Constructing Cointegrated Cryptocurrency Portfolios for

Statistical Arbitrage
Tim Leung∗ Hung Nguyen†

Abstract
In this paper, we analyze the process of constructing cointegrated portfolios of cryp-
tocurrencies. Our procedure involves a series of statistical tests, including the Johansen
cointegration test and Engle-Granger two-step approach. Among our results, we construct
cointegrated portfolios involving four cryptocurrencies: Bitcoin (BTC), Ethereum (ETH),
Bitcoin Cash (BCH), and Litecoin (LTC). We develop a number of trading strategies under
different entry/exit thresholds and risk constraints, and examine their performance in de-
tails through backtesting and comparison analysis. Our methodology can be applied more
generally to create new cointegrated portfolio using other cryptocurrencies.

Contents
1 Introduction 2

2 Data 4

3 Mean-Reverting Portfolio 5
3.1 Johansen Test . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
3.2 Engle-Granger Two-Step Method . . . . . . . . . . . . . . . . . . . . . . . . . 7
3.3 Portfolio Time Series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

4 Statistical Arbitrage Strategies 11

4.1 Trading the spread . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
4.2 Trading the spread with stop-loss . . . . . . . . . . . . . . . . . . . . . . . . . 13
4.3 Trading the spread with trailing stop-loss . . . . . . . . . . . . . . . . . . . . 15

5 Conclusion 18

A Appendix: statistics for trading strategies 20

∗
Applied Mathematics Department, University of Washington, Seattle WA 98195. Email: timleung@uw.edu.
Corresponding author.
†
Computational Finance & Risk Management Program, Applied Mathematics Department, University of Wash-
ington, Seattle WA 98195. Email: hn11@uw.edu.

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=3235890
https://ssrn.com/abstract=3235890
 1 Introduction
In 2009, Bitcoin (BTC) was introduced to become the first decentralized digital currency.1
Since then, Bitcoin has been accepted by thousands of merchants and vendors, and the
cryptocurrency market has seen enormous growth (Hileman and Rauchs, 2017). Its market
capitalization measured in USD has grown from about $10bn in 2014 to over $200bn in 2018,
with a peak around $828bn in Jan 2018.2 Today, there are more than a thousand of cryp-
tocurrencies besides Bitcoin, such as Ethereum (ETH), Litecoin (LTC), Bitcoin Cash (BCH),
and more that are being traded on major cryptocurrency exchanges internationally.3 Many
alternative coins, also known as altcoins, were invented to be substitutes of Bitcoin for various
reasons. For instance, Ethereum, a decentralized platform that run smart contracts,4 was
created to address the lack of a scripting language of Bitcoin for application development.5
As another example, Bitcoin Cash was invented to address the scalability issue of Bitcoin
due to its block size limit which put a cap on the amount of transactions Bitcoin network can
process on a given time frame, etc. Such innovations have led to the rapid growths of many
altcoins together with the rise in market size of cryptocurrency. However, being known as
the earliest decentralized currency, Bitcoin have consistently been ranked on top in terms of
traded volume, price and market capitalizations (see Table 1).
The rise of Bitcoin and many altcoins has created trading opportunities for investors and
speculators. In fact, a recent study by Chuen et al. (2017) finds very low correlations between
the CRyptocurrency Index (CRIX), a portfolio of cryptocurrencies, and traditional assets.
Also, their empirical results show that including CRIX in an initial portfolio consisting only
traditional assets can expand the efficient frontier and give additional utility to investors.
Recall that S&P 500 returned around 21% in 2017, but Bitcoin gained an astounding 1,318%
in one year. In an earlier study, Cheah and Fry (2015) consider the Bitcoin phenomenon as a
speculative bubble with zero fundamental value. In mid December 2017, Bitcoin price reached
all-time-high at over $19 thousand, with 24h volume approximately $23 billion. However, as
price plummeted to a low $6000 in early February, Bitcoin has been considered one of the
greatest bubbles in finance history.
The price surges of Bitcoin and Ethereum in early 2017 generated much attention to the
cryptocurrency market and even triggered a massive series of Initial Coin Offerings (ICOs),
which are similar to Initial Public Offerings (IPOs), but are for cryptocurrencies or tokens
and less regulated. Many technology startups launched their own cryptocurrencies as a means
of crowdfunding, though some ICOs were completely frauds or had been heavily targeted by
hackers (Labbe, 2017). To purchase tokens, investors must have one of the mainstream coins
which typically were Bitcoin and Ethereum. Such events created an increased demand that
pushed the prices of these two coins even higher. One of the most popular ways for investors
to gain access to the cryptocurrency market was through Coinbase6 where investors could
purchase coins directly from their bank accounts and then transfer the coins to different
digital wallets and exchanges for trading purposes. Some exchanges even allow short selling
of cryptocurrencies.
Unlike traditional equity market, most cryptocurrencies are highly correlated. For exam-
1
See the bitcoin white paper, Bitcoin: A Peer-to-Peer Electronic Cash System, by Satoshi Nakamoto. Available
at https://bitcoin.org/bitcoin.pdf
2
Source: https://coinmarketcap.com/charts/
3
Source: https://coinmarketcap.com/rankings/exchanges/
4
Smart contracts are applications that allows exchanges of money, property, or anything of value to take place
without the middleman once certain conditions are met.
5
See the white paper, A Next-Generation Smart Contract and Decentralized Application Platform, by Vita-
lik Buterin, the founder of Ethereum. Available at https://github.com/ethereum/wiki/wiki/White-Paper#
decentralized-autonomous-organizations
6
Founded in 2012, Coinbase was one of the earliest online platform in the U.S. that allowed users to store,
trade, and transfer digital currencies.

