cases achieves better results than the Corwin and Schultz Corwin and Schultz (2012) and Abdi and Ranaldo (2017) liquid-
(2012) and Abdi and Ranaldo (2017) estimators. ity measures because these estimators best capture time-series

16
A. Brauneis, R. Mestel, R. Riordan et al. Journal of Banking and Finance 124 (2021) 106041

Table 8
This table reports the percentage of matching orders of proxy and benchmark liquidity measures for paired
exchange-wise comparisons for the pair BTCUSD. Panels depict results for hourly, daily and 15-daily intervals
respectively. Each panel reports results for the three respective pairs and the four benchmark measures. For
instance, 96.84 in the hourly panel, for the pair Bitfinex / Bitstamp (BF / BS), for the quoted spread (QS)
and the Abdi and Ranaldo estimator (AR) implies that the Abdi and Ranaldo estimator correctly matches the
ranking of the quoted spread for this exchange pair in 96.84% of all intervals. An asterisk (∗ ) indicates that
the corresponding value is significantly higher (at the 1% level) than 50% (the value that would obtain from
guessing). The sample period is 12/16/2017 to 12/16/2019.

Hourly BF / BS TX $Vol Amihud Roll_r Roll_p Kyle CS AR

QS 5.32 24.19 91.78∗ 75.71∗ 75.65∗ 89.84∗ 92.64∗ 96.84∗

ES 4.35 19.55 76.06∗ 61.60∗ 61.56∗ 74.35∗ 78.01∗ 80.79∗
PI 32.46 37.11 50.13 43.92 43.91 49.80 53.04∗ 52.57∗
CRT 5.64 24.16 91.70∗ 75.54∗ 75.48∗ 89.84∗ 92.47∗ 96.60∗
BF / CB
QS 55.80∗ 77.34∗ 46.97 38.61 38.60 30.12 69.26∗ 72.34∗
ES 53.52∗ 60.60∗ 45.48 37.86 37.79 39.60 67.57∗ 69.63∗
PI 38.60 50.71 57.81∗ 35.49 35.52 47.69 48.80 54.04∗
CRT 49.39 44.65 63.25∗ 37.10 37.13 59.65∗ 57.16∗ 58.93∗
BS / CB
QS 4.03 32.90 97.32∗ 76.70∗ 76.70∗ 86.83∗ 87.86∗ 96.28∗
ES 4.09 29.24 84.31∗ 65.92∗ 65.92∗ 74.77∗ 78.46∗ 84.17∗
PI 24.96 37.60 63.18∗ 53.23∗ 53.23∗ 59.14∗ 58.98∗ 62.81∗
CRT 4.07 32.92 97.28∗ 76.67∗ 76.67∗ 86.86∗ 87.87∗ 96.27∗
daily BF / BS
QS 0.30 20.36 81.00∗ 37.86 37.56 79.03∗ 75.72∗ 52.34
ES 0.45 20.06 78.43∗ 36.50 36.20 76.47∗ 72.85∗ 50.98
PI 21.42 35.29 61.09∗ 34.69 34.69 60.03∗ 65.16∗ 47.51
CRT 0.30 20.36 81.00∗ 37.86 37.56 79.03∗ 75.72∗ 52.34
BF / CB
QS 56.22∗ 85.16∗ 21.89 42.43 42.88 17.99 75.11∗ 61.62∗
ES 59.07∗ 74.96∗ 30.73 41.98 41.98 26.24 68.97∗ 58.92∗
PI 30.43 54.57 45.58 39.88 40.03 45.88 57.87∗ 54.42
CRT 44.23 44.38 59.07∗ 42.28 42.43 57.87∗ 53.52 52.62
BS / CB
QS 0.30 17.07 84.59∗ 42.90 42.45 83.08∗ 91.39∗ 61.18∗
ES 0.45 16.92 82.33∗ 41.24 40.79 80.21∗ 88.82∗ 59.82∗
PI 16.16 24.47 73.11∗ 40.33 39.88 72.51∗ 75.98∗ 56.95∗
CRT 0.30 17.07 84.59∗ 42.90 42.45 83.08∗ 91.39∗ 61.18∗
15-daily BF / BS
QS 0.00 19.57 82.61∗ 32.61 28.26 78.26∗ 58.70 58.70
ES 0.00 19.57 82.61∗ 32.61 28.26 78.26∗ 58.70 58.70
PI 13.04 28.26 69.57∗ 32.61 32.61 69.57∗ 58.70 58.70
CRT 0.00 19.57 82.61∗ 32.61 28.26 78.26∗ 58.70 58.70
BF / CB
QS 65.22 93.48∗ 8.70 39.13 43.48 15.22 78.26∗ 60.87
ES 69.57∗ 84.78∗ 17.39 45.65 50.00 23.91 73.91∗ 60.87
PI 17.39 50.00 52.17 39.13 36.96 50.00 56.52 43.48
CRT 41.30 43.48 54.35 41.30 41.30 56.52 50.00 41.30
BS / CB
QS 0.00 13.04 86.96∗ 30.43 26.09 89.13∗ 84.78∗ 47.83
ES 0.00 13.04 86.96∗ 30.43 26.09 89.13∗ 84.78∗ 47.83
PI 10.87 15.22 84.78∗ 23.91 19.57 86.96∗ 82.61∗ 50.00
CRT 0.00 13.04 86.96∗ 30.43 26.09 89.13∗ 84.78∗ 47.83

