Cryptocurrency Trading Program Using Machine Learning

by

Dante Knight, Jack Gerulskis, Sean Dandeneau

A Major Qualifying Project

Submitted to the Faculty

of the

WORCESTER POLYTECHNIC INSTITUTE

In partial fulfillment of the requirements for the

Degree of Bachelor of Science

in Computer Science

by

May 2021

APPROVED:

Smith, Therese
 Abstract

Cryptocurrency is a digital asset that has been historically volatile. This volatil-
ity allows traders to capitalize on short term price movement. More specifically,
cryptocurrency is vulnerable to “epidemic-like price bubbles” from social media fac-
tors compared to traditional assets (Phillips, 2018)[11]. Social media’s influence
on the price of cryptocurrency gives traders a unique opportunity for predicting
price movements. Traditionally, traders have used technical analysis to predict the
best opportunities to buy and sell, but sentiment analysis of posts on social media
can help improve their accuracy. While humans are capable of manually conduct-
ing technical analysis, it is near impossible for them to understand the trends and
consensus of an asset from posts on social media. A computer can conduct both
technical and sentiment analysis more efficiently and use them as indicators to make
accurate predictions on future price movements. The goal of this project is to create
an automated trading program that uses technical and sentiment analysis as inputs
for a machine learning model which can predict profitable opportunities to buy and
sell cryptocurrency.
Contents

1 Introduction 1

2 Background 4
2.0.1 Technical Analysis Trading Strategies . . . . . . . . . . . . . . 4
2.0.2 Machine Learning Trading Strategies . . . . . . . . . . . . . . 6
2.0.3 Comparing Trading Strategies . . . . . . . . . . . . . . . . . . 7
2.0.4 Comparing Tradable Assets . . . . . . . . . . . . . . . . . . . 9
2.0.5 Cryptocurrency Markets vs Traditional Markets . . . . . . . . 11
2.0.6 Trading Algorithms . . . . . . . . . . . . . . . . . . . . . . . . 13
2.0.7 Sentiment Analysis . . . . . . . . . . . . . . . . . . . . . . . . 17

3 Methodology 20
3.0.1 Collect Historical Sentiment Data . . . . . . . . . . . . . . . . 20
3.0.2 Collect Historical Price Data . . . . . . . . . . . . . . . . . . . 20
3.0.3 Train Child Models . . . . . . . . . . . . . . . . . . . . . . . . 21
3.0.4 Train Parent Model Using Child Models . . . . . . . . . . . . 22
3.0.5 Stream Sentiment Data . . . . . . . . . . . . . . . . . . . . . . 23
3.0.6 Automate Trading using Parent Model . . . . . . . . . . . . . 23

4 Results and Conclusions 27

5 Limitations and Future Work 30

Appendices 34

i
List of Figures

2.1 RSI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.2 Moving Average . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.3 Human Designed Strategy . . . . . . . . . . . . . . . . . . . . . . . . 8
2.4 Machine Learning Strategy . . . . . . . . . . . . . . . . . . . . . . . . 10
2.5 Bitcoin Price Graph . . . . . . . . . . . . . . . . . . . . . . . . . . . 11
2.6 XRP Price Graph . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.7 US Equities Order Flow . . . . . . . . . . . . . . . . . . . . . . . . . 13
2.8 Cryptocurrency Order Flow . . . . . . . . . . . . . . . . . . . . . . . 14
2.9 Bitcoin’s 5 day moving average from 2015-2020. . . . . . . . . . . . . 15

3.1 Share of 8 popular exchanges by volume . . . . . . . . . . . . . . . . 25

3.2 Share of 8 popular exchanges by volume . . . . . . . . . . . . . . . . 26
3.3 Taker fees for Kraken and Coinbase at different monthly volumes . . 26

4.1 Fit of MultiLayer perceptron during testing, 24 hour time frame . . . 29

4.2 Fit of MultiLayer perceptron during testing, 2 hour time frame . . . . 29

ii
Chapter 1

Introduction

Machine learning trading algorithms have had success in the stock markets for
some time now (Chang, Lui, 2009)[2]. More recently, they also have outperformed
traditional buy-and-hold strategies for cryptocurrencies (Jiang, Liang, 2017)[10].
The successes of machine learning models trading cryptocurrency has inspired more
research in the field. Prior research in the field has shown sentiment analysis from so-
cial media sites is a useful indicator for machine learning models. (Phillips, 2018)[11].
Other researchers have successfully created multiple profitable models using techni-
cal analysis. This report intends to build on the previous research by finding new
ways to conduct meaningful sentiment analysis and also by creating an ensemble
model to increase accuracy and effectiveness.
Different artificial neural networks have shown to outperform traditional statis-
tical models in 72% of cases, but the highly volatile nature of the cryptocurrency
market makes it difficult for a single model to be effective in every instance (Chang
et al. 2017)[3]. In the other 28% of situations, statistical models outperform ma-
chine learning (Chang et al. 2017)[3]. The other challenge with creating models
that uses sentiment analysis is accurately converting posts on social media to data
the models can understand. A few of the challenges with this conversion is under-
standing slang, sarcasm, and irony. Selecting an optimal model depending on the
volatility of the market and properly analysing sentiment will be critical to creating
an effective trading program.