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=3235890
https://ssrn.com/abstract=3235890
 Rank Name Symbol Price ($) MarketCap ($)
1 Bitcoin BTC 8239.57 141534282604
2 Ethereum ETH 467.785 47238602070
3 XRP XRP 0.454269 17859896217
4 Bitcoin Cash BCH 828.355 14299695631
5 EOS EOS 8.3154 7451841487
6 Stellar XLM 0.314209 5896895909
7 Litecoin LTC 84.5161 4869209762
8 Cardano ADA 0.163161 4230286756
9 IOTA MIOTA 1.02566 2850853030
10 Tether USDT 0.998571 2503557642
11 TRON TRX 0.036809 2420122242
12 Monero XMR 140.502 2285241135
13 NEO NEO 34.0835 2215427500
14 Dash DASH 241.932 1988855554
15 Ethereum Classic ETC 16.9501 1752226368
16 NEM XEM 0.177986 1601874000
17 VeChain VEN 2.6677 1479361013
18 Binance Coin BNB 14.2613 1362132746
19 Tezos XTZ 2.07289 1259257958
20 Zcash ZEC 220.618 983003486
Table 1: Summary of top 20 cryptocurrencies ranked by market capitalization
as of August 16, 2018. Source: CoinMarketCap.

ple, correlations on daily returns between the 4 major digital currencies, BTC, ETH, LTC,
and BCH, were over 75% as shown in Figure 1. The high volatility of cryptocurrencies calls
for the development of market-neutral trading strategies, and the high correlations among
cryptocurrencies motivates us to investigate cointegration or mean-reversion strategies.
Mean-reversion trading strategies have been extensively researched and studied by aca-
demic researchers and practitioners who seek to understand the long-term co-movements of
different asset prices and arbitrage from the mean-reverting property of the price spread.
Strategies have been developed for many areas of traditional finance, including fixed income
Wagner (2005), commodities and futures (Leung and Ward, 2015), as well as equities and
ETFs (Kang and Leung, 2017; Leung and Li, 2015). However, very little research has been
done for trading pairs or baskets of cryptocurrencies. In fact, most research on trading cryp-
tocurrencies are based on technical analysis (Ha and Moon, 2018; Detzel et al., 2018), or
arbitraging from price differences across exchanges (Makarov and Schoar, 2018).
In this paper, we analyze the process of constructing a cointegrated set of cryptocurren-
cies. The process involves a series of statistical tests, including the Johansen cointegration
test (Johansen, 1988), and the classical 2-step approach developed by Engle and Granger
(1987). Among our results, we construct cointegrated portfolios involving four cryptocur-
rencies: Bitcoin (BTC), Ethereum (ETH), Bitcoin Cash (BCH), and Litecoin (LTC), and
the corresponding trading strategies are shown to be profitable under different entry/exit
thresholds and risk constraints. To our best knowledge, the proposed cointegrated portfolio
of 4 cryptocurrencies is new, and our methodology can be applied more generally to create
new cointegrated portfolio using other cryptocurrencies.

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=3235890
https://ssrn.com/abstract=3235890
 1

0.8
BTC
0.6

0.4
0.81 ETH 0.2

−0.2
0.83 0.8 LTC
−0.4

−0.6
0.79 0.75 0.75 BCH −0.8

−1

Figure 1: Correlation in daily returns of BTC, ETH, LTC, and BCH

2 Data
Before designing the trading systems, we first make general observations on how prices (in
USD) of Bitcoin (BTC), Ethereum (ETH), Bitcoin Cash (BCH), and Litecoin (LTC) have
co-moved since BCH was first introduced. To identify a cointegrating relationship among
these crypto-assets, we employ both Johansen and Engle-Granger approaches to construct
a mean-reverting portfolio to be tested for stationarity and tradability. Lastly, we design
and back test trading systems for the mean-reverting portfolio under five different entry/exit
threshold levels and incorporate a stop-loss exit and trailing stop exit.
A majority of cryptocurrency exchanges provides traders with user-friendly trading plat-
forms with the intention to attract more users by reducing the complexity of buying and
selling coins. For instance, Coinbase, in addition to the web application, offers mobile app
that allows traders to store coins and make trades directly from their phones without hav-
ing to know much about different types of order (i.e. limit, market order, etc.). For larger
exchanges, such as Binance, Bitfinex and Bittrex, historical and live data are also accessible
through RESTful APIs that allow traders to send price requests for analysis and manage their
trading accounts. In this paper, we use the Open High Low Close Volume (OHLCV) data
accessed through an API provided by CryptoCompare, a database in UK that provides raw
and preprocessed cryptocurrency market data from multiple exchanges around the world.
Historical data aggregated from all listed exchanges are also available in their inventory.
However, due to price differences across exchanges, daily prices in U.S. dollars of Bitcoin
(BTC), Ethereum (ETH), Litecoin (LTC) and Bitcoin Cash (BCH) were gathered only from
Coinbase from December 20, 20177 to June 20, 2018 to construct our trading models.
7
Bitcoin Cash only became available on Coinbase on December 20, 2017

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=3235890
https://ssrn.com/abstract=3235890
 10

logPrices 8
BTC
7 ETH
LTC
BCH
6

01/18 02/18 03/18 04/18 05/18 06/18

Date
Figure 2: Log-prices of 4 main cryptocurrencies in Coinbase

As shown in Figure 2, price movements of all 4 cryptocurrencies during the testing period
are nearly identical. This could imply a potential linear relationship between these 4 coins.
Furthermore, if the time series of all 4 cryptocurrencies are stationary after 1-differencing,
i.e. their prices are I(1) processes, and there exists a linear combination whose residuals
are stationary (or mean reverting), then these crypto-assets are said to be cointegrated.
The linear combination (or number of shares to purchase for each crypto-asset8 ) that forms
a stationary portfolio from non-stationary asset prices is often referred to as cointegrating
vector. From a trading perspective, an investor can profit from purchasing this stationary
portfolio, or so called the spread, when price is low and realize a profit when its price return
to the mean price or cross above certain threshold. Similarly, an investor can short sell the
spread when its price is high and realize a profit when price reverts to the mean price or
cross below certain level. This characteristic of mean-reverting in price is extensively used
by hedge fund managers and proprietary trading firms to construct pair-trading strategies,
or so-called statistical arbitrage. One can think of the spread as a mutual fund or a basket of
crypto-assets where the number of shares on each asset (can be positive or negative) would
be predetermined by the cointegrating vector. And to purchase or sell the spread, an investor
must be able to long and short on certain assets by the amount of shares suggested their
cointegrating relationships.