variation in liquidity. These differing findings suggest that the the view & editing, Visualization. Roland Mestel: Methodology, Formal
setting is important in determining the best liquidity proxy. analysis, Writing - original draft, Writing - review & editing, Visu-
Our results can be used by researchers, investors, traders, and alization, Project administration. Ryan Riordan: Conceptualization,
regulators to understand liquidity levels and dynamics with rela- Methodology, Formal analysis, Writing - original draft, Writing -
tively easy to acquire and process aggregate price and volume data. review & editing. Erik Theissen: Conceptualization, Methodology,
In many applications, the transactions-based aggregate measures Formal analysis, Writing - original draft, Writing - review & edit-
perform adequately when describing high-frequency measures de- ing.
rived from order book data. The use of these low-frequency mea-
sures is far less time-consuming and memory-intensive, offering
a reasonable compromise between accuracy and computational Acknowledgements
workload. Strategies that require more granular data such as tri-
angular arbitrage or market-making will of course require higher We thank Yakov Amihud and Hans Manner for valuable com-
frequency measures. ments. We gratefully acknowledge financial support from the Uni-
versity of Graz.

CRediT authorship contribution statement

Appendix
Alexander Brauneis: Methodology, Software, Formal analysis,
Investigation, Data curation, Writing - original draft, Writing - re- This appendix reports the results for the currency pair ETHUSD.

17
A. Brauneis, R. Mestel, R. Riordan et al. Journal of Banking and Finance 124 (2021) 106041

Table 9
Descriptive data for benchmark liquidity measures for the pair ETHUSD used in the empirical analysis at a daily res-
olution. The table reports descriptive statistics for the quoted spread (QS), the effective spread (ES), the price impact
(PI) and the percentage cost of a roundtrip trade (CRT). The unit of measurement is basis points. The sample period is
12/16/2017 to 12/16/2019.

exchange mean std. dev. Q1 median Q3 num daily obs.

