=Paper=
{{Paper
|id=Vol-3191/paper30
|storemode=property
|title=Predicting Bitcoin Volatility Using Machine Learning Algorithms and Blockchain Technology
|pdfUrl=https://ceur-ws.org/Vol-3191/paper30.pdf
|volume=Vol-3191
|authors=Mimoza Mjoska,Blagoj Ristevski,Snezana Savoska,Vladimir Trajkovik
|dblpUrl=https://dblp.org/rec/conf/isgt2/MijoskaRST22
}}
==Predicting Bitcoin Volatility Using Machine Learning Algorithms and Blockchain Technology==
https://ceur-ws.org/Vol-3191/paper30.pdf
Predicting Bitcoin Volatility Using Machine
Learning Algorithms and Blockchain Technology
Mimoza Mjoska 1, Blagoj Ristevski 1, Snezana Savoska 1 and Vladimir
Trajkovik 2
1
Faculty of Information and Communication Technologies, University “St. Kliment
Ohridski”, ul. Partizanska bb, Bitola, 7000 RN Macedonia
2
Faculty of Computer Science and Engineering, Ss. Cyril and Methodius University,
Skopje, RN Macedonia
Abstract
Blockchain technology has the potential to be applied in a variety of areas
of our daily life. Blockchain is the foundation of cryptocurrency, but the
applications of blockchain technology are much more expansive. This
technology is considered to be a revolutionary solution for the financial
industry. Also, it can be successfully applied in scenarios involving data
validation, auditing, and sharing. On the other hand, machine learning is
one of the most noticeable technologies in recent years. Both technologies
are data-driven, and thus there are rapidly growing interests in integrating
them for more secure and efficient data sharing and analysis. This paper
shows how these two technologies, blockchain and machine learning,
can be combined in predicting bitcoin volatility. To analyze and predict
bitcoin volatility, bitcoin data from real-time series and random forests as a
machine learning algorithm were used. When predicting bitcoin volatility,
low statistical errors were obtained in the training and test set. This confirms
that the forecasting model is well designed.
Keywords:
blockchain technology, machine learning algorithms, random forests, bit-
coin, time-series data.
1. Introduction
In 2008, powerful US financial institutions and insurance companies were on
the edge of bankruptcy. These circumstances called for immediate intervention
by the federal government to avoid domestic and possibly global financial col-
Information Systems & Grid Technologies: Fifteenth International Conference ISGT’2022, May 27–28, 2022, Sofia, Bulgaria
EMAIL: mijoska.mimoza@uklo.edu.mk (M. Mjoska); blagoj.ristevski@uklo.edu.mk (B. Ristevski); snezana.savoska@
uklo.edu.mk (S. Savoska); vladimir.trajkovik@finki.ukim.mk (V. Trajkovik)
ORCID: 0000-0002-4248-2760 (M. Mjoska); 0000-0002-8356-1203 (B. Ristevski); 0000-0002-0539-1771 (S. Savoska);
0000-0001-8103-8059 (V. Trajkovik)
© 2022 Copyright for this paper by its authors.
Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR Workshop Proceedings (CEUR-WS.org)
lapse. Moreover, these events illustrated the dangers of living in a digital, inter-
connected world that depends on transactional intermediaries and leaves people
vulnerable to digital exploitation, greed, and crime.
Blockchain technology is a fast-growing part of the Internet and cloud com-
puting. Similar data structures existed long before the famous bitcoin was con-
ceived, but the main theories about blockchain architectures used today were
originally defined in the original article on bitcoin written and published by a
person under the pseudonym Satoshi Nakamoto in 2008 [1].
Nowadays, we live in a world where an abundance of data surrounds us.
Therefore, it is interesting to develop software systems that can use this informa-
tion, learn from it, and offer valuable behaviours based on it. This is possible by
using machine learning algorithms.
Realized volatility measures the actual performance of an asset in the past
and helps to understand the stability of the asset based on its past performance.
It indicates how an asset’s price has changed in the past and the period in which
it has changed. Higher the volatility, the higher the price risk associated with the
stock and, therefore, the higher the premium attached to the stock. The realized
volatility of the asset may be used to forecast future volatility, i.e., implied vola-
tility of the asset. While entering into transactions with complex financial prod-
ucts such as derivatives, options, etc., the premiums are determined based on the
underlying volatility and influence the prices of these products [26].
This paper describes an algorithm that combines blockchain technology and
machine learning algorithms to predict bitcoin volatility. The remainder of the
paper is organized as follows. Section 2 highlights the principles of blockchain
technology and its application. Section 3 elaborates on the machine learning al-
gorithms focusing on the random forest algorithm. Next, an application in the
programming language R is used to predict the bitcoin volatility, as described in
section 4. Finally, concluding remarks and directions for further works are given
in the last section.