3 Mean-Reverting Portfolio
Cointegration reflects how asset prices are tied together in long-term by a common stochas-
tic trend even though they might drift apart in the short run. Therefore, cointegrated asset
prices may appear to be more useful in developing quantitative trading models. Theoreti-
cal framework for constructing cointegrating time series in this paper include the Johansen
cointegration test (Johansen, 1988), and the classical 2-step approach developed by Engle
and Granger (1987). Note that, although the two approaches are somewhat equivalent in
terms of establishing cointegrating relationships between given crypto-assets, the Johansen
test tends to be more appropriate when it comes to modeling multivariate time series data.
However, there are good reasons why the Engle-Granger approach is chosen for most finan-
cial applications, in particular, risk management. While the Johansen criterion seeks for
8
Often referred to as weighted sum vector; however, since we perform regression directly on price series, this
linear combination implies the number of shares to go long/short on each crypto-asset

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=3235890
https://ssrn.com/abstract=3235890
the linear combination that is most stationary, Engle-Granger approach, based on ordinary
least squares regression (OLS), seeks for the linear combination with minimum variance. As
a result, Engle-Granger approach is generally a preferred method for most risk managing
purposes, especially when involving risk measures. Also, one might be able to find more than
one linear combination to form a mean-reverting portfolio; but to determine the appropriate
model, we will need to examine both approaches and choose the cointegrating vector that
would give the spread a faster rate of mean-reversion. In other words, we construct the spread
in such way that yields highest profit when trading the mean-reverting crypto-portfolio.

3.1 Johansen Test

The main advantage of the Johansen test over the Engle-Granger approach is the ability to
perform hypothesis testings directly on I(1) price series themselves and derive cointegrating
vectors without having to constantly check for stationarity. The order of cointegration,
denoted rank(r), from Johansen criterion indicates the number of independent portfolios
that can be formed by various linear combinations of the price time series, whereas using
Engle-Granger approach we can only infer at most one linear combination if there exists one.
From Table 2, the Johansen test estimates of the order of integration from given price
series with 3 confidence levels, i.e. 10%, 5%, and 1%. We reject the null hypothesis for r = 0
since the test value is greater than confidence level’s value at 1%. Similarly, we accept the
null hypothesis for r = 1 at 95% confidence. We infer that BTC, ETH, LTC and BCH are
of 1st order cointegrated, i.e rank(r) = 1. Reading off the first column of Table 3, we find a
set of linear combinations for constructing a mean-reverting portfolio:

SPREADt = BTCt − 2.22ETHt − 29.43LTCt − 0.19BCHt (1)

This portfolio is essentially a spread between BTC and the other three cryptocurrencies
(ETH, LTC, BCH). Somewhat surprisingly, the names of Bitcoin (BTC) and BCH (Bitcoin
Cash) would suggest a comoving relationship, but in this portfolio BCH carries the least
weight compared to the other two cryptocurrencies, ETH and LTC, both in terms of units
and cash amount. During this 6-month period from December 20, 2017 to June 20, 2018,
prices of BTC were relatively high in comparison to ETH, LTC and BCH, as shown in
Figure 2. Thus, it is sensible to have a relatively high (short) stake on LTC as suggested
from equation (1). More importantly, we have established with 95% confidence that there
exists 1 linear combination of these 4 crypto-assets whose prices are stationary in the long-
term. This result is extremely important when constructing a mean-reverting portfolio using
Engle-Granger approach since we already knew beforehand the existence of cointegration
between crypto-assets’ prices. In the next section, we apply the Engle-Granger approach to
construct a cointegrating portfolio using again closing prices of BTC, ETH, LTC, and BCH.

test 10% 5% 1%
r≤3 1.97 6.50 8.18 11.65
r≤ 2 11.25 12.91 14.90 19.19
r≤ 1 22.25 18.90 21.07 25.75
r= 0 43.30 24.78 27.14 32.14
Table 2: Johansen estimates for testing cointegration

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=3235890
https://ssrn.com/abstract=3235890
 BTC ETH LTC BCH
BTC 1.00 1.00 1.00 1.00
ETH -2.22 0.75 -25.79 69.87
LTC -29.43 -28.66 85.76 1298.44
BCH -0.19 -2.79 -6.16 -94.57
Table 3: Coefficients from the Johansen test for cointegration

3.2 Engle-Granger Two-Step Method

The classical approach for cointegration testing developed by Engle and Granger (1987) has
the ability to capture long-term co-movements. We first perform a linear regression on some
given I(1) time series data, then test for stationarity of the residuals. It is crucial to ensure
that residuals from the OLS model are stationary; otherwise, we may run into a potential
issue, known as spurious regression. To assure stationarity, we then perform a series of
unit-root tests on the OLS’s residuals including the Augmented Dickey Fuller (ADF) test
(Dickey and Fuller, 1979), the Phillips-Peron (PP) test (Phillips and Perron, 1988) and the
Kwiatkowski-Phillips-Schmidt-Shin (KPSS) test (Kwiatkowski et al., 1992).
As the goal is to perform stationarity check on price series, first order price-differences
and residuals from the OLS model, testing hypotheses are designed as follows:
(i) Augmented Dickey Fuller test and Phillips-Peron test:
Ho : Presence of unit root in observable price series
Ha : Unit root does not exist in observable price series

(ii) Kwiatkowski-Phillips-Schmidt-Shin test:

Ho : Unit root does not exist in observable price series
Ha : Presence of unit root in observable price series

The Engle-Granger method involves a linear regression on given time series data which,
in this case, is BTC; but it is not important which cryptocurrency is taken as the dependent
variable. We propose the following OLS model:

BTCt = c + β1 ETHt + β2 LTCt + β3 BCHt + t

Once a cointegrating relationship is established, we perform 3 stationarity tests on the resid-

uals. Again, prices of these crypto-assets are said to be cointegrated only if prices are all I(1)
processes and the residuals t from the above OLS model are stationary.
In many cases, financial data follow a random walk. BTC, ETH, LTC, and BCH in our
study are not exempt from this assumption. According to the results from Table 4, we accept
the null hypothesis (there exits unit-root in time series) from both ADF and PP tests which
indicate that daily closing prices of all 4 cryptocurrencies are non-stationary and reject the
null hypothesis from KPSS test which again suggests that prices are non-stationary. However,
they all become stationary after the first differencing. Note that, one could theoretically reject
the null hypothesis from KPSS test performed on ∆BCH series since its p-value is right on
the edge of rejecting and accepting the null. Both ADF and PP tests, otherwise, suggest the
non-existence of unit root from the time series. Therefore, we conclude BTC, ETH, LTC,
and BCH are I(1) processes.