Bitfinex QS 1.367 1.433 0.646 1.020 1.420 699

ES 1.929 1.622 0.901 1.538 2.320
PI 0.578 0.440 0.274 0.461 0.736
CRT 7.224 3.809 3.854 7.634 9.885
Bitstamp QS 13.17 4.681 10.11 12.04 15.31 651
ES 13.45 4.806 10.32 12.39 15.43
PI 0.770 0.584 0.411 0.624 0.929
CRT 30.33 9.934 23.17 29.15 35.24
Coinbase QS 1.232 0.928 0.647 0.979 1.475 676
ES 2.177 2.294 1.157 1.790 2.580
PI 0.601 0.587 0.227 0.458 0.769
CRT 7.447 3.125 5.351 6.916 8.988

Table 10
Descriptive data for proxy liquidity measures for the pair ETHUSD used in the empirical analysis at a daily resolution.
The table reports descriptive statistics for the number of transactions (TX), the dollar volume ($ Vol, million USD), the
Amihud measure (Amihud, values ∗ 1e6), Roll’s returns based measure (Roll_r, basispoints), Roll’s price based measure
(Roll_p), the Kyle and Obizhaeva measure (Kyle, values ∗ 1e3), the Corwin and Schultz measure (CS, basispoints) and the
Abdi and Ranaldo measure (AR, basispoints). The sample period is 12/16/2017 to 12/16/2019.

exchange mean std. dev. Q1 median Q3 no. daily obs

Bitfinex TX 33,985 25,895 13,948 27,570 47,358 699

$ Vol 59.24 72.87 13.88 33.20 77.84
Amihud 0.044 0.436 0.003 0.006 0.016
Roll_r 48.51 56.57 0 37.30 70.78
Roll_p 2.006 3.882 0 0.725 1.906
Kyle 0.131 0.041 0.101 0.124 0.156
CS 26.82 21.74 13.16 20.58 34.00
AR 30.32 24.20 15.60 23.83 36.51
Bitstamp TX 7,889 8,127 2,947 5,451 10,075 651
$ Vol 12.47 17.26 3.633 7.136 14.21
Amihud 0.286 1.157 0.019 0.041 0.110
Roll_r 49.01 54.73 0 36.74 69.82
Roll_p 2.089 3.957 0 0.747 2.046
Kyle 0.212 0.060 0.171 0.203 0.242
CS 28.46 22.14 14.63 22.35 35.32
AR 31.13 23.34 16.13 24.50 38.08
Coinbase Pro TX 35,894 31,307 14,997 26,626 46,208 676
$ Vol 41.60 63.87 10.11 21.31 45.18
Amihud 0.568 10.38 0.005 0.009 0.017
Roll_r 46.42 54.08 0 33.90 67.08
Roll_p 1.919 3.715 0 0.659 1.826
Kyle 0.143 0.040 0.116 0.137 0.163
CS 26.10 20.54 13.13 20.65 32.34
AR 29.80 23.62 15.07 23.43 36.01

Table 11
This table reports mean correlations among the first principal component of proxy liquidity measures and the four
benchmark measures for the pair ETHUSD. Column ’expl. var’ shows the percentage of total variance explained by the
first principal component. Rows denoted ’mean’ contain equally-weighted averages across exchanges. The sample period
is 12/16/2017 to 12/16/2019.

expl. var QS ES PI CRT

hourly Bitfinex 50.98 0.696 0.660 0.752 0.414

Bitstamp 50.23 0.636 0.687 0.692 0.548
Coinbase Pro 51.80 0.533 0.438 0.836 0.359
mean 51.00 0.621 0.595 0.760 0.440
daily Bitfinex 52.35 0.679 0.626 0.820 0.231
Bitstamp 54.05 0.660 0.719 0.823 0.536
Coinbase Pro 52.24 0.345 0.457 0.838 0.177
mean 52.88 0.561 0.600 0.827 0.315
15-daily Bitfinex 56.29 0.475 0.295 0.725 −0.191
Bitstamp 50.21 0.715 0.740 0.783 0.493
Coinbase Pro 54.80 −0.032 0.137 0.851 −0.293
mean 53.76 0.386 0.391 0.786 0.003

18
A. Brauneis, R. Mestel, R. Riordan et al. Journal of Banking and Finance 124 (2021) 106041

Table 12
This table reports average root mean squared errors (RMSE) and average mean absolute errors (MAE) for the
three exchanges for the pair ETHUSD. The panels in the table refer to hourly, daily and 15-daily results re-
spectively. Lines depict proxy measures whilst columns in the table refer to the benchmark liquidity measures
effective spread (ES) and price impact (PI). Note that Amihud is conceptually a proxy for the price impact,
while the other low-frequency measures Roll_r, Kyle, CS and AR estimate the spread. All values are multiplied
by 1,0 0 0. The sample period is 12/16/2017 to 12/16/2019.

