Informatica Economică vol.22, no. 2/2018 35

DOI: 10.12948/issn14531305/22.2.2018.04

Arbitrage Trading Systems for Cryptocurrencies.

Design Principles and Server Architecture

Cristian PĂUNA

The Bucharest University of Economic Studies, Romania

cristian.pauna@ie.ase.ro

When dot.com has become a quaint idea, when electronic shops have lost the mass attention,

while classical and margin trading has become obsolete, something new is coming: cryptocur-

rencies. Hundreds of virtual coins have been invented for a single reason: the profit. The high

price volatility of these new markets and the fact that the virtual coins price is not regulated by

a central bank or a single exchange, gives us opportunities for arbitrage trading. The existence

of important price differences makes possible the profit when an automated system buy cheaper

and sell more expensive in the same time. This paper will present the general principles under-

pinning the implementation of arbitrage trading software for virtual coins market. The very

large number of cryptocurrencies and exchanges fundamentally change the server architecture

of the trading software. The distributed price data in hundreds of sources and the technical

differences of each of these data providers make all the things difficult to be implemented in a

single application. The low-latency order calculation needed for the fast delivery before a sig-

nificant price change, in the presence of thousands of price quotes coming from hundreds of

distributed servers makes everything special.

Keywords: Cryptocurrency (CRYC), Arbitrage trading (ART), Algorithmic trading (AT), High

Frequency Trading (HFT), Automated trading software (ATS)

Introduction

Bitcoin (BTC), Litecoin (LTH), Ethereum

(ETH), Ripple (RIP), thousands of virtual

coins already exists. The Bitcoin “introduced

in 2009”, “caught the interest of the main-

stream media in 2012” [1]. After that, an un-

defined number of virtual coins were in-

vented. All these coins are named virtual just

because they exist only in the electronic rep-

resentation. They are also called cryptocurren-

cies because they are designed to work as a

currency like a medium of value exchange and

the “cryptography is used to secure financial

transactions” [2], to control the creation of

new units, to verify the transfer of funds and

to set the owner for each unit.

The reason involving real money to buy this

kind of virtual assets is of course the profit.

The virtual coins value is fluctuating on the

free markets as any asset, powered by specu-

lative and marketing actions and sustained at

least hypothetical by some economical pro-

cesses involved in the background of each

crypto coin. However, even it is a free market,

the virtual coin market has its own character-

istics, especially when it is about the volatility

and “the flexibility in the supply schedule”

[3].

We are not so far away the days when Bitcoin

almost touched 20.000 USD in a relative short

period of time starting from couple of bucks

and after that plunged back to 7.000 - 8.000

USD in an even shorter interval. This kind of

movements makes the investment in crypto-

currencies an attractive one today. However,

the high price volatility makes this investment

to be a very risky one, drawdowns of more

than 61.8% being very easily assimilated with

a gambling game. For all that, keeping the

money into a virtual currency hoping the

value will increase after a while can be a prof-

itable idea, but it is not the only one. All these

assets are operating independently of a central

bank or a central exchange.

Without a central price regulator the quotes

can be different from an exchange to another.

That means we can find cases when and where

we can buy cheap and sell more expensive the

same thing in the same moment of time. This

will be called an arbitrage trade. The unbal-

anced price of the same asset in two different

places will give us an instant profit.

36 Informatica Economică vol.22, no. 2/2018

DOI: 10.12948/issn14531305/22.2.2018.04

The question is how to find these opportuni-

ties and how can we profit from all of these in

the real time. The theory is easy, just find two

exchanges with different quotes for the same

equity, buy where the equity is cheaper and

sell where it is more expensive.

The problems start when we will consider the

facts that there are hundreds of crypto pairs,

hundreds of thousands of combination of

these coins and hundreds of exchanges where

you can buy and sell all of these assets. That

means a human being cannot compare all

these prices in order to find the best trade op-

portunities, to build the orders and to send

them to those exchanges before the price is

changed. Only a computer can to manage all

of these actions.