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=3235890
https://ssrn.com/abstract=3235890
 BTCt ETHt LTCt BCHt ∆BTCt ∆ETHt ∆LTCt ∆BCHt
ADF-test 0.49 0.36 0.24 0.41 0.01 0.01 0.01 0.01
PP-test 0.33 0.58 0.16 0.35 0.01 0.01 0.01 0.01
KPSS-test 0.01 0.01 0.01 0.01 0.10 0.10 0.10 0.05
Table 4: Summary of p-values from stationarity tests on cryptocurrencies’
prices and their first differences

An OLS model is introduced to establish linear relationships between BTC with ETH,
LTC and BCH, see Table 5. Notably, all the regressing coefficients appear to be statistically
significant with p-values less than 1%. Also, the adjusted R-squared of 94.5% is considered
remarkably high, especially in finance.

Dependent variable:
BTC
ETH 0.795∗∗∗
(0.264)
LTC 27.020∗∗∗
(1.770)
BCH 1.838∗∗∗
(0.140)
Constant 1,959.585∗∗∗
(176.725)
Observations 183
R2 0.950
Adjusted R2 0.949
Residual Std. Error 576.961 (df = 179)
F Statistic 1,138.254∗∗∗ (df = 3; 179)
∗
Note: p<0.1; ∗∗ p<0.05; ∗∗∗ p<0.01
Table 5: Ordinary least square regression model for BTC

From Table 5, we obtain the following linear model:

BTCt = 1, 959.59 + 0.795ETHt + 27.02LTCt + 1.838BCHt + t
Next, we conduct stationarity tests, i.e. ADF, PP and KPSS tests, on the residuals t from
this model. The results are as follows:
ADF test: rejecting the null-hypothesis ⇒ no unit-root
PP test: rejecting the null-hypothesis ⇒ no unit-root
KPSS test: accepting the null-hypothesis ⇒ residuals are stationary
All 3 stationarity tests confirm that the residuals from linear regression model are stationary
time series. Hence, prices of BTC, ETH, LTC and BCH are indeed cointegrated by our
procedure. With this confirmation, the spread can be constructed using the suggested linear
coefficients:
SPREADt = BTCt − 0.795ETHt − 27.02LTCt − 1.838BCHt (2)
This is the portfolio we will use to trade.

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=3235890
https://ssrn.com/abstract=3235890
3.3 Portfolio Time Series
The main purpose of our cointegration testing procedure is to construct a tradable mean-
reverting portfolio. Recall that there are two linear models given in (1) and (2) in which the
portfolio is a spread between BTC and three other cryptocurrencies. The fact that the linear
combinations in the two models are highly similar further supports that BTC, ETH, LTC
and BCH are cointegrated. Observing from Figures 3, it is hard to see which spread process
appears to be more stationary or mean-reverting. However, from a trading perspective,
periodic movements or fluctuations that lead to frequent crossing of the equilibrium levels
from both directions are desired, especially given that entry and exit rules are based on
deviations from the mean price (Chan, 2013). For this reason, Figure 3(b) suggests that
model (2) from the Engle-Granger approach is more practical to trade, and is thus chosen
for back-testing our trading system.

5000
8000

Mean Mean
Lower−Upper Bounds Lower−Upper Bounds
SPREAD SPREAD
6000

3000
USD

USD
4000

1000
2000

Jan Mar May Jan Mar May

(a) (b)
Figure 3: Spread constructed from two approaches: (a) Johansen methodology
(b) Engle-Granger methodology

Next, we construct a time series model to capture dynamic of the spread in order to
design the trading rules. Generally, we expect the spread to be highly mean-reverting since
that is the main criterion we used to choose the 4 crypto-assets in the first place. In discrete
time, the spread is often assumed to be stationary Autoregressive Moving Average (ARMA)
processes since they are also mean-reverting by design. This is the basic requirement we set
forth to designing our trading signals. From previous section, we derived from stationarity
tests that the spread is reverting around the $1,959.058 equilibrium. This helps us determine
the amount of inventory that are willing to hold on the spread, but more importantly, it
satisfies the stationarity requirement before we propose an appropriate ARMA model.
To estimate the orders and parameters of an ARMA process, we rely on sample autocor-
relation function (ACF) and partial autocorrelation function (PACF) plots with test bounds,
(see Box et al. (2015)). As suggested by ACF and PACF plots respectively in Figures 4(a)
and 4(b), the residuals can be modeled by ARMA(1, 0) or AR(1). We summarize the test in
Table 6.

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=3235890
https://ssrn.com/abstract=3235890
 1.0

0.8
0.8

0.6
0.6

Partial ACF

0.4
ACF

0.4

0.2
0.2

0.0
0.0

−0.2
−0.2

0 5 10 15 20 5 10 15 20

Lag Lag

(a) (b)
Figure 4: (a) ACF plot of the residuals from OLS model; (b) PACF plot of
residuals from OLS model

Dependent variable:
residuals
AR(1) Coeff. 0.874∗∗∗
(0.036)
Intercept 1,919.026∗∗∗
(160.731)
Observations 183
Log Likelihood −1,294.312
σ2 80,800.910
Akaike Inf. Crit. 2,594.625
∗ ∗∗ ∗∗∗
Note: p<0.1; p<0.05; p<0.01

Table 6: Fitting the spread to the ARIMA(1, 0, 0) model

To ensure that AR(1) is an appropriate model, Ljung-Box test is performed on residuals

from the ARMA model to check for the overall randomness and no autocorrelation in the
residual time series (see Box et al. (2015)). Ljung-Box test hypothesis:
Ho : Price series are independent, or no serial correlations
Ha : Price series are not independent, or they exhibit serial correlations
As shown in Figures 5(a) and 5(b), even with some lags outside the 95% confidence levels,
the p-value of 0.1094 for the Ljung-Box test on residuals implies the failure to reject the null
hypothesis that there is no autocorrelation in the residuals. We conclude that prices of the
spread are AR(1) processes.