1
 As mentioned, prior research on trading cryptocurrency using machine learning
has already been done. Hegazy and Mumford created a model using a supervised
learning strategy that had a 57% accuracy in predicting price fluctuations (Hegazy,
Mumford, 2016). Jiang and Liang used deep reinforcement learning to increase their
initial investment by 1000% (Jiang, Liang, 2017)[10]. Shah and Zhang achieved
a 200% return on their inital investment in 2 months by using Bayesian regres-
sion(Shah, Zhang, 2014)[13]. Fischer and others created a profitable model using an
arbitrage approach (Fischer et al, 2019)[4]. Stenqvist and Lonno used deep learning
algorithms to analyze 2.27 million tweets to predict Bitcoin price fluctuations with
a 79% accuracy (Stenqvist, Lonno, 2017)[14].
In order to differentiate ourselves from the work that has already been done,
we have identified two gaps in the research that we will attempt to fill. First, we
will implement an ensemble learning model that can operate accurately regardless
of the volatility. This will improve the reliability of the system because the outputs
of the models in the ensemble system will reduce the effect of bad decisions from
varying volatility. It is important to note that the best model in the group might
still outperform the ensemble in the short term, but the ensemble will outperform
the individual models in the long term (Fumera, Roli, 2005)[5]. The second gap is
using the impressions metric when converting posts from social media to machine
readable data. This should enable the model to differentiate between posts that
have lots of interactions with people compared to tweets that nobody interacts
with. This is something that we didn’t see in our prior research and will help
identify posts that have a stronger influence. By implementing these solutions to
on top of successful models that have already been created, we should be able to
enhance their performance.
The goal of this project is to create an automated trading program that uses

2
technical and sentiment analysis as inputs for a machine learning model which can
predict profitable opportunities to buy and sell cryptocurrency. Our objectives
to achieve this goal were to collect data from Twitter, conduct accurate sentiment
analysis, create profitable individual models, and lastly to create an ensemble model.
This report begins with an overview of cryptocurrency trading strategies and
machine learning trading. We then discuss sentiment analysis and the key techniques
involved with properly performing this analysis. Following the background, our
report details how we gathered our data and created our ensemble model. We then
finish by discussing the results and implications of our study. This research benefits
cryptocurrency traders and prior studies, who can use our research to improve their
trading decisions.

3
Chapter 2

Background

2.0.1 Technical Analysis Trading Strategies

The most prevelant analysis for cryptocurrency trading is Technical Analysis.

Technical Analysis is, ”the study, practice, and analysis of chart patterns, indicators
and oscillators, and the candlesticks themselves that make up price charts of as-
sets (https://primexbt.com/blog/cryptocurrency-trading-strategies/)”. This form
of analysis will identify factors such as Relative Strength Index (RSI) and mov-
ing averages that can give a trader an edge in the market. Understanding these
strategies is critical before using them as indicators for our model.
RSI is the most common technical cryptocurrency trading strategy. RSI identifies
when an asset is being over or under valued and bought in the market. It watches
the price movement of a tradable asset set to a scale of 0-100, with two triggers at an
upper and lower bound. As a baseline, these triggers are set at 70 and 30. When the
stock reaches or goes above 70, RSI tells us that this stock is being overbought and
overvalued, and history tells us this stock will revert back towards a lower price. On
the other hand, when a stock drops below 30, it is being oversold and undervalued
with a high likelihood of trending back to the mean of 50. An example of RSI is
depicted in the figure below.
The moving average trading strategy identifies the moving average of the stock
and, like RSI, has two triggers to identify when to buy or sell. Shown in the graph
below, when the price crosses the moving average line it tells us when to buy or short

4
 Figure 2.1: RSI

the stock. Moving averages are help identify trends by comparing the averages from
a long and short time frame. Time frames for moving averages commonly follow the
Fibonacci series. For example a short term moving average may be 5 days and that
would be compared with a longer moving average like 13 days.

Figure 2.2: Moving Average

5
2.0.2 Machine Learning Trading Strategies

To help remove the human element out of trading, we look towards machine
learning strategies. These strategies remove human emotion and bias to make a
calculated decision. There are numerous different machine learning trading methods,
so we will just cover the top three.
The first is graph neural networks. Graph neural networks (GNNs) are
a deep learning discipline that focuses on models that operate efficiently on
graph data structures (https://www.coindesk.com/five-machine-learning-methods-
crypto-traders-should-know-about). Though this is a relatively new area of deep
learning, it is widely used for numerous applications in companies like Uber,
Google, Microsoft and DeepMind. GNNs use ”a graph as input representing
the flows in and out of exchanges and infer relevant knowledge relevant to its
impact on price (https://www.coindesk.com/five-machine-learning-methods-crypto-
traders-should-know-about).” GNNs open the potential for new quant methods
based on blockchain datasets.
The second machine learning strategy is generative models. Generative models
look at historical data to develop their own synthetic data that closer mirrors the
distribution of a training data set. When we combine the real dataset with the
synthetic one that was generated, we then have a large enough dataset to train a
complex deep learning model. One technique used in generative models is generative
adversarial neural networks (GANs). Although not directly related to cryptocur-
rency trading, GANs have proven very successful in image classification. The ability
to memorize previous trends and identify them happening in real time separate ma-
chine learning trading strategies from technical strategies.
Semi-supervised learning is the last machine learning trading strategy we will talk

