=Paper=
{{Paper
|id=Vol-2105/10000453
|storemode=property
|title=None
|pdfUrl=https://ceur-ws.org/Vol-2105/10000453.pdf
|volume=Vol-2105
}}
==None==
https://ceur-ws.org/Vol-2105/10000453.pdf
Cryptocurrencies Prices Forecasting With Anaconda
Tool Using Machine Learning Techniques
Oleksandr Snihovyi1, Oleksii Ivanov1 and Vitaliy Kobets1[0000-0002-4386-4103]
1 Kherson State University, 27, Universitetska st., Kherson, 73000, Ukraine
snegovoy@hotmail.com, sink2385@gmail.com, vkobets@kse.org.ua
Abstract. The research goal is to study criteria which affect the price of a cryp-
tocurrency and their usage for the price forecasting using Machine Learning
techniques. The subject of research is forecasting of most famous cryptocurren-
cy Bitcoin using the most popular Python Data Science Platform – Anaconda.
Research Methods are machine learning, data analysis, and multiple regression.
A set of criteria which affect the price of the cryptocurrency were defined based
on the analysis of public information. Their correctness was tested using Ma-
chine Learning algorithms. As a result of simulation experiment through the
application using real data from open sources, we have found that selected
combination of criterion can explain more than 70% of cryptocurrencies prices
variation using either Multiple Regression or Random Trees or Long Short-
Term Memory networks.
Keywords: cryptocurrency, bitcoin, forecasting, machine learning, multiple re-
gression, random forests, lstm, long short-term memory, python, data science.
1 Introduction
The story of cryptocurrencies had started in 2009 when Bitcoin was born. It was the
first decentralized digital currency with emphasis on cryptography [1]. After nine
years, almost 1500 live cryptocurrencies exist [2], and their number continues to
overgrow [3].
Another trending sphere that is developing rapidly in the last few years is Machine
Learning (ML). It is a method of data analysis that automates analytical model build-
ing. ML allows applications to find hidden insights without being explicitly pro-
grammed [4]. One of the most common task ML solves correctly is forecasting.
The purpose of the paper is to formalize criteria which affect the price of crypto-
currencies and use them in the price forecasting experiment developed with Python
Anaconda Data Science tool with ML algorithms.
The paper is organized as follows: part 2 describes dataset and ML algorithms used
in an experiment and the results of the experiment, and the last part concludes.
2 Cryptocurrencies Dataset and ML Prediction Algorithms
Information about cryptocurrencies like their prices per date, their supply, mining
difficulty, and other is open source. So we combined data into one dataset [5]. It has
next columns:
• date, from 25th of January 2017 to 22 of January 2018 separated by weeks;
• price - the price of Bitcoin for each date (in USD);
• supply - the Bitcoin's total number of coins for each date;
• difficulty - the Bitcoin's mining difficulty for each date (hash rate)[8-hashrate];
• trading_volume - the Bitcoin's trading volume for each date;
• reaction - average society reaction on Bitcoin for each date ("-1" is negative, "1" is
positive, and "0" is neutral);
Our dataset has 105 rows, and it was separated into two subsets: training and testing.
70% of data is training set (it is 73 rows) and 30% if data is verification set (it is 32
rows).
In our research we decided to use three supervised ML algorithms to verify if these
criteria can be used in cryptocurrency’s price prediction in such combination:
• Multiple Linear Regression with default configuration sklearn configuration;
• Random Forests with 100 trees;
• Long Short-Term Memory Networks with 50 neurons on a hidden layer and 100
epochs of training.
Linear regression shows relations between variables and how changes affect them.
Random Forests algorithm uses a bagging approach to create a bunch of decision trees
with a random subset of data. LSTMs are capable of learning long-term dependencies.
The research question is: Which criteria affect the price of the cryptocurrency and
how? We have found next points:
1. Total number of mined coins;
2. Mining difficulty level;
3. Cryptocurrency’s trading volume;
4. Perceptions of the cryptocurrency’s value by the society;
5. Price of Bitcoin.
After all three predictions, we were able to find mean squad error and coefficient of
determination. Linear regression application output was:
Mean squared error: 2274869.02
R^2 score: 0.79
But, Random Forests application output was:
Mean squared error: 585482.78
R^2 score: 0.97
And, finally LSTM network application output was:
Mean squared error: 3248292.00
R^2 score: 0.80
In the graphic below we display the results of both predictions in contrast to real val-
ues. (fig. 3).
25000
20000
15000
10000
5000
0
6.18.17 7.18.17 8.18.17 9.18.17 10.18.17 11.18.17 12.18.17 1.18.18
Original Linear Regression Random Forests LSTM
Fig. 1. Comparison of real Bitcoin prices, predicted with Linear Regression and predicted with
Random Forests [6]
3 Conclusions and Outlook
Thus we have tested the correctness of the selected criteria combination on their effect
on the price of Bitcoin. For our experiment, we used Multiple Linear Regression,
Random Forests, and LSTM ML algorithms implemented with Python in Anaconda
Data Science tool. As a result, we have found that selected combination of criterion
can explain more than 70% of Bitcoin’s price.
On this base, we plan to study additional criteria which affect prices of cryptocur-
rencies to be able to forecast their prices more accurate.
References
1. Bitcoin: A Peer-to-Peer Electronic Cash System, https://bitcoin.org/bitcoin.pdf, last ac-
cessed: 2018/1/27.
2. All cryptocurrencies, https://coinmarketcap.com/all/views/all/, last accessed: 2018/1/27
3. 2018 Will See Many More Cryptocurrencies Double In Value,
https://www.forbes.com/sites/kenrapoza/2018/01/02/2018-will-see-many-more-
cryptocurrencies-double-in-value, last accessed: 2018/1/27
4. Nilson, N.J.: Introduction to Machine Learning. An early draft of a proposed textbook,
Robotics Laboratory Department of Computer Science, Stanford University, Stanford, CA
94305 (2005)
5. Dataset, https://drive.google.com/open?id=1pBCtHw4DEGJnqiy-R8Ko3cgD2NRS-qw8
6. Chart dataset, https://drive.google.com/open?id=15CbWUZ2d-
ZLYoYentP5qGcF4GWgdn8yG