This paper will present how an arbitrage trad-

ing system for cryptocurrencies can be built,

will reveal the main principles and practices

and will also present some results in order to

open or to grow the interest for this kind of

software systems.

2 Arbitrage trading

As we already presented, the arbitrage trading

means to buy cheap and to sell more expen-

sive the same equity in the same moment of

time in two places where the quote prices are

different. The arbitrage trading is not a new

idea. It was used since years to trade shares.

“Stocks are matched into pairs with minimum

distance between normalized historical

prices” [4] and arbitrage trading strategies can

be used with any stock exchange in order to

make profit.

Usual, the classical arbitrage trade is involv-

ing the same volume of equity in the both

trades. For each trade (i), the profit (P

) is

equal with the difference between the sell

price (sell

) and the buy price (buy

) multiplied

by the volume traded (V

For this simple trade idea, the trading software

must compare continuously the quotes from

different exchanges and find the highest price

differences. But the classical arbitrage trading

is not the only one trading opportunity which

can be found in the cryptocurrency markets.

The virtual coins are traded in pairs. The price

of a pair, for example BTCETH if we will

consider Bitcoin and Ethereum, means how

much ETH we need to sell in order to buy a

specified unit of BTC. When an exchange ask

you a specified price to buy BTCETH and an-

other exchange offers you a higher price for

the same pair, a classical arbitrage trade can

be executed and the profit will result from the

price difference. But sometimes the depend-

ency between two coins is not expressed in the

same way in all exchanges. There are cases

when an exchange will quote the price for

BTCETH and another exchange will quote the

price for ETHBTC. The virtual coins involved

are the same but the price in built in the re-

versed way. For this case, to find an arbitrage

trade the software must compare the price

from the first exchange with the reversed price

into the second. Now we are talking about a

reversed arbitrage trade. In this case the profit

will be:

The execution of the trades is usual simple. A

trading order means to buy or sell the speci-

fied pair at a specified price with an estab-

lished volume. Of course an order can be exe-

cuted by the exchange only if there are enough

coin units to sell in that exchange in order to

buy the ordered volume from the first coin.

This validation must to be done by the trading

software before to send the orders to the ex-

changes, based on the real-time capital data

provided by each exchange. An arbitrage

trade means at least two orders in two differ-

ent exchanges.

Another type of arbitrage trades is involving

more pairs. Sometimes the product of two,

three or more pair quotes is imbalanced be-

tween two exchanges or even in the same ex-

change. For understanding, let's consider the

next inequalities:

BTCETH*ETHLTC<BTCLTC (3)

and

BTCETH*ETHLTC*LTCRIP<BTCRIP (4)

Informatica Economică vol.22, no. 2/2018 37

DOI: 10.12948/issn14531305/22.2.2018.04

In cases like these, multiple arbitrage trades

can be possible in order to make profit. In the

first exchange will be executed more trades

and in the second only one. The software will

buy the lowest price for multiple pairs and sell

the last one on a higher price. The profit for

multiple arbitrage trades will be expressed in

this case by the formula:

For the next level of multi arbitrage trades,

when we buy more pairs and sell more pairs

in order to earn the difference between those

operations, the profit will be expressed by the

next equation:

where (nBUY) is the number of buy trades

and (nSELL) is the number of sell trades made

simultaneous. This kind of trades can be found

using two different exchanges but sometimes

the difference can be found even in the same

exchange. The multiple arbitrage trades are

more complex to be validated and executed

and a serious computational power is needed

in order to deliver all the trading signals and

orders in real time. Only low-latency trading

software can manage this process in order to

deliver the orders in that short time needed to

execute them.