6
about. Semi-supervised learning focuses on creating models that learning with small
labeled data sets and a large amount of unlabeled data. ”Semi-supervised learning is
analogous to a teacher presenting a few concepts to a group of students and leaves
the other concepts to homework and self-study (https://www.coindesk.com/five-
machine-learning-methods-crypto-traders-should-know-about).” A semi-supervised
learning model will identify important features in the labeled data and use that
information to incorporate the unlabeled data into the training.
Machine learning trading strategies get much more complex that technical strate-
gies but can be much more accurate with far less risk. All of these models will be
able to automatically calculate and interpret technical indicators mentioned in the
previous section to make predictions.

2.0.3 Comparing Trading Strategies

As discussed above, there are two groups of trading strategies that we will be
looking at: human designed and machine learned. The first group is the type of
strategy that can be understood well because they are the creation of a human
who designed it with a certain theory in mind. A typical human designed trading
strategy will issue orders in cases like when a certain price point is reached or a
technical indicator crosses some value or some combination of one or more of both.
One can imagine a strategy which uses RSI, a common technical indicator which
indicates positive and negative momentum on a price series from 0-100 with 100
being the highest positive momentum and 0 being the highest negative momentum.
A common way to use RSI is to have two thresholds near the extremes of the
spectrum like 30 and 70 or 18 and 86. When RSI crosses the chosen threshold in
this strategy, at the higher threshold it would sell the asset and at the lower it
would buy. This is an attempt to buy when the price is moving most down and

7
sell when price is moving most up. A more complex example of this can be seen in
the table below which has columns for buy and sell signals which are determined by
the functions shown in each row of the two columns. The strategy described can be
understood and rationalized by the people who implement it either through trading
manually or by automatically.

Figure 2.3: Human Designed Strategy

The same cannot be said for trading strategies that have been learned using
machine learning. While it may be possible to guess how a model with a simple
architecture works, as the complexity increases, the guess becomes more rough and
less reflective of the actual model. This isn’t to say that machine learning isn’t

8
desirable for a trading strategy because it certainly is. Only that one’s understanding
of the connection between the input and the output becomes less accurate as the
model becomes more complex. The example in the table below is set up similar to
the human designed strategy above but the buy and sell signal columns are missing
because the nice functions that we had before can’t be defined as easily now (Anghel,
2020)[1]. The equivalent to the buy and sell signals in this example would be the
last dense layer in each of the networks that outputs only one value at the end which
would be interpreted as buy sell or hold. In the example before, if there was a buy
signal, we would be able to point to the RSI being below the threshold and say with
certainty that that is what caused the buy signal unlike with the machine learning
example.

2.0.4 Comparing Tradable Assets

The correlation between price of bitcoin and the many other cryptocurrencies has
been studied in the past and has increased since bitcoin’s all time high price near the
end of 2017. This is an important consideration because it has implications on the
prospect of diversifying an all cryptocurrency portfolio, which could be a strategy
we consider. The findings tell us that all of the cryptocurrencies prices are corre-
lated which makes diversification less effective at spreading out risk over multiple
independent investments (Ferreira and Pereira, 2019)[ferreira˙contagion˙nodate].
This is because when the assets are correlated, investing in two different assets in-
stead of one isn’t going to shield you from the price fluctuation of one because a
similar fluctuation will most likely be occurring in the other asset. An example can
be seen below showing the last year of prices of bitcoin and ripple (XRP).
The figures have green and red boxes signifying upwards and downwards price
trends respectively. The figure showing the price of XRP has additional grey vertical

9
 Figure 2.4: Machine Learning Strategy

lines which represent the starts and ends of the green and red boxes for bitcoin. It
can be seen by the eye that these assets have some serious price correlation but
that being said not everything is the same. Another thing to note is that certain
cryptocurrencies seem to have higher volatility meaning that they are prone to
fluctuations of greater percentages than bitcoin. This can be seen in the figures
above. The percentage changes for the boxes in the XRP graph are generally greater
than those of bitcoin. This is an important consideration because it can help us to

10
 Figure 2.5: Bitcoin Price Graph

gauge how much risk a strategy is taking on when it invests in each cryptocurrency
and knowing that allows the strategy to regulate its own risk. This can most likely
be attributed to the lower market cap of XRP in comparison to Bitcoin. In general
as market cap increases, price fluctuations become lower in percentage change.