10

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=3235890
https://ssrn.com/abstract=3235890
 1.0

0.10
0.8
0.6

Partial ACF

0.00
ACF

0.4

−0.10
0.2
0.0

−0.20
−0.2

0 5 10 15 20 5 10 15 20

Lag Lag

(a) (b)
Figure 5: (a) Autocorrelation graph of the residuals from OLS model; (b)
Partial autocorrelation graph of residuals from OLS model

4 Statistical Arbitrage Strategies

In trading practices, finding the right times for entry and exit is key to a profitable trading
system (see Leung and Li (2016); Vidyamurthy (2004)). As previously mentioned, the default
entry and exit thresholds in back-testing stage are typically set around 1 deviation, denoted by
σ, above and below the mean spread level, assuming that traders can go both long and short
on the spread. It is crucial to assure that these entry and exit levels are set to maximize
profits in terms of the trading costs and transaction frequency. Thus, by back-testing a
trading system with multiple entry/exit levels, we can get a sense of how profitable the
system is by looking at different performance criteria such as profit and loss (P&L), Sharpe
ratio, and more.
Define St as the spread at time t. A mean reversion strategy for trading the spread is
described by the following linear combination:

St = BTCt − β1 ETHt − β2 LTCt − β3 BCHt (3)

where (1, β1 , β2 , β3 ) := (1, −0.795, −27.02, −1.838) are the coefficients from the linear model
(2).
To purchase 1 “unit” of the spread S, we buy 1 unit of BTC, -0.795 unit of ETH, -27.02
unit of LTC, and -1.838 unit of BCH. In other words, in addition to buying 1 BTC, we need to
short sell 0.795 shares of ETH, 27.02 shares of LTC and 1.838 shares of BCH.9 Different from
traditional equity market, traders can purchase fractional shares up to 8 decimal places in
the cryptocurrency market. It boils down to how the inventor of Bitcoin, Satoshi Nakamoto,
decided to name Bitcoin in such way. When developing the Bitcoin, the creator Satoshi had
decided its base unit to be 108 and the smallest fraction is called Satoshi. Many recent altcoins
have adapted the same philosophy, though some have extended their smallest units up to 16
decimal places. For simplicity, we only go up to 3 decimal places on each cryptocurrency.
In addition, we only expose to 1 unit of the spread at any given time. That is to say
when if we already held 1 unit of the spread and observe a deviation of the spread from the
mean price, we do not add to our inventory any additional unit of S. As competitions in the
cryptocurrency market continues to rise, many recent exchanges start charging very minimal
9
Some exchanges, such as Bitfinex, allow for margin trading and short selling.

11

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=3235890
https://ssrn.com/abstract=3235890
fee per transaction (typically 0.1% or less); thus, we assume no transaction cost associated
with trading the spread.
Let us now state our trading rule:
(i) Long the spread/Exit short position when
St < µ − c · σ
(ii) Short the spread/Exit long position when
St > µ + c · σ
where µ is the mean value of the spread, σ is the standard deviation of the spread, and
c is a multiple we choose to set our entry/exit thresholds. At any given time t where St
falls under/over the lower/upper bound, the system will buy/sell 1 share of S then wait
until the price revert to realize a profit from that particular trade. Here, we backtest and
compare 5 different thresholds for entry/exit including ±0.5σ, ±0.75σ, ±1σ, ±1.25σ, and
±1.5σ (deviations from the mean).

4.1 Trading the spread

Testing for multiple entry/exit levels allows us to analyze the advantages and disadvantages
of the trading rules via multiple performance criteria. We summarize some trade statistics
in Table 7 (see Appendix for definitions). When setting lower/upper bound at ±1.5σ, our
system seems to outperform the rest in terms of annual Sharpe ratio, maximum drawdown,
max/min/end equities, and gross profits.

Threshold level ±0.5σ ±0.75σ ±σ ±1.25σ ±1.5σ

Num.Txns 17 13 13 9 7
Num.Trades 8 6 6 4 3
Net.Trading.PL 7601.87 8112.87 8307.29 6313.43 8422.57
Avg.Trade.PL 950.06 1200.60 1233.00 1351.04 2518.30
Largest.Winner 1599.35 1599.35 1599.35 1605.06 2773.72
Largest.Loser 0.00 0.00 0.00 0.00 0.00
Gross.Profits 7600.50 7203.60 7398.03 5404.16 7554.91
Gross.Losses 0.00 0.00 0.00 0.00 0.00
Percent.Positive 100.00 100.00 100.00 100.00 100.00
Percent.Negative 0.00 0.00 0.00 0.00 0.00
Avg.Win.Trade 950.06 1200.60 1233.00 1351.04 2518.30
Avg.Losing.Trade 0.00 0.00 0.00 0.00 0.00
Avg.Daily.PL 950.06 1200.60 1233.00 1351.04 2518.30
Ann.Sharpe 37.22 56.98 71.88 81.09 109.00
Max.Drawdown -2846.51 -2846.51 -2846.51 -2885.34 -1536.31
Profit.To.Max.Draw 2.67 2.85 2.92 2.19 5.48
Max.Equity 7996.42 8136.47 8330.90 6337.03 9334.01
Min.Equity -1020.73 -1020.73 -1020.73 -1009.32 -152.05
End.Equity 7601.87 8112.87 8307.29 6313.43 8422.58
Table 7: Statistics for trading strategies with different entry and exit thresholds

12

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=3235890
https://ssrn.com/abstract=3235890
 spread 2017−12−21 / 2018−06−20
5000