RMSE MAE

Hourly ES PI ES PI

Amihud 63.14 63.05 2.013 1.447

Roll_r 0.855 1.103 0.544 0.715
Kyle 0.693 0.085 0.549 0.062
CS 0.578 0.339 0.460 0.173
AR 0.485 0.584 0.347 0.424
daily ES PI ES PI

Amihud 0.655 0.085 0.575 0.065

Roll_r 6.873 7.246 4.556 4.774
Kyle 0.532 0.117 0.431 0.103
CS 2.914 3.380 2.159 2.647
AR 3.311 3.778 2.479 2.977
15-daily ES PI ES PI

Amihud 0.641 0,077 0.599 0.068

Roll_r 41.49 41.93 32.69 33.00
Kyle 0.482 0.147 0.424 0.134
CS 18.84 19.35 17.14 17.67
AR 19.08 19.58 17.06 17.60

Table 13
This table reports the percentage of matching orders of proxy and benchmark liquidity measures for paired exchange-wise comparisons for the
pair ETHUSD. Panels depict results for hourly, daily and 15-daily intervals respectively. Each panel reports results for the three respective pairs for
Bitfinex (BF), Bitstamp (BS) and Coinbase Pro (CB). For instance, 84.62 in the hourly panel, for the pair Bitfinex / Bitstamp (BF / BS), for the quoted
spread (QS) and the Abdi and Ranaldo estimator (AR) implies that the Abdi and Ranaldo estimator matches the order of the contemporaneous
order of the quoted spread for this exchange pair in 84.62% of all intervals. An asterisk (∗ ) indicates that the corresponding value is significantly
higher (at the 1% level) than 50% (the value that would obtain from guessing). The sample period is 12/16/2017 to 12/16/2019.