2.0.5 Cryptocurrency Markets vs Traditional Markets

A common example of a traditional market would be US equities. In this market,

brokers use to make their money by charging large fees. Recently, brokers like
Robinhood have popularized the concept of $0 commission trades. Now most big
firms offer the same deals to investors. The dropping fees forced brokers to find
more ways to generate revenue. Brokers often sell orders that don’t specify a specific
exchange to execute the trade on to market makers. The market makers are given
the opportunity to trade with individual investors the lack the advantages high

11
 Figure 2.6: XRP Price Graph

frequency trading firms have. This creates an issue for individual traders trying
to profitable trade intra-daily because most of the time they will be trading with
market makers or the high frequency trading firms on the exchanges. Market makers
aren’t that interested in institutional investors because they have a better advantage
than individuals. These markets are highly efficient, have low spreads in the orders,
and news propagates instantly.
In cryptocurrency markets, the biggest difference is that there are no longer bro-
kers. Instead of brokers managing orders, the investors handle the orders manage-
ment themselves in a peer to peer fashion. The settlement is then handled without
intervention from the National Securities Clearing Corporation and is handled by
the exchange. If the investor is trading on a decentralized exchange the trade is then
conducted smart contracts. Unlike centralized exchanges where buy and sell orders
are paired by an order book, decentralized exchanges operate by matching the in-

12
 Figure 2.7: US Equities Order Flow

dividuals that want to execute the buy and sell orders and allow them to exchange
their assets. By removing the middleman between the investor and the exchanges,
the high frequency trading firms loose their critical advantage between directed and
non directed orders. Decentralized exchanges make it even harder for high frequency
traders because they operate using smart contracts which lowers transaction speeds
and have higher fees. This is constantly changing as the technology improves but for
now those factors and the lower liquidity result in less high frequency trading being
done on decentralized exchanges. Contrary to traditional markets, the markets are
more inefficient have wider spreads and slower news propagation.
Another common difference the two markets volatility. As seen in Figure 2.9,
the 5 day moving for Bitcoin’s variance is high. (Correlate this to twitter activity?)

2.0.6 Trading Algorithms

Machine learning trading algorithms have been shown to have success trading
in traditional stock markets for some time now (Chang, Lui, 2009)[2]. Even with

13
 Figure 2.8: Cryptocurrency Order Flow

the recent emergence of various cryptocurrencies, machine learning has also shown
to outperform traditional buy and hold strategies over a one year period with 12
cryptocurrencies (Jiang, Liang, 2017)[10]. One of the biggest differences cryptocur-
rencies compared with traditional stock markets is the short term volatility. This
volatility has been partially attributed to online factors. The strongest correlation
between an online factor and the price of a cryptocurrency was the polarization of
opinions on Twitter about the digital asset (Phillips, 2018)[11]. The polarization of
opinions on Twitter often indicated that positive price action for the cryptocurrency
in the medium-term while other factors like security breaches had a greater impact
on the price in the short term (Phillips, 2018)[11]. In the medium term, opinions
seem to have the greatest impact on price action. In the short term, facts, or news,
have the greatest impact on the price action. Ideally, an algorithm using these fac-
tors after completing a sentiment analysis would have an advantage over algorithms
that do not have any understanding of public opinion.
“The relationships link online activity increases to price falls (the converse is not
observed). It is not surprising that occasionally discussion is associated with price
falls, as negative events (e.g. blockchain bugs, and exchange hacks) are newsworthy
in the community.” (Phillips, 2018)[11] A long term indicator for online activity that

14
 Figure 2.9: Bitcoin’s 5 day moving average from 2015-2020.

has proven to be positively correlated with price movement is activity on Reddit.

[Aside: this article proposes that this long term factor is because as there is more
technological progress, there is more discussion and news related to it in these fo-
rums. Additionally the author suggests a future project looks at the github pages
because the project are all open source. Could be a cool new factor] There are
also periods in the medium to long time range that create bubbles from a positive
feedback loop where an increase in discussion online about a cryptocurrency results
to more usage and higher prices which in turn creates more discussion about the
cryptocurrency (Phillips, 2018)[11]. This often creates a bubble because the price
is exceeding the actual value of the token from the positive feedback loop.
Between 7-10 days an amount of positive user replies were successful in predicting
price fluctuations (Kim et al, 2016)[7]. This report found that smaller communities
had less accuracy in predicting price fluctuations. Since the time the report was