4500

4000

3500

3000
2700
2650
2600
2600
2550
2500 2500
2450
2400

2000

1500 1500
1400
1400
1300
1300
1200
1100
1000 1000
900

500

1.0 Positionfill 1 1.0

0.5 0.5

0.0 0.0

−0.5 −0.5

−1.0 −1.0
CumPL 8307.2932
8000 8000

6000 6000

4000 4000

2000 2000

0 0

0 Drawdown −23.60296 0

−2000 −2000

−4000 −4000

−6000 −6000

−8000 −8000

Dec Jan Feb Mar Apr May Jun

Dec 21 2017 Jan 08 2018 Jan 22 2018 Feb 05 2018 Feb 19 2018 Mar 05 2018 Mar 19 2018 Apr 02 2018 Apr 16 2018 Apr 30 2018 May 14 2018 May 28 2018 Jun 11 2018

Figure 6: Trading system setting entry/exit level at ±σ

The mean-reverting behavior of the spread, as we can see in Figure 6, means that trades
from all threshold levels yield positive profits, i.e. the percentages of profitable trades are
all 100%. When St moved away from its mean in early January 2018, we see the advantage
of setting a wide trading band (±1.5σ). In fact, the yield from that particular trade had
resulted in the highest realized profit, yet lower drawdowns in comparison to cases with
narrower trading bands. As the band widens, the number of trades and transactions decreases
as expected; but more importantly, net trading profit & loss (P&L) seems to be optimal
when we set the bandwidth at ±1.5σ which is consistent with all trading statistics from
the table. Average P&L and average daily P&L also go up together with the largest gain
when we increase the band. From asset management perspective, trading system with largest
threshold level yield astonishing Sharpe ratio of 109 with lowest maximum drawdown. As a
result, we observe that the trading system is optimal with the entry/exit level set at ±1.5σ.

4.2 Trading the spread with stop-loss

In the cryptocurrency market, we often observe larger price swings not only on daily basis,
but also potentially anytime of the day. This is the reason why setting a larger entry/exit
threshold could potentially boost our trading profits. At the same time, we cannot set
it too large that the trading system would result in no-trades. On the other hand, when
the entry/exit threshold levels get too narrow, traders would suffer larger drawdowns due
to extremely high volatility or short term deviation in crypto-assets’ prices. Therefore, in
addition to testing multiple entry/exit levels, we propose adding stop-loss order on top of the
existing 5 threshold levels trading systems, mainly to mitigate the effects of higher drawdown.
Here, we set an arbitrary stop-loss order at 5%, meaning our system will automatically
execute the stop-loss order a day after if today’s price falls below 5% of the buy-in price.

13

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=3235890
https://ssrn.com/abstract=3235890
 According to Table 8, stop-loss orders prevent the trading system from unrealized gains
that we could have earned when not implementing stop-loss orders. However, maximum
drawdown from all 5 threshold levels systems are reduced. Noticeably, numbers of trades
and transactions go up by a significant amount with lower net trading P&L, gross profits,
and most importantly, the Sharpe ratio. Without zero-transaction cost assumption, making
trades frequently could ultimately hurt the net trading P&L and gross profits of our system.
Notably, when imposing a 5% stop-loss, most trades got executed at a lower price and thus,
we do not fully capture the mean reverting benefit from the spread. In contrast to not having
stop-loss, more than 70% of trades from all 5 systems resulted in losses. Both Sharpe ratios
and end equities are significantly reduced.

Threshold level ±0.5σ ±0.75σ ±σ ±1.25σ ±1.5σ

Num.Txns 31 19 23 15 9
Num.Trades 15 9 11 7 4
Net.Trading.PL 2233.63 2285.58 3400.36 555.88 1672.08
Avg.Trade.PL 88.29 152.92 226.46 -50.48 201.10
Largest.Winner 954.18 1816.26 1816.26 1266.85 1506.03
Largest.Loser -145.28 -533.75 -533.75 -625.11 -380.16
Gross.Profits 2539.48 2615.69 3978.36 1266.85 1506.03
Gross.Losses -1215.11 -1239.38 -1487.26 -1620.24 -701.62
Percent.Positive 26.67 22.22 27.27 14.29 25.00
Percent.Negative 73.33 77.78 72.73 85.71 75.00
Avg.Win.Trade 634.87 1307.85 1326.12 1266.85 1506.03
Avg.Losing.Trade -110.46 -177.05 -185.91 -270.04 -233.87
Avg.Daily.PL 88.29 152.92 226.46 -50.48 201.10
Ann.Sharpe 3.64 3.39 4.84 -1.29 3.64
Max.Drawdown -791.28 -1325.03 -1325.03 -2010.37 -1225.63
Profit.To.Max.Draw 2.82 1.72 2.57 0.28 1.36
Max.Equity 2257.24 2640.13 3754.91 910.43 2583.51
Min.Equity -267.28 -801.02 -801.02 -1486.36 -701.62
End.Equity 2233.63 2285.58 3400.36 555.88 1672.08
Table 8: Statistics for the trading strategy with 5% stop-loss

14

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=3235890
https://ssrn.com/abstract=3235890
 spread 2017−12−21 / 2018−06−20
5000 5000

4500 4500

4000 4000

3500 3500
3500

3000 3000
3000
3000

2600
2500 2500
2500
2500
2400

2000 2000
2000
2000

1500 1500
1500
1500
1400
1300
1000 1000
1000
1000

500 500
500
500

1.0 Positionfill 1 1.0

1.0

0.5 0.5
0.5

0.0 0.0
0.0

−0.5 −0.5
−0.5

−1.0 −1.0
−1.0
CumPL 3400.36377

3000 3000

2000 2000

1000 1000

0 0

−1000 −1000
0 Drawdown −354.54648 0

−1000 −1000

−2000 −2000

−3000 −3000

−4000 −4000
Dec Jan Feb Mar Apr May Jun

Dec 21 2017 Jan 08 2018 Jan 22 2018 Feb 05 2018 Feb 19 2018 Mar 05 2018 Mar 19 2018 Apr 02 2018 Apr 16 2018 Apr 30 2018 May 14 2018 May 28 2018 Jun 11 2018