Hourly BF / BS TX $Vol Amihud Roll_r Roll_p Kyle CS AR

QS 0.34 4.99 93.75∗ 68.41∗ 68.21∗ 97.02∗ 57.29∗ 84.62∗

ES 0.16 1.35 38.91 26.64 26.54 39.76 27.28 36.55
PI 11.65 12.54 28.02 20.99 21.03 28.34 25.98 27.95
CRT 0.37 4.97 93.73∗ 68.39∗ 68.18∗ 97.05∗ 57.27∗ 84.60∗
BF / CB
QS 51.52∗ 50.98 54.76∗ 40.24 40.26 54.25∗ 68.28∗ 68.81∗
ES 34.28 28.49 44.56 29.06 29.01 45.87 52.98∗ 53.45∗
PI 38.25 35.73 39.61 28.23 28.16 39.03 41.98 42.90
CRT 46.87 31.08 69.00∗ 41.22 41.23 71.84∗ 65.22∗ 66.29∗
BS / CB
QS 0.55 6.30 97.27∗ 66.39∗ 66.44∗ 96.98∗ 47.54 76.10∗
ES 0.36 1.91 40.90 26.53 26.53 40.70 24.17 34.77
PI 16.05 16.87 25.28 20.37 20.37 25.22 23.01 24.19
CRT 0.55 6.30 97.27∗ 66.39∗ 66.44∗ 96.98∗ 47.54 76.10∗
daily BF / BS
QS 0.00 0.31 97.85∗ 37.23 36.92 99.69∗ 54.00 53.08
ES 0.00 0.31 87.54∗ 31.69 31.54 88.77∗ 49.69 47.08
PI 20.46 20.77 67.69∗ 32.77 33.38 68.31∗ 46.77 47.85
CRT 0.00 0.31 97.85∗ 37.23 36.92 99.69∗ 54.00 53.08
BF / CB
QS 47.02 48.81 52.08 37.50 36.90 54.17 56.70∗ 51.49
ES 50.30 41.96 55.80∗ 34.97 35.12 56.55∗ 57.29∗ 49.11
PI 48.36 48.66 47.47 39.73 40.18 49.11 51.49 49.70
CRT 38.54 27.53 73.96∗ 35.86 36.76 75.74∗ 50.00 52.23
BS / CB
QS 0.00 0.31 97.69∗ 44.31 44.77 99.69∗ 58.92∗ 54.62∗
ES 0.15 0.31 86.15∗ 38.92 39.38 88.15∗ 54.15∗ 48.46
PI 19.85 20.00 66.92∗ 37.85 38.00 68.31∗ 48.46 51.38
CRT 0.00 0.31 97.69∗ 44.31 44.77 99.69∗ 58.92∗ 54.62∗
15-daily BF / BS
QS 0.00 0.00 100.00∗ 39.13 39.13 100.00∗ 45.65 45.65
ES 0.00 0.00 100.00∗ 39.13 39.13 100.00∗ 45.65 45.65
PI 17.39 17.39 82.61∗ 30.43 30.43 82.61∗ 50.00 32.61
CRT 0.00 0.00 100.00∗ 39.13 39.13 100.00∗ 45.65 45.65
BF / CB
QS 43.48 50.00 50.00 43.48 32.61 54.35 43.48 50.00
ES 52.17 45.65 54.35 43.48 36.96 58.70 33.78 54.35
PI 58.70 43.48 56.52 52.17 52.17 52.17 45.65 39.13
(continued on next page)

19
A. Brauneis, R. Mestel, R. Riordan et al. Journal of Banking and Finance 124 (2021) 106041

Table 13 (continued)

Hourly BF / BS TX $Vol Amihud Roll_r Roll_p Kyle CS AR

∗ ∗
CRT 34.78 28.26 71.74 43.48 39.13 76.09 39.13 45.65
BS / CB
QS 0.00 0.00 100.00∗ 30.43 34.78 100.00∗ 69.57∗ 56.52
ES 0.00 0.00 100∗ 30.43 34.78 100.00∗ 69.57∗ 56.52
PI 21.74 21.74 78.26∗ 32.16 36.96 78.26∗ 52.17 43.48
CRT 0.00 0.00 100.00∗ 30.43 34.78 100.00∗ 69.57∗ 56.52

Fig. 10. This figure shows time series correlations between benchmark measures of liquidity and low-frequency proxy measures for the pair ETHUSD. Values represent
simple averages across the three exchanges Bitfinex, Bitstamp, and Coinbase Pro. Liquidity measures are calculated on an hourly basis over the sample period 12/16/2017 to
12/16/2019.

Fig. 11. This figure shows time series correlations between benchmark measures of liquidity and low-frequency proxy measures for the pair ETHUSD. Values represent simple
averages across the three exchanges Bitfinex, Bitstamp, and Coinbase Pro. Liquidity measures are calculated on a daily basis over the sample period 12/16/2017 to 12/16/2019.

20
A. Brauneis, R. Mestel, R. Riordan et al. Journal of Banking and Finance 124 (2021) 106041

Fig. 12. This figure shows time series correlations between benchmark measures of liquidity and low-frequency proxy measures for the pair ETHUSD. Values represent
simple averages across the three exchanges Bitfinex, Bitstamp, and Coinbase Pro. Liquidity measures are calculated on a 15-daily basis over the sample period 12/16/2017 to
12/16/2019.

Fig. 13. This figure plots the quantile dependence (averaged across the three trading venues) as a function of the quantile q in steps of 0.01 for the four benchmark liquidity
measures and each of the 8 proxy measures for the pair ETHUSD for daily data. The sample period is 12/16/2017 to 12/16/2019.