15
published, 2016, there are many more active communities with many daily active
users.
Figure 2.1 shows the relation between price movement of the cryptocurrency
PinkCoin and number of tweets mentioning PinkCoin. PinkCoin has a small market
cap so the sample size could be considered small, but from the graph it is clear that
a huge spike in the number of tweets about the cryptocurrency preceded jumps in
price which was commonly followed by a dramatic drop. This is probably due to
the bubble like behavior described by Phillips.
Our analysis reveals that increases in opinion polarization and exchange volume
precede rising Bitcoin prices, and that emotional valence precedes opinion polariza-
tion and rising exchange volumes (Phillips, 2018)[11]. Emotional valence refers to
the negativity or positivity of a comment. Opinion polarizations refers to conflicting
viewpoints about a topic.
A group researching cryptocurrency trading using machine learning got positive
results from a model that they trained to trade by letting it control trading decisions
on a sliding scale from -1 to 1. -1 represents completely short which means opening
a short trade worth 100% of the accounts value. This means that if the price drops
10%, the trade will be worth 1% of the starting value. 1 means the same but for the
long side, a price increase of 1% means the trade value increases by 1%. By using a
scheme like this, the model is able to learn how to risk different amounts in different
scenarios instead of either relying on a traditional method of risk calculation or by
simply using only the cases -1 or 1. This group trained their model with a reward
system that calculated reward considering the length of a trade. For example it
would be a better idea to take two trades that give 1.5% in 1 day each rather than
take one trade that returns 2% in two days. By dividing their reward by time they
effectively have given the model a way to interpret opportunity cost. This means

16
that the model will be incentivized to choose time effective trades instead of purely
basing decisions on predicted profit or loss (Koker and Koutmos, 2020)[8]. In a
similar fashion, another group made a trading model that was different than all the
others so far. They used a portfolio management system instead of just trading
one asset pair. The model would output percentages for each asset to be traded
including cash. This way the model would output allocation specifications and the
group simulated what the account value would be after executing trades to reach
the models allocation numbers. Like the other group, they also had a time element
to their reward but in this case they also divided the profit by the initial investment
as well to give ROI over time (Laura et al, 2018)[9].
Machine learning algorithms face a new set of challenges with cryptocurrency
that are not present in traditional exchanges. Machine learning algorithms trading
on traditional exchanges have access to many macroeconomic variables that have
been strong indicators of price movement (Koker and Koutmos, 2020)[8]. However,
cryptocurrency does not have strong macroeconomic variables.
As an alternative to machine learning, technical analysis has also proven to
be easier to implement and still effective (Anghel, 2020)[1]. In some scenarios,
certain inputs for machine learning algorithms actually underperform traditional
technical analysis (Anghel, 2020)[1]. Many current reports on effective machine
learning algorithms are victims of data dredging (Anghel, 2020)[1]. Data dredging is
misusing data analysis to find patterns and present results as statistically significant.

2.0.7 Sentiment Analysis

Sentiment Analysis has proven to be a useful indicator when determining price

fluctuations of an asset in the past. One of the earlier studies to do this was a
study that tried to predict cryptocurrency price fluctuations based on user com-

17
ments and replies. They collected data from multiple online sources where users
can post text and placed them into one of five categories: Very negative, negative,
neutral, positive, very positive (Kim et al, 2016)[7]. They were able to predict fluc-
tuations with 79% accuracy using a 7 day lag. Another study using deep learning
algorithms achieved also a 79% accuracy in predicting price fluctuations of Bitcoin
by conducting similar sentiment analysis on over 2 million tweets (Stenqvist, Lonno,
2017)[14]. They found the sentiment analysis to be most effective when splitting
the data into 30 minute time blocks. There was also varying level of success by
pairing the sentiment analysis to price data from up to 90 minutes later. The study
definitively concluded that sentiment analysis of relevant tweets is a useful price
predictor. A later study achieved 90% accuracy in predicting price fluctuations by
conducting sentiment analysis on popular social media platforms about a specific
asset (Colianni et al, 2018)[12].
As the cryptocurrency ecosystem becomes more popular, the studies are finding
more and more success with less lag. Additionally, the volume of data to be analyzed
has increased. Polling for tweets about any cryptocurrency will return a collection of
recent tweets about the topic. The mechanisms for collecting the data and analyzing
it have changed, but currently one of the most prevalent techniques is using Twitter’s
APIs along with a sentiment analyzing technique called VADER (Hutto, 2015)[6].
VADER, or Valence Aware Dictionary and sEntiment Reasoner, is a rules based
module that can accurately interpret sentiment for social media type content. It is
capable of interpreting slang, emojis, and hashtags. It understands negations such
as ”not good” and ”wasn’t that good”. It properly analyzes punctuations and all
caps. It understands intensity modifiers like ”very” and ”kind of”. A high level
overview of this process is as follows.
VADER is easy to use and returns a few metrics about the sentence it analyzed.

18
It scores the positivity, neutrality, and negativity. It also has a metric called com-
pound that is the sum of the other three metrics. The compound score ranges from
-1 to 1 with -1 being very negative and 1 being very positive.