Figure 7: Trading system with 5% stop-loss, setting entry/exit level at ±σ

4.3 Trading the spread with trailing stop-loss

While a stop-loss order limits the traders’ net loss, a trailing stop is triggered by a large
drawdown and is adjusted upward as the portfolio value increases. Hence, it has the advantage
of continuously moving with the portfolio value and allowing the trader to potentially end
with a gain even when the trailing stop is triggered by a drawdown.
When implementing a strategy with trailing stop, there is also a possibility that trades
get exited too early before a bigger upward price movement. This is the opportunity cost of
having trailing stop rather than just implementing stop-loss alone. In our case, since trading
signals are only triggered when prices of the spread fall above or drop below the entry/exit
level, it could potentially prevent the trading system from executing additional trades as
price fall within one deviation threshold. Also, if we know with certain level of confidence
that the spread is a stationary process, it may not be ideal to implement either stop-loss or
trailing stop . Nonetheless, we design additional trading systems setting a trailing stop at
the same level as we did for stop-loss, which is 5%, to compare 3 trading systems with 5
threshold levels and evaluate the benefit of having one over the other using multiple trading
performance criteria.
From Table 9, the number of trades and transactions increase significantly from the
original trading system which could potentially be an expense when there are transaction
costs associated with each trade. However, net trading P&L, average trade P&L, gross
profits are higher than from the system with stop-loss. Here, the trailing stop has led to
multiple exits with small profits as the spread reverts to the mean. As a result, percentages
of profitable trades have gone up two times from the stop-loss trading systems and average

15

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=3235890
https://ssrn.com/abstract=3235890
daily P&L are also improved. Notably, maximum drawdown are drastically reduced while
Sharpe ratios and end equities significantly increase.

Threshold level ±0.5σ ±0.75σ ±σ ±1.25σ ±1.5σ

Num.Txns 36 24 26 20 10
Num.Trades 18 12 13 10 5
Net.Trading.PL 3509.31 3872.88 5080.65 4682.81 2114.03
Avg.Trade.PL 194.96 322.74 390.82 468.28 422.81
Largest.Winner 954.18 1816.26 1816.26 1266.85 1303.67
Largest.Loser -137.35 -137.35 -137.35 -125.72 -179.62
Gross.Profits 3991.64 4440.42 5367.59 4849.46 2334.59
Gross.Losses -482.33 -567.55 -286.94 -166.66 -220.56
Percent.Positive 66.67 58.33 76.92 80.00 60.00
Percent.Negative 33.33 41.67 23.08 20.00 40.00
Avg.Win.Trade 332.64 634.35 536.76 606.18 778.20
Avg.Losing.Trade -80.39 -113.51 -95.65 -83.33 -110.28
Avg.Daily.PL 194.96 322.74 390.82 468.28 422.81
Ann.Sharpe 8.85 8.33 11.03 14.44 9.82
Max.Drawdown -676.06 -1021.65 -676.06 -1040.13 -1066.16
Profit.To.Max.Draw 5.19 3.79 7.52 4.50 1.98
Max.Equity 3657.30 4020.86 5228.64 4830.80 2365.69
Min.Equity -152.05 -497.65 -152.05 -516.12 -542.16
End.Equity 3509.31 3872.88 5080.65 4682.81 2114.03
Table 9: Statistics for the trading strategy with 5% trailing stop-loss

Starting with an initial amount of $50,000 and restricting the exposure of our trading
systems to only 1 unit of the spread at a time, Figure 8 shows how the three systems per-
form with 5 proposed entry/exit threshold levels. Overall, the trading system with neither
stop-loss nor trailing stop produced maximum profits with highest Sharpe ratios. Although
traders may not prefer drawdowns from the initial trade, the spread reverted to the mean
price and crossed above the exit level, yielding a total net profit of over $1,599 which is
equivalent to 175% gain on the period from December 28, 2017 to January 14, 2018. By
setting larger bands, we were able to capture full profit from the initial trade with little to
no drawdown, see Figure 8(e). In addition, by not incorporating a stop-loss nor trailing stop,
setting threshold at ±1.5σ has yielded highest profit and Sharpe ratio out of all the proposed
trading systems. Although when implementing trailing stop and stop-loss, we are able to re-
duce drawdowns from the initial system, trading performances also decline. In terms of profit
maximization, the trading system without stop-loss nor trailing should be preferred since we
have established that a portfolio comprised of BTC, ETH, LTC, and BCH is mean-reverting.

16

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=3235890
https://ssrn.com/abstract=3235890
 Equity Line With.Stop.Loss With.Trailing.Stop.Loss Equity Line With.Stop.Loss With.Trailing.Stop.Loss

58000 58000

56000 56000

54000
USD

USD
54000

52000 52000

50000 50000

Jan Feb Mar Apr May Jun Jan Feb Mar Apr May Jun
Date Date

(a) (b)

Equity Line With.Stop.Loss With.Trailing.Stop.Loss

58000

56000
USD

54000

52000

50000

Jan Feb Mar Apr May Jun

Date

(c)

Equity Line With.Stop.Loss With.Trailing.Stop.Loss Equity Line With.Stop.Loss With.Trailing.Stop.Loss

56000

57000
54000
USD

USD

52000 54000

50000 51000

Jan Feb Mar Apr May Jun Jan Feb Mar Apr May Jun
Date Date

(d) (e)
Figure 8:
Equity lines: (a) Entry/exit threshold = 0.5σ; (b) Entry/exit threshold
= 0.75σ; (c) Entry/exit threshold = σ, (d) Entry/exit threshold = 1.25σ.
(e) Entry/exit threshold = 1.5σ.