21
A. Brauneis, R. Mestel, R. Riordan et al. Journal of Banking and Finance 124 (2021) 106041

Fig. 14. This figure shows time series correlations between benchmark measures of liquidity and low-frequency proxy measures for the first and second half of the sample
for the pair ETHUSD. Values represent simple averages across the three exchanges Bitfinex, Bitstamp, and Coinbase Pro. Liquidity measures are calculated on a daily basis
over the sample period 12/16/2017 to 12/16/2019.

22
A. Brauneis, R. Mestel, R. Riordan et al. Journal of Banking and Finance 124 (2021) 106041

Fig. 15. This figure shows time series correlations between benchmark measures of liquidity and low-frequency proxy measures in the subset of high and low return intervals
for the pair ETHUSD. Values represent simple averages across the three exchanges Bitfinex, Bitstamp, and Coinbase Pro. Liquidity measures are calculated on a daily basis
over the sample period 12/16/2017 to 12/16/2019.

23
A. Brauneis, R. Mestel, R. Riordan et al. Journal of Banking and Finance 124 (2021) 106041

Fig. 16. This figure shows time series correlations between benchmark measures of liquidity and low-frequency proxy measures in the subset of high and low volatility
intervals for the pair ETHUSD. Values represent simple averages across the three exchanges Bitfinex, Bitstamp, and Coinbase Pro. Liquidity measures are calculated on a daily
basis over the sample period 12/16/2017 to 12/16/2019.

24
A. Brauneis, R. Mestel, R. Riordan et al. Journal of Banking and Finance 124 (2021) 106041

Fig. 17. This figure shows time series correlations between benchmark measures of liquidity and low-frequency proxy measures in the subset of high and low volume
intervals for the pair ETHUSD. Values represent simple averages across the three exchanges Bitfinex, Bitstamp, and Coinbase Pro. Liquidity measures are calculated on a daily
basis over the sample period 12/16/2017 to 12/16/2019.

References Corwin, S.A., Schultz, P., 2012. A simple way to estimate bid-ask spreads from daily
high and low prices. J. Finance 67 (2), 719–760.
Abdi, F., Ranaldo, A., 2017. A simple estimation of bid-ask spreads from daily close, Dimpfl, T., 2017. Bitcoin Market Microstructure. Available at SSRN: https://ssrn.com/
high and low prices. Rev. Financ. Stud. 30 (12), 4437–4480. abstract=2949807.
Alexander, C., Dakos, M., 2019. A critical investigation of cryptocurrency data and Duan, F., Manner, H., Wied, D., 2019. Model and moment selection in factor copula
analysis. Quant. Finance 20 (2), 173–188. models. J. Financ. Econom. 28, 1–31.
Amihud, Y., 2002. Illiquidity and stock returns: cross-section and time-series effects. Dyhrberg, A., Foley, S., Svec, J., 2018. How investible is bitcoin? Analyzing the liq-
J. Financ. Mark. 5 (1), 31–56. uidity and transaction costs of bitcoin markets. Econ. Lett. 171, 140–143.
Amihud, Y, Noh, J, 2002. Illiquidity and stock returns II: cross– Fink, C., Johann, T., 2014. Bitcoin Markets. Available at SSRN: https://ssrn.com/
section and time-series effects. Review of Financial Studies abstract=2408396.
https://doi.org/10.1093/rfs/hhaa080. In press. Fong, K.Y., Holden, C.W., Trzcinka, C.A., 2017. What are the best liquidity proxies for
Bogousslavsky, V., Collin-Dufresne, P., 2020. Liquidity, volume, and volatility. Avail- global research? Rev. Finance 21 (4), 1355–1401.
able at SSRN: https://ssrn.com/abstract=3336171. Goyenko, R.Y., Holden, C.W., Trzcinka, C.A., 2009. Do liquidity measures measure
Brauneis, A., Mestel, R., 2018. Price discovery of cryptocurrencies: bitcoin and be- liquidity? J. Financ. Econ. 92 (2), 153–181.
yond. Econ. Lett. 165, 58–61. Harris, L., 1990. Statistical properties of the roll serial covariance bid/ask spread es-
Brauneis, A., Mestel, R., Riordan, R., Theissen, E., 2021. Bitcoin unchained: Determi- timator. J. Finance 45 (2), 579–590.
nants of cryptocurrency exchange liquidity. Available at SSRN: https://ssrn.com/ Hasbrouck, J., Seppi, D.J., 2001. Common factors in prices, order flows, and liquidity.
abstract=3249477. J. Financ. Econ. 59 (3), 383–411.
Conrad, J., Wahal, S., 2020. The term structure of liquidity provision. J. Financ. Econ.
136, 239–259.