19
Chapter 3

Methodology

3.0.1 Collect Historical Sentiment Data

Before we can perform sentiment analysis we need content to analyze. We chose

Twitter as our source because we can collect a large number of data points within
a small period. We set up a script to run for 6 months and gather tweets every 60
seconds. The script was written in python and utilizes the Twitter API. Our script
maximizes the API’s limits. As a result, our script queried 16,665 tweets a day,
694 an hour, and 11 a minute. By limiting our query to 11 tweets every minute we
ensured a more even distribution of data.
The endpoint (https://api.twitter.com/2/tweets/search/recent) allows
for the request to specify relevant search terms to be returned. Also, the request can
include the tweet’s timestamp, likes, retweets, replies, language, and quotes. The
query also gets the followers of the original poster. Lastly, all the tweets collected
were in English.

3.0.2 Collect Historical Price Data

In order to train our trading models we first need to figure out what types of
data are useful for predicting trading returns. Some of this work was already done
by researching past attempts at trading using machine learning. Since it’s almost
impossible to tell if a type of data will be useful for predicting trading returns, we

20
will most likely include some data into our set that we don’t end up using for our
final model. The types of data that we will consider include: price data, technical
indicators and user text entries about cryptocurrency scraped from various sources.
Using just price data and technical indicators would be a traditional way of training
a model like ours but we hope to add some utility with the user text data being how
we can gauge market sentiment.
To get our price data, we downloaded the historical candle data as a csv for
multiple different crypto trading pairs in USD. This gives us a series of candles with
their times. One candle consists of 4 prices: the open, high, low, and close price,
where the open price is the first price that a trade is made at during that candle
period whereas the close price is the last price traded at during the candle period.
This data could serve as input data to the model as well as be used to calculate
returns.
Technical indicators are certain values derived from things like price and volume
data. We will use existing technical indicators as well as possibly creating new ones
to use for our model. Since a technical indicator is calculated by using existing price
info, it wouldn’t be too hard to calculate custom indicators if we ended up using
them. The advantages to using technical indicators over price data is that since
many technical indicators oscillate from one region of values to another, it is easier
to map the indicator output to a range of 0 to 1 with proper distribution which
makes training models easier.

3.0.3 Train Child Models

The next step is to train our models. We want to train our models with different
sets and combinations of data so we can learn what data is more able to be used
for our problem. It might be the case that certain trading pairs and timeframes

21
might do better with models that use different input data. For example if you used
a daily time frame trading strategy to trade on the 5 minute timeframe, you might
not do well. It turns out that different data is sometimes better suited for a specific
timeframe or trading pair.
The child models can be any assortment of traditional trading strategies to a
trained neural network. The idea is that we find methods of trading that work well
either in general or under certain circumstances. For example, traders consider the
200 day moving average of an asset to be a line in the sand between being “bearish”
or “bullish”. Bearish and bullish are terms used to describe if price is generally
trending down or up respectively. Taking this in mind, we can figure that strategies
that take more longs than shorts above the 200 day moving average would in general
perform better than a similar strategy which took more shorts than longs.
We will be using the programming language Python because it has many modules
that make data collection and manipulation as well as machine learning much easier
and faster to do. We hope that using tools like Keras and TensorFlow will help
us to quickly create machine learning architectures without having to worry about
implementing them ourselves.
If possible it would be extremely useful to be able to make a trading pair agnostic
model. This would allow us to train the model using data from multiple trading
pairs instead of just keeping it to one pair to a model. This would help because it
would allow us to use the same model and strategy for all or most of the trading
pairs we are looking at while vastly increasing the amount of test data available.

3.0.4 Train Parent Model Using Child Models

Once the lower models are trained, we can now take the output of those models
and feed it to a new model along with their respective effectivenesses. This new

22
model should then be able to interpret the input and come up with a trading decision.
This step may not be feasible or even practical depending on the type of output we
get from our previous models and whether or not we are able to develop asset
agnostic models.

3.0.5 Stream Sentiment Data

The model will be making decisions in real time, therefor the historical tweets
API previously used is not sufficient. Instead the model will be directly interacting
with Twitter’s streaming endpoint that opens a connection and delivers content that
matches a specific rule set as their tweeted. After specifying a timeframe, the script
gathers and analyzes the tweets. It finally finds the average normalized, weighted
composite sentiment and variance at the end of the time frame. This will become a
single input data point for the machine learning algorithm to use.
As the project goes on tailoring the rule set will be an important part of opti-
mizing our model. The current configuration retrieves all tweets containing ”Eth”,
”Ethereum”, ”ETH”, or ”$ETH”. There’s two rules that use those keywords. The
first returns tweets that have been annotated by Twitter’s algorithm as content
that is referring to politics. The second just returns any tweet with the keyword.
Tweets gathered that match the first rule will be given a slightly higher weight when
calculating overall sentiment because they are generally more significant tweets.