3. Cryptocurrency dictionaries

One of the most important parts of an arbi-

trage trading system for cryptocurrencies is

the dictionary. Due to a very high number of

cryptocurrencies quotes involved and because

there is not a standard for naming the coins

and pairs, a lot of false trading signals can be

built without a strict codification methodol-

ogy. Each exchange has its own coin and pair

codes. We can find cases when the same coin

can be named or coded differently.

For example BTC can be the symbol for

Bitcoin in the most exchanges but some of

them will code the Bitcoin with XBC or BTX.

In order to compare the price of the same cur-

rencies, a codification dictionary must to be

implemented. The trading software must have

its own codification rules and all codes from

all exchanges must to be included into the

trading software database.

In addition there are cases when the code of a

pair into an exchange can be the code of an-

other pair into another exchange. For example

SCAPPC is representing the pair of SC re-

ported to the APPC coin in the most ex-

changes but it others it is the pair between

SCA and PPC. All these codes SC, SCC, PPC

and APPC are valid crypto coin codes. This is

the case when a profitable arbitrage trade can

be found if we will compare the price of the

SCAPPC pair into two exchanges where that

code represent different coin pairs. To avoid

these type of false signals, in the trading soft-

ware dictionary must to be inserted each pair

code from each exchange related with the real

pairs that are quoted. A codification using a

separator is a very good implementation. With

this additional rule, the code will looks like

SC-APPC and SCA-PPC and the confusion is

avoided.

From this reason, the cryptocurrencies dic-

tionary implementation and continuous opti-

mization is a very significant part of the arbi-

trage trading software for virtual coins. More,

the current practice added to this reason an-

other important one. In fact more exchanges

are changing the codification of the crypto

coins and pairs occasionally. This means reli-

able trading software must have the possibility

for an easy change of a coin code or pair code,

in order to save time and to trade the most sig-

nificant profit opportunities without a signifi-

cant change in the code of that software.

In this moment (2018 year) there are more

than 3000 virtual coins on the free market

listed in more than 300 exchanges and we are

still counting. Not on the last place, the fact

that new crypto coins are invented every day

and they are added often into the exchanges

quotes makes the dictionary to be a significant

tool in order to keep the trading software up-

dated with the reality of cryptocurrency mar-

kets.

38 Informatica Economică vol.22, no. 2/2018

DOI: 10.12948/issn14531305/22.2.2018.04

4 Data structure

The price quotes are received from the ex-

changes or from the brokerage companies in

real-time in different data structures. Each ex-

change has its own API application and the

data structure of each exchange is different.

Even all data sources are available by an API

access, the differences between data structures

delivered by each exchange make everything

more complex. The arbitrage software must

be adapted to import all the piece quotes from

each exchange according each technical spec-

ifications.

Another difference is given by the fact that

some exchanges will deliver all the price

quotes on a single interrogation of their API,

others exchanges will deliver the quotes in

different interrogations, one request for each

currency pair price. The third level of diffi-

culty is occurred for those exchanges that ac-

cept a limited number of requests in a speci-

fied time interval. All of these facts make con-

nection with each exchange to be a dedicated

and laborious job.

Whatever are the difficulties to implement the

data connection with the exchanges, once the

price quote is received in real-time, it must to

be recorded into the application database. Be-

cause the update interval is random, each

quote will be inserted as a separate record re-

lated with the exchange code, with the pair

code and with the updated time. For those pro-

cesses that need the historical price quotes, be-

cause of the large number of the records, usual

the history is simplified and represented sepa-

rately as historical records which will include

the next components, that become a standard

for a historical price data series [5]: open price

of a time interval (OPEN), maximal price of a

time interval (MAX), minimal price of a time

interval (MIN), close price of a time interval

(CLOSE), where the time interval has usual

one of the next values: one minute (1M), five

minutes (5M), fifteen minutes (15M), thirty

minutes (30M), one hour = 60 minutes (1H),

four hour = 240 minutes (4H), one day = 1440

minutes (1D), one week = 10080 minutes

(1W), one calendar month (1M), one calendar

year (1Y).