17

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=3235890
https://ssrn.com/abstract=3235890
5 Conclusion
In this study, we have identified price cointegration among the prices of four cryptocurrencies,
BTC, ETH, LTC, and BCH, using both the Johansen test and Engle-Granger procedure.
Although the Johansen test implies a stationary portfolio, it is not a preferred approach for
trading application because the resulting slow mean-reverting portfolio would result in very
few to no trades. The Engle-Granger approach gives a mean-reverting and tradable portfolio
since the fast mean-reverting spread would cross the entry and exit levels frequently. We find
that setting greater entry and exit levels leads to larger profits. In comparison to trading
systems with a stop-loss exit or trailing stop, the one without yields the highest profit but
also largest drawdowns.
As the cryptocurrency market continues to grow with new coins and new exchanges, it
is very important for individual investors, crypto-fund managers, as well as regulators to
understand the price dependency among all cryptocurrencies, along with their derivatives.
There are many directions for future research. For example, do the prices and volatilities of
cryptocurrencies interact with the equity, bond, or commodity markets (see Narayan et al.
(2017))? Can they be used as an inflation hedge like gold (see (Ghosh et al., 2004))? In this
paper, we have considered several cryptocurrencies with the largest market capitalizations,
but one can also develop a machine learning approach that automatically selects a small
portfolio from a vast universe of cryptocurrencies (see Zhang et al. (2018)). Moreover, one
can analyze and compare the price impacts of trading different cryptocurrencies. Results on
this front will shed light on the process of selecting a cointegrated portfolio and the execution
of trades.

References
Box, G. E., Jenkins, G. M., Reinsel, G. C., and Ljung, G. M. (2015). Time series analysis:
forecasting and control. John Wiley & Sons.

Chan, E. (2013). Algorithmic trading: winning strategies and their rationale, volume 625.
John Wiley & Sons.

Cheah, E.-T. and Fry, J. (2015). Speculative bubbles in bitcoin markets? an empirical
investigation into the fundamental value of bitcoin. Economics Letters, 130:32–36.

Chuen, K., David, L., Guo, L., and Wang, Y. (2017). Cryptocurrency: A new investment
opportunity? Working paper.

Detzel, A. L., Liu, H., Strauss, J., Zhou, G., and Zhu, Y. (2018). Bitcoin: Learning, pre-
dictability and profitability via technical analysis. Working paper.

Dickey, D. A. and Fuller, W. A. (1979). Distribution of the estimators for autoregressive

time series with a unit root. J. Am. Stat. Assoc., 74(366):427–431.

Engle, R. F. and Granger, C. W. J. (1987). Co-integration and error correction: Represen-

tation, estimation, and testing. Econometrica, 55(2):251–276.

Ghosh, D., Levin, E. J., MacMillan, P., and Wright, R. E. (2004). Gold as an inflation hedge?
Studies in Economics and Finance, 22(1):1–25.

Ha, S. and Moon, B.-R. (2018). Finding attractive technical patterns in cryptocurrency
markets. Memetic Computing, pages 1–6.

Hileman, G. and Rauchs, M. (2017). Global cryptocurrency benchmarking study. Cambridge

Centre for Alternative Finance. Available at https://ssrn.com/abstract=3040224.

18

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=3235890
https://ssrn.com/abstract=3235890
Johansen, S. (1988). Statistical analysis of cointegration vectors. Journal of Economic
Dynamics and Control, 12(2-3):231–254.

Kang, J. and Leung, T. (2017). Asynchronous ADRs: overnight vs intraday returns and
trading strategies. Studies in Economics and Finance, 34(4):580–596.

Kwiatkowski, D., Phillips, P. C., Schmidt, P., and Shin, Y. (1992). Testing the null hypothesis
of stationarity against the alternative of a unit root: How sure are we that economic time
series have a unit root? Journal of Econometrics, 54(1-3):159–178.

Labbe, A. (2017). PRC regulator blames fraud for ICO ban. International Financial Law
Review.

Leung, T. and Li, X. (2015). Optimal mean reversion trading with transaction costs and
stop-loss exit. International Journal of Theoretical & Applied Finance, 18(3):15500.

Leung, T. and Li, X. (2016). Optimal Mean Reversion Trading: Mathematical Analysis
and Practical Applications. Modern Trends in Financial Engineering. World Scientific,
Singapore.

Leung, T. and Ward, B. (2015). The golden target: Analyzing the tracking performance of
leveraged gold ETFs. Studies in Economics and Finance, 32(3):278–297.

Makarov, I. and Schoar, A. (2018). Trading and arbitrage in cryptocurrency markets. Work-
ing paper. Available at https://ssrn.com/abstract=3171204.

Narayan, P. K., Thuraisamy, K. S., and Wagner, N. F. (2017). How do bond, equity and
commodity cycles interact? Finance Research Letters, 21:151 – 156.

Phillips, P. C. and Perron, P. (1988). Testing for a unit root in time series regression.
Biometrika, 75(2):335–346.

Vidyamurthy, G. (2004). Pairs Trading: quantitative methods and analysis, volume 217.
John Wiley & Sons.

Wagner, N. (2005). Autoregressive conditional tail behavior and results on government bond
yield spreads. International Review of Financial Analysis, 14(2):247 – 261. Special issue:
Quantitative Issues in Finance.

Zhang, J., Leung, T., and Aravkin, S. (2018). Mean reverting portfolios via penalized OU-
likelihood estimation. In Proceedings of the 57th IEEE Conference on Decision and Con-
trol, 2018.

19

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=3235890
https://ssrn.com/abstract=3235890
A Appendix: statistics for trading strategies

Num.Txns Total number of transactions

Num.Trades Total number of trades
Net.Trading.PL Total profit & loss made due to all trades
Avg.Trade.PL Average profit & loss made due to all trades
Largest.Winner The maximum profit of trades
Largest.Loser The maximum loss of trades
Gross.Profits Total profits
Gross.Losses Total losses
Percent.Positive Percentage of profitable trades
Percent.Negative Percentage of loss trades
Avg.Win.Trade The average of profits of trades
Avg.Losing.Trade The average of losses of trades
Avg.Daily.PL Average profit & loss on daily basis
Ann.Sharpe Sharpe ratio with 0 risk-free rate
Profit.To.Max.Draw The ratio of profit to maximum drawdown
Avg.WinLoss.Ratio The ratio of average profitable trades and average loss trades
Max.Equity Maximum equity
Min.Equity Mimum equity
End.Equity Ending equity

Table 10: Trade statistics explanations

20

Electroniccopy
Electronic copyavailable
available at:
at: https://ssrn.com/abstract=3235890
https://ssrn.com/abstract=3235890