2. Blockchain technology
“Blockchain” is a coined word composed of the words “block” and “chain”.
Blockchain is a distributed replicated database organized in the form of a single
linked list – chain, where nodes are blocks of data for transactions. These blocks,
after grouping, are protected by using cryptographic methods [3][4]. Blockchain
technology enables the accomplishment of digital transactions without interme-
diaries.
Blockchain (distributed ledger technology) is a network software protocol
that enables the secure transfer of money, assets, and information through the
Internet without a third-party organization as an intermediary [3]. It can safe-
347
ly store transactions such as digital cryptocurrencies or data/information about
debt, copyrights, equity, and digital assets. Furthermore, the stored data cannot
be easily forged and tampered with because it requires individual approval of all
distributed nodes. This significantly reduces the cost of trust and accounting in
non-digital economies and other social activities [2].
Blockchain is a technology that is constantly evolving. The most common
types of blockchains are:
• public Blockchain [6],
• private Blockchain and
• hybrid Blockchain [10].
Researchers pointed out that applying blockchain technology to cross-border
payment has a high potential effect. Holotiuk et al. in [14] stated that Blockchain
technology will improve the payment system by providing a solid structure for
cross-border transactions, removing expensive intermediary costs, and gradually
weakening or altering the business model of the existing payment industries [11].
Because modifying transactions or whole blocks is almost impossible in block-
chain, blockchain technology makes it easy to prove the integrity of the electronic
files used in accounting and auditing[13]. This technology is considered to be a
revolutionary solution for the financial industry. Also, this technology is the basis
of the digital currency bitcoin whose volatility is analyzed in this paper.
3. Machine learning algorithms
Among the numerous applications of machine learning in healthcare, science,
industry, etc., is the timely detection of diseases such as cancer, glaucoma and other
conditions that take human lives at high speed. Another application is the visu-
alization of smart cars, efficient web browsing that facilitates Internet searches,
language translations that help immensely in world communications and limit the
significant language barrier between countries, and the implementation of fraud
detection and face-recognition systems. Well-known machine learning algorithms
are the k-nearest neighbors (k-NN) algorithm, artificial neural networks, random
forests, support vector machines, the Naïve Bayesian classifier, etc. Machine learn-
ing systems can be classified according to the amount and type of supervision they
receive during the training. There are three main categories: supervised learning
[18], unsupervised learning [16], and reinforcement learning [20] [16].
3.1. Random forests
The Random Forest is a supervised learning algorithm used for regression and
classification. Technically, it is an ensemble algorithm. The algorithm generates
individual decision trees through an attribute selection indication. Each tree relies
348
on an independent random sample [19]. In a classification problem, each tree votes,
and the most popular class is the end result. On the other hand, in a regression prob-
lem, the average of all trees’ outputs is calculated, which will be the result [19]. This
algorithm is used to predict the realized volatility of Bitcoin, in this paper.
4. Discussion and analysis of the results obtained in creating a
model that overlooks the time series of realized volatility of the
market price of bitcoin
4.1. Realized volatility
Volatility forecasting plays a critical role in financial modelling and financial
decision-making. Realized volatility assesses variation in returns for an invest-
ment product by analyzing its historical returns within a defined period [26]. As-
sessment of the degree of uncertainty and/or potential financial loss/gain from
investing in a company may be measured using variability/volatility in the stock
prices of the entity. In statistics, the most common measure to determine variabil-
ity is by measuring the standard deviation, i.e., the variability of returns from the
mean. It is an indicator of the actual price risk [26].
The realized volatility or actual volatility in the market is caused by two
components – a continuous volatility component and a jump component, which
influence the stock prices. Continuous volatility in a stock market is affected by
intra-day trading volumes. For example, a single high-volume trade transaction
can introduce a significant variation in the price of an instrument [26].
This paper predicts the realized volatility of bitcoin. Analysts use High-vari-
ance daily data to estimate hourly/daily/weekly or monthly frequency levels. The
data can then be used to estimate volatile sales movement. During the analysis,
data whose frequency is 1 hour from the Gemini platform were taken [21], and
then using that data, the realized volatility is calculated with a daily frequency.
There is OHLC (Open / High / Low / Close) pricing data in each file that is updat-
ed daily. For this paper, granular hourly data are taken back to the 2015 year, for
the pair of bitcoin/dollar. The estimation of variability is calculated by measuring
the standard deviation from the average price of the monitored object in a given
period. Because volatility is nonlinear, the realized variance is first measured by
translating the values taken
from the stock market into logarithmic values and
then calculating the standard deviation. The realized variance is calculated by
calculating the sum of the squares of the standard deviation. The next step is to
calculate the realized volatility, which is the square root of the realized variance:
(1)
349
To calculate the realized volatility of bitcoin, an application was created in
the programming language R. For the necessary analysis, the package Rstudio
Version 1.4.1717 for ubuntu 20.04 and the necessary packages dplyr, forecastML,
ggplot2, glmnet, DB, randomForest and caret were used. The application first
loads the hourly market price of bitcoin downloaded from the Gemini platform
using the read.csv() function.