Usual start time interval and data about the

traded volumes are also available together

with the price quotes. Advanced software will

record also the current spread for the equities,

which is the difference between the buy and

the sell price at a moment of time.

All price data quotes from hundreds of ex-

changes and for hundred of thousand of cryp-

tocurrency pairs must to be made in real time.

In addition the data history must to be recorder

for future operations. Using a relational data-

base is a fact. In addition, the huge number of

CRUD operations and the low-latency needed

for the trading signals and orders push the ap-

plication to use multi server architecture pre-

sented in Figure 1. In the current practice,

more servers will do the acquisition processes

of the real time price data from the exchanges,

a different server will makes all CRUD (cre-

ate, record, update, delete) operations and a

different one will compute the trading signals

and will build and deliver the low-latency

trading orders. All of these operations must to

be completed in real-time, before the price is

changing significantly the values in order to

make possible the arbitrage trades.

Fig. 1. Server architecture for a cryptocur-

rency arbitrage system

Informatica Economică vol.22, no. 2/2018 39

DOI: 10.12948/issn14531305/22.2.2018.04

5 SQL needed

Due to the large amount of data, first reasons

to use a relational database are to control data

redundancy, to provide the data integrity and

to avoid data inconsistency. The concurrency

control needed by using several servers to up-

date the same data warehouse with the real-

time price quotes will be also provided by a

relational database. In addition, to restrict the

data access to different levels of authorization

and to provide reliable data backup and recov-

ery, the usage of an SQL relational database

will be the solution, even today when NoSQL

databases are in trend.

In the context of a large volume of CRUD op-

erations and data requests, SQL will provide

the low-latency data responses. The most sig-

nificant example is the signal generator

server. To see how much can be the time re-

sponse difference we have compared a repeti-

tive process to build the arbitrage trading sig-

nals and an SQL procedure to extract the same

signals from the database. In the first case, a

classical procedure to build the trading signals

was wrote to find better prices for each price

quote in the database using only repetitive

code statements.

For a database including more than 1,000,000

real-time quotes, the process to build the arbi-

trage trading signals was longer than two

minutes on a dedicated server with 2x3GHz

processor core and 4 GB RAMM Burst. This

time is too long for reliable trading software.

In more than two minutes the quotes can be

significantly changes, meaning those signals

have a low chance to be executed after the de-

livery on the brokerage account.

In the second case it was implemented an SQL

procedure to build the arbitrage trading sig-

nals with the same amount of data. The results

using SQL are conclusive; the time response

was about 0.25 seconds which is a realistic

time for the trading software. In Figure 2 is

presented a sample code for node.js using

callback functions and SQL to build the arbi-

trage trading signals.

Fig. 2. node.js callback functions for the classical arbitrage trading signals module.

40 Informatica Economică vol.22, no. 2/2018

DOI: 10.12948/issn14531305/22.2.2018.04

6 API data server

A special particularity of the arbitrage trading

systems is the API server that collects real-

time quotes from each data source. Each ex-

change has organized an API application in

order to permit restricted and real-time access

to the data. To interrogate that API, to receive

low-latency data and to insert it in the local

database or warehouse, we need to organize a

special server process for each exchange. In

the Figure 3 it is presented the code for a data

acquisition process using an API. The “Ex-

change” object represents the Rest API proce-

dure module distributed by each exchange.

“bookTickers” is the function from this mod-

ule that exports the price data. Each exchange

has its own exporting functions, the name of

this function and the “data” structure deliv-

ered by this function are different, and conse-

quently the code function must to be adapted

for each particular case.

The data delivered by each exchange is usual

a JSON object [6] including different data re-

lated with the current state of the market. Be-

cause the structure of this JSON object is dif-

ferent from an exchange to another, a general

procedure cannot be built. In the example pre-

sented, the “setInterval” function is a prede-

fined node.js function to repeat the process. In

our case we set to repeat the price quotes im-

port each minute (60 sec * 1000 = 60000).