3.0.6 Automate Trading using Parent Model

We will most likely employ the simple method of having the trading states of
completely short and long mapped from -1 to 1 respectively. The benefit of doing it
like this would be that our model doesn’t have to output any sort of price, it only
needs to output a number from -1 to 1 to signal to our trading algorithm to market

23
sell or buy depending on the change in output between the two steps. For example
if the model starts at an output of 1 but drops to 0 we know that we need to exit the
long position by selling the same amount used to open the trade at market price.
These trades would all be executed on margin so that we don’t need to worry
about actually buying and selling the assets in question. The way this works is that
if we had 1000 dollars on an exchange and we want to short BTC, if we weren’t
using margin we wouldn’t be able to enter the trade because in order to enter a
short trade, you must sell the asset which at the time was not in the account. By
using margin, the trader is able to borrow the asset of the exchange to sell based
on the amount of capital in their account. At some exchanges, the amount lent can
be up to 100 times the capital in the account. Obviously using leverage like that is
dangerous so we will only use margin to borrow a value that matches the account
value simply for the utility of being able to both short and long with just cash.
The exchange we will be using to trade is Kraken. We will be using this exchange
because it is one of the highest volume exchanges that services American citizens.
There are many other exchanges that might have been better but since they aren’t
legally allowed to work with American citizens, we would be risking being kicked
off the platform as well as possible loss of control of funds. For these reasons, it
is not worth it to use an exchange like Binance, even though it would be better in
every way. The one other clear option would be Coinbase Pro because it has higher
volume than Kraken. As can be seen from the pie chart in Figure 4.1 below, Kraken
and Coinbase Pro make up less than five percent of the volume of the 8 popular
global exchanges included in the chart.
This is bad because it means that the exchange is more likely to be manipulated
as the order books are thinner meaning that price can be moved the same amount
for much less on low volume exchanges. Despite being a low volume exchange when

24
 Figure 3.1: Share of 8 popular exchanges by volume

compared to global exchanges, Coinbase Pro and Kraken both take up a majority
share of the volume from exchanges that are US friendly, which can be seen in Figure
4.2. The problem with Coinbase in particular is that its trading fees are higher than
Kraken’s for nearly all volumes of trading except after $50 million in 30 day trading
volume as shown in Figure 4.3 below. Since we will almost certainly be trading
under $50 million volume each month, Kraken is the superior choice in terms of
fees. This is important especially for strategies which trade for small movements as
well as strategies which make a higher number of trades than normal.

25
 Figure 3.2: Share of 8 popular exchanges by volume

Figure 3.3: Taker fees for Kraken and Coinbase at different monthly volumes

26
Chapter 4

Results and Conclusions

At the conclusion of our development we were able to create a model that could
trade profitably. However, getting to that point took a lot of refinement. One of
the biggest challenges was first getting our sentiment analysis to have a meaningful
correlation with Ethereum price movement. From the start in October to the end
of the collection in January, we were able to collect around 500,000 tweets that
included the keyword ”Ethereum”, ”Eth”, ”$Eth”, or ”Eth”. Our background ex-
plored grouping the data into various time frames and delays. One of the most
common time frames was 2 hours with a 30 min delay. Our data unfortunately
had too many gaps at that interval. Our data collection was done on a embedded
system that was not connected to a monitor, so when unforeseen events like WiFi
interruption or power outages occurred, it took human intervention to reboot the
collection script. Trying to fill in the remaining data was difficult and dropping
rows with missing data was not an option. It was critical for the model to be able
to see all of the price data to accurately understand the trends from our technical
indicators. Our best option was to modify the time frames. The smallest feasible
time frame was 13 hours but we opted for a 24 hour time frame because it had the
strongest correlation with our data. Even though the market is open 24 hours, the
tweets we collected mostly originated from the US. The majority of trades from the
original posters of the tweets are conducted between 7am to 8pm. This reasoning is
most likely why this time frame was the most useful for our model. Our algorithm

27
was then refined to look at the previous days sentiment and technical analysis and
make a decision at 7am EST to either hold a position, buy, or sell an active po-
sition. While the algorithm may miss out on intraday opportunities, it was much
more reliable at the 24 hour time frame. We implemented a 7% stop loss to prevent
too much damage from incorrect decisions.
Our best model was a Multilayer Perceptron but we also included the three other
models discussed in the background into the ensemble. It achieved an r-squared
value of 0.37 during testing and continually improved during the testing phase as
more data from twitter was gathered in real time. Although the model ultimately
was making a classification decision as to if the price would increase or decrease over
the next time frame, regression was more useful. The algorithm predicted the %
increase or decrease for the next day and if it was within a 90% confidence interval
it would make a decision, otherwise it would ignore or close the current position if
it was not confident enough.
Figure 4.1 was the final result of our model but to further illustrate the issues
we had with trying to fit the model with a 2 hour time a figure of the fit using that
time frame is in figure 4.2. The model had a r-squared of -0.12. Obviously, this was
not sufficient to trade profitably.
Regardless of the time frames, we were able to successfully extract and analyze
data, create meaningful technical indicators, and finally make a profitable automated
trading model. The model will benefit from additional testing over a longer period
to understand its sustainability. The model is currently still running and additional
testing is being done.