A date sequence is then added using the seq function, defining the start
and end time points with a frequency of 1 hour. In the time series “2015-10-08
13:00:00” is taken as the starting date, and “2022-01-12 12:00:00” is taken as the
end date. The price of bitcoin was taken at the close of the calculations, and other
data were omitted. The logarithmic values of the bitcoin price are calculated to
make the results more accurate. Using the library (dplyr) are created a new col-
umn with a unique date for each day. It is an identification column, so later, using
the group_by, sum and arrange functions, the squares of the standard deviation
corresponding to one date for one day are summed so that we get the realized
variance. Using the sqrt() function, the realized volatility in daily frequency (1)
is calculated.
For the needs of the research, free data is downloaded from the website
https://www.blockchain.com/charts in .csv format for the last 3 years for the
properties shown in Table 1.
Table 1
Properties of bitcoin are downloaded from https://www.blockchain.com/charts
Variable Description
cost_per_transaction Miners’ income is divided by the number of transac-
tions.
cost_per_transaction_percent Miners’ income as a percentage of transaction volume.
estimated_transaction_volume_usd The total estimated value in US dollars of blockchain
transactions.
miners_revenue The total US value of block rewards and transaction
fees paid to miners.
n_transactions The total number of confirmed transactions per day.
n_transactions_per_block The average number of transactions per block in the
past 24 hours.
output_volume The total value of all outgoing transactions per day.
trade_volume The total value in US dollars on the trading volume of
major bitcoin exchanges.
transaction_fees_usd The total value in US dollars of all transaction fees
paid to the miners.
350
All the data for the properties with the read.csv() command are loaded into
separate variables in the R programming language. Then all the data is loaded
into a data table with the command data.frame(). Finally, the volatility of bitcoin
is added to that data, which was previously calculated with data downloaded from
the cryptocurrency exchange platform www.gemini.com.
Once the database is complete, some transformations need to be made to bal-
ance the data to begin data analysis. Each vector x is replaced by its natural loga-
rithm with function in R, log(x). This transformation is done so that the results of
statistical analysis can become more accurate.
To prepare the data for machine learning, the data is normalized. The pur-
pose of the normalization is to adjust the values of the numeric columns in the
database to a standard scale, without distorting the variations in the range of val-
ues. The min-max normalization method is selected. The data is normalized to
have values between 0 and 1.
A machine learning algorithm randomForest is used to predict the realized
bitcoin volatility. The random forest algorithm creates a large number of decision
trees during training. From the created set of decision trees, the random forest
method gets a prediction from each tree individually and the average of all trees’
outputs is calculated, which will be the end result.
The algorithm for predicting bitcoin volatility uses the forecastML package
in the programming language R. When machine learning algorithms are used, the
model is first generated using training data, and then the values for the test data
are predicted. Once the data is ready, the next step in the research is to divide the
data into a training and test set. The data used to predict the volatility of bitcoin
contain 1095 observations, starting from 14.01.2019 to 12.01.2022. The first 915
observations are taken for the training set, and the remaining 180 observations
are used as a test set.
The forecasting method uses four different forecasting horizons. These dif-
ferent horizons are used to predict in the short and long term to combine the
predictions in the final forecast and thus minimize the error. The randomForest
function is then defined with its arguments.
The first step in the prediction process is to create some validation windows
to perform nested cross-validation. After training the model the predictions, re-
siduals, and some error metrics are presented. The test set is then predicted using
the validation windows, and the actual versus predicted values are displayed. Ini-
tially, the size of each forecast horizon is defined. The first horizon is seven steps
ahead, the second is 30 steps forward, the third is 90 steps forward, and the last is
180 steps forward. The horizons to be predicted are chosen, and the view through
certain time steps in the past is also chosen.
Ten validation windows are created. This means that ten models will need
to be trained for each direct forecast horizon, each theoretically selecting differ-
351
ent optimal hyperparameters and having different coefficients from the internal
cross-validation process [27]. Estimating variations between these models is a
good way to estimate stability under the dynamics of different time series in a
given modeling method.
In the given research the elbow method is used to select which number of
trees will be used in the given test. After approximately 100 trees, there is almost
no difference in error reduction. That is why it is chosen to use 100 trees in this
research.
The importance of the variables is used so that we can understand which
variables affect the volatility of bitcoin.
The importance of the variables used in the forecast is presented in Figure 1.