Setting this time is also a procedure depending

on each exchange. There are exchanges that

allows one per minute requests, other ex-

changes will allow interrogations one time at

two or five minutes and others at ten or larger

interval. Usually when the request interval is

not respected, the API can cancel the connec-

tion with the server and no data will be re-

ceived from that source. In the example

above, the “updateQuote” function is a com-

mon function to update the records with the

piece quotes, this function will be used for any

exchange server, once the data is reduced to

an established data structure.

Fig. 3. node.js call back functions for the real-time price data acquisition server

Important is the fact that the real-time repeti-

tive price process is a server process, orga-

nized for each exchange and depending on the

particularities of each API data source. The

update time is also depending on the data

source. The usage of the node.js server “is

quite lightweight and efficient” [7] and permit

a fast and simple implementation of these en-

tire server processes. When the number of

these modules is a significant one, using

node.js we can distribute the repetitive proce-

dures on more distributed servers, as a meas-

ure to keep the data delivery time under those

limits to ensure the low-latency response.

7 Arbitrage trading signals

In order to consider why we have to do all

these technical efforts, in this chapter it will

be presented a real case of arbitrage trading

signals with cryptocurrencies. Using theCryp-

toTrader [8] the next signals were obtained at

25 May 2018 using 787 cryptocurrencies

listed on 7 different exchanges.

Informatica Economică vol.22, no. 2/2018 41

DOI: 10.12948/issn14531305/22.2.2018.04

Table 1. Arbitrage trading signals made by theCryptoTrader on 25.05.2018

Arbitrage

trading sig-

nal

Buy

exchange

Buy price

Sell

exchange

Sell price

Profit [%]

BTG-BTC

Cryptopia

0.00101454

Binance

0.00604100

495.44

CTM-BTC

Cryptopia

0.00000929

Binance

0.00004469

381.05

BTC-STEEM

Poloniex

0.00029422

Bittrex

0.00036467

23.94

BTC-SBD

Bittrex

0.00023894

Poloniex

0.00026653

11.55

BTC-OMNI

Bittrex

0.00338255

Poloniex

0.00367822

8.74

BTC-SYS

Poloniex

0.00004503

Bittrex

0.00004895

8.71

BTC-GBP

Kraken

5398.10

Gdax

5653.55

4.73

BTC-NAV

Bittrex

0.00012728

Poloniex

0.00013294

4.45

BTC-DOGE

Poloniex

0.00000046

Bittrex

0.00000048

4.35

BTC-PINK

Bittrex

0.00000250

Poloniex

0.00000258

3.20

As we can see in the Table 1, the profit expec-

tation is a considerable one. The signals pre-

sented are the best ten arbitrage trading sig-

nals from a list of about 186 available signals

found at that moment by the trading software.

Any investor who has available money into

those exchanges can trade that signal in order

to make profit. Usual the signal is not availa-

ble for a very long period of time. This is the

reason to use trading software that has to find

the trading opportunities and to trade them

with low-latency orders sent direct to the ex-

changes.

The arbitrage trading software permits a very

fast order delivery for each signal direct to the

exchanges into the investor's accounts. If the

orders arrived before the significant change of

the price quotes, they will be executed and the

profit will be immediately received into the

investor's account. However, time to time,

some signals are false. Some exchanges de-

liver price quotes for some cryptocurrency

pairs but they refuse to execute orders at those

prices. This is usual the fact when the profit

expectation of a signal is too high. Even so,

signals with profit expectation under 999%

can be well executed in the real practice. The

limit is imposed by the exchange by the traded

volume which cannot exceed a specified

value. Some exchanges limit the trading vol-

ume depending on the investor's profile. How-

ever, with all that limits, scheduling the exe-

cution of a signal to be repeated time to time

as a function of the trading software will per-

mit the profit for capitals higher than the trad-

ing limits imposed by each exchange. Usual

signals with the profit expectation around

20% are well traded.