28
Figure 4.1: Fit of MultiLayer perceptron during testing, 24 hour time frame

Figure 4.2: Fit of MultiLayer perceptron during testing, 2 hour time frame

29
Chapter 5

Limitations and Future Work

Like every other project, we were limited on both the scope and time available
for our project. We initially thought we would be able to trade based on our model
live for longer but due to time limitations we haven’t been able to trade live for long
enough to be fully confident in its success yet. This is something we will continue
to do in the future and is certainly the next step for this project.
One improvement to make could be further developing the algorithm for taking
the output from the model and turning that into trading decisions that consistently
profit. This work is the final bridge between the model and making profitable trades
but hopefully that will become easier as our model improves over time.
Another consideration is making sure that the system that makes live trades has
a notification system of some kind to make sure that any technical problems like lost
power or internet are solved as soon as possible. We could also switch the hosting
of our live trading to a hosting service instead, which comes with its own problems,
but it could be worth it. This test will give further proof that our system works
even with real world problems like human error or technical problems. We focused
on getting our model to successfully trade ETH mostly due to time concerns.
A cryptocurrency we were looking at, XRP, had its trading ability shut down
on all but a few small exchanges in the US, which made it nearly impossible to
trade live, making it not worth it to pursue. In the future, XRP trading should be
available again in the US and we can trade and train our model on it and other

30
cryptocurrencies.

31
Bibliography

[1] Dan-Gabriel Anghel. “A reality check on trading rule performance in the cryp-
tocurrency market: Machine learning vs. technical analysis”. In: Finance Re-
search Letters (June 20, 2020), p. 101655. issn: 1544-6123. doi: 10.1016/
j . frl . 2020 . 101655. url: http : / / www . sciencedirect . com / science /
article/pii/S1544612320304414 (visited on 09/14/2020).

[2] P Chang et al. An Ensemble of Neural Networks for Stock Trading Decision
Making. Berlin: Springer Berlin Heidelberg, 2009, pp. 1–10.

[3] Yung-Ho Chang, Chia-Ching Jong, and Sin-Chong Wang. “Size, trading vol-
ume, and the profitability of technical trading”. In: International Journal of
Managerial Finance Vol.13.4 (2017), pp. 475–494.

[4] Thomas Fischer, Christopher Krauss, and Alexander Deinert. “Statistical Ar-
bitrage in Cryptocurrency Markets”. In: Journal of risk and financial man-
agement 12.1 (Feb. 13, 2019), p. 31.

[5] G Fumera and F Roli. “A theoretical and experimental analysis of linear com-
biners for multiple classifier systems”. In: IEEE transactions on pattern anal-
ysis and machine intelligence 27.6 (June 2005), pp. 942–956.

[6] CJ Hutto. VADER: A Parsimonious Rule-based Model for Sentiment Analysis

of Social Media Text. Jan. 2015. url: https : / / www . researchgate . net /
publication/275828927_VADER_A_Parsimonious_Rule-based_Model_for_
Sentiment_Analysis_of_Social_Media_Text (visited on 09/14/2020).

[7] Young Bin Kim et al. “Predicting Fluctuations in Cryptocurrency Transac-

tions Based on User Comments and Replies”. In: PLOS ONE 11.8 (Aug. 17,

32
 2016), e0161197. issn: 1932-6203. doi: 10 . 1371 / journal . pone . 0161197.
url: https : / / journals . plos . org / plosone / article ? id = 10 . 1371 /
journal.pone.0161197 (visited on 09/14/2020).

[8] Thomas Koker and Dimitrios Koutmos. “Cryptocurrency Trading Using Ma-
chine Learning”. In: 13.8 (2020). doi: DOI:10.3390/jrfm13080178.

[9] Alessandretti Laura et al. “Anticipating cryptocurrency prices using machine

learning”. In: Alessandretti, Laura; ElBahrawy, Abeer; Aiello, Luca Maria;
Baronchelli, Andrea (Nov. 9, 2018). doi: DOI:10.1155/2018/8983590.

[10] J Liang and Z Jiang. Cryptocurrency portfolio management with deep rein-
forcement learning. London, Sept. 2017, pp. 905–913.

[11] RC Phillips and D Gorse. Cryptocurrency price drivers: Wavelet coherence

analysis revisited. 2018.

[12] Otabek Sattarov et al. “Recommending Cryptocurrency Trading Points with

Deep Reinforcement Learning Approach”. In: Applied Sciences 10.4 (Jan.
2020), p. 1506. doi: 10.3390/app10041506. url: https://www.mdpi.com/
2076-3417/10/4/1506 (visited on 09/25/2020).

[13] Devavrat Shah and Kang Zhang. “Bayesian regression and Bitcoin”. In:
(Oct. 6, 2014). url: https://arxiv.org/abs/1410.1231.

[14] Evita Stenqvist and Jacob Lönnö. Predicting Bitcoin price fluctuation with
Twitter sentiment analysis. 2017. url: http://urn.kb.se/resolve?urn=
urn:nbn:se:kth:diva-209191 (visited on 09/25/2020).

33
Appendices

34