25
A. Brauneis, R. Mestel, R. Riordan et al. Journal of Banking and Finance 124 (2021) 106041

Hautsch, N., Scheuch, C., Voigt, S., 2018. Building Trust Takes Time: Limits to Arbi- Liu, Y., Tsyvinski, A., Wu, X., 2019. Common risk factors in cryptocurrency. NBER
trage in Blockchain-Based Markets. Available at SSRN: https://ssrn.com/abstract= Working Paper 25882. Available at: https://www.nber.org/papers/w25882
3302159. Makarov, I., Schoar, A., 2020. Trading and arbitrage in cryptocurrency markets. J.
Hougan, M., Kim, H., Lerner, M., 2019. Economic and non-economic trad- Financ. Econ. 135 (2), 293–319.
ing in bitcoin: exploring the real spot market for the world’s first digi- Mancini, L., Ranaldo, A., Wrampelmeyer, J., 2013. Liquidity in the foreign exchange
tal commodity. Working Paper. Available at: https://www.sec.gov/comments/ market: measurement, commonality, and risk premiums. J. Finance LXVIII,
sr-nysearca-2019-01/srnysearca201901-5574233-185408.pdf 1805–1841.
Hu, A.S., Parlour, C.A., Rajan, U., 2019. Cryptocurrencies: stylized facts on a new in- Marshall, B., Nguyen, N., Visaltanachoti, N., 2019. Bitcoin Liquidity. Available at
vestible instrument. Financ. Manag. 48 (4), 1049–1068. SSRN: https://ssrn.com/abstract=3194869.
Irvine, N., Benston, G., Kandel, E., 20 0 0. Liquidity Beyond the Inside Spread: Mea- Marshall, B.R., Nguyen, N.H., Visaltanachoti, N., 2012. Commodity liquidity measure-
suring and Using Information in the Limit Order Book. Available at SSRN: https: ment and transaction costs. Rev. Financ. Stud. 25 (2), 599–638.
//ssrn.com/abstract=229959. Roll, R., 1984. A simple implicit measure of the effective bid-ask spread in an effi-
Johann, T., Theissen, E., 2020. The Best in Town: A Comparative Analysis of Low- cient market. J. Finance 39 (4), 1127–1139.
Frequency Liquidity Estimators. Available at SSRN: https://ssrn.com/abstract= Schestag, R., Schuster, P., Uhrig-Homburg, M., 2016. Measuring liquidity in bond
2905032. markets. Rev. Financ. Stud. 29 (5), 1170–1219.
Karnaukh, N., Ranaldo, A., Söderlind, P., 2015. Understanding FX liquidity. Rev. Fi- Shi, S., 2018. The Impact of Futures Trading on Intraday Spot Volatility and Liquid-
nanc. Stud. 28, 3073–3108. ity: Evidence from Bitcoin Market. Available at SSRN: https://ssrn.com/abstract=
Kyle, A.S., Obizhaeva, A.S., 2016. Market microstructure invariance: empirical hy- 3094647.
potheses. Econometrica 84, 1345–1404.

26