8 Conclusions

Cryptocurrency market price is not regulated

by a central bank. Arbitrage trading opportu-

nities can be found on the free markets in or-

der to make profit. Because of the large num-

ber of virtual coins and exchanges, finding

and computing the arbitrage trading signals

must to be done by the automated trading soft-

ware. A crypto dictionary is a must to have in

the trading software with cryptocurrencies be-

cause of the important codes differences be-

tween exchanges. In addition, this module will

permit all frequent changes in the codification

and new virtual coins to be managed.

The technical differences between data

sources impose an individual server process to

be organized for each data source. A high

number of data sources mean a high number

of server processes. This implies implementa-

tion of more data acquisition process servers

distributed in different technical resources.

Due to the large amount of relational data, re-

lational database and SQL is a reliable solu-

tion in order to ensure the low-latency re-

sponse of the trading software. Arbitrage trad-

ing software with cryptocurrencies delivers a

significant profit.

42 Informatica Economică vol.22, no. 2/2018

DOI: 10.12948/issn14531305/22.2.2018.04

References

[1] N. Gandal, H. Halaburda, “Can We Predict

the Winner in a Market with Network Ef-

fects? Competition in Cryptocurrency

Market Design and Implementation”,

2016, NET Institute Working Paper No.

14-17, Available: https://ssrn.com/ab-

stract= 2506463, May 9, 2018

[2] Wikipedia Encyclopedia - “Financial

Cryptography” - https://en.wikipe-

dia.org/wiki /Financial_cryptography,

May 10, 2018

[3] M. Iwamura, Y. Kitamura, T. Matsumoto,

K.Sailo, “Can We Stabilize the Price of a

Cryptocurrency?: Understanding the De-

sign of Bitcoin and Its Potential to Com-

pete with Central Bank Money”, Institute

of Economic Research Hitotsubashi Uni-

versity Kunitachi, Tokyo, 186-8603 Japan,

2014 Available: https://ssrn.com/abstract

=2519367, May 9 2018

[4] E. Gatev, “Pairs Trading: Performance of

a Relative-Value Arbitrage Rule” - The

Review of Financial Studies, Volume 19,

Issue 3, 2006, Pages 797–827, Available:

https://doi.org/10.1093/rfs/hhj020, May

10, 2018.

[5] MetaTrader 4 (MT4) - Historical price

data structure - Available:

https://docs.mql4.com/ constants/struc-

tures/mqlrates, May 10, 2018.

[6] JavaScript Object Notation (JSON) – “The

javascript object notation (json) data inter-

change format” - Available:

https://buildbot.tools.ietf.org/html/rfc825

9, May 10, 2018.

[7] K. Ley, Y. Ma, Z. Tan, “Performance

Comparison and Evaluation of Web De-

velopment Technologies in PHP, Python,

and Node.js”, 2014 - Computational Sci-

ence and Engineering (CSE), 2014 IEEE

17th International Conference, Available:

https://ieeexplore.ieee.org/ docu-

ment/7023652, May 11, 2018

[8] Cryptocurrency Trading software Platform

- theCryptoTrader - Available:

https://pauna. biz/thecryptotrader, May

12, 2018

Cristian PĂUNA graduated the Faculty of Cybernetics, Statistics and Eco-

nomic Informatics of the Academy of Economic Studies in 1999 and also he

is a graduate of the Faculty of Aircrafts of the Polytechnic University of Bu-

charest in 1995. He got the title of Master of Science in Special Aerospace

Engineering in 1996. In the last decades he had a sustained activity in software

development industry, especially applied in the financial trading domain.

Based on several original mathematical algorithms he is the author of more

automated trading software. At present he is the Principal Software Developer of Algo Trading

Service Ltd. and he is involved as PhD student in Economic Informatics Doctoral School of the

Academy of Economic Studies.