Figure 1: Importance of the properties of bitcoin
From Figure 1, it can be seen that of the selected features, cost.per.transac-
tion is the most important in increasing the percentage of the mean square error
(% IncMSE), and miners.revenue is the most important in increasing the node
purity (IncNodePurity).
The following Figure 2 shows the standard errors in the training set:
Figure 2: Standard set errors in the training set
The next step is to predict the test set. The error metric for the test set by
horizons, the predicted values versus the actual values, is presented in Figure 3.
352
Figure 3: Standard errors in the test set
Compared to the training set error metric, the test set error metric is smaller.
The error values are satisfactory, indicating that the model is good.
In the second part of the analysis, it is necessary to train the model through-
out the entire training database without nested cross-validation. Without nested
cross-validations and retention windows, the forecast graph basically fits the
model.
The next step is to predict the test set and re-display the real versus predicted
values and error metrics for the test set.
The predictions on each horizon of the predicted values of the test set are
then combined. The final part is to predict the off-sample for each horizon using
the training and test set and re-combine the out-of-sample forecasts for each ho-
rizon to display the final combined forecast.
A new model is now being re-trained using only one validation window.
The standard prediction errors in the training set with a single validation
window are shown in Figure 4.
Figure 4: Standard errors in the training set generated by R
The table from Figure 5 shows the error metrics of the predicted values in the
test set without validation windows.
Figure 5: Standard errors in the test set
353
If we compare the data in Figure 3 and Figure 5, it is noticed that the test set
is error is more significant than the training set. That result is expected because
test data are not used when the model is being trained. The fact that the error in
the test set does not become extremely large indicates that our model predicts
well. The mean absolute error in the training set is 0.006, and the test set is 0.015,
proving that the model has a good prediction. According to the random forest
algorithm, the next step is to combine the predictions of each direct forecast ho-
rizon using the combine_forecasts () function.
The next step according to the random forest algorithm is to make an off-
sample prediction, which is presented in Figure 6.
Figure 6: Prediction outside the sample Predict the horizons of each out-of-
sample horizon again using the combine_forecasts() function
The obtained combined forecast scheme outside the sample for 180 steps
forward is presented in Figure 6. The command combine_forecasts() predicts 180
steps forward, where different predictions are combined to get a better result.
Using the standard type = “horizon”, forecasts are combined so that short-term
forecasts are produced from short-term models and long-term forecasts from the
long-term model.
5. Conclusion
Bitcoin is the most popular cryptocurrency, nowadays. It’s widely researched
in economics and computer science. This paper uses a random forest machine-
learning algorithm to predict the time series of realized volatility of the bitcoin.
This algorithm is one of the best machine learning algorithms and predicts very
well with high accuracy, gives better and more accurate results than traditional
354
predictions, such as autoregressive and vector autoregressive methods. This pa-
per also examines whether the information obtained from blockchain can be used
to predict the volatility and price of bitcoin. Many people in the world use bitcoin
as an investment due to its high volatility, and in this way, they can get enormous
profits for a certain period. The R package uses the forecastML library, which
contains many time-series modeling visualizations. It is concluded that block-
chain information may be included as a variable in further research into the price
and volatility of Bitcoin.
According to the literature the smaller the mean absolute error, the better the
prediction model. In our model mean absolute error (mae) in the training set using
10 validation windows is 0.019, and in the test set, the mean absolute error from
the different horizons is 0.015. The mae in the training set without validation win-
dows is 0.006, and in the test set is 0.015. The paper uses data on variables only
from blockchain information and concludes that the obtained empirical results
correspond to the literature that blockchain information can be used to examine
the price and volatility of bitcoin. From this research, we can conclude that if in
the real data the volatility of bitcoin is big, then the prediction error will be big-
ger. From the example shown, it can be concluded that the error in predicting the
data is smaller when the forecast horizon is shorter and when volatility is smaller.
For further research different numbers of trees in the random forest, and
different forecast horizons can be used. In addition, different machine learning
algorithms can be used for further work. In this way, the accuracy of different
time series algorithms can be tested and the one with the best results and accuracy
can be selected for prediction. To determine the error with which they predict and
compare that error to choose which machine learning algorithm is most suitable
to predict the relevant data.
6. References
[1] Satoshi Nakamoto (2008). “Bitcoin: A Peer-to-Peer Electronic Cash Sys-
tem”.
[2] Dragana Tadić Živković (2018). “Blockchain technology: opportunity or a
threat to the future development of banking”, Proceedings of Ekonbiz.
[3] Swan M. (2015). “Blockchain: Blueprint for a new economy”, https://
www.goodreads.com/work/best_book/44338116-blockchain-blueprint-for-
a-new-economy (last accessed 06/10/2021).
[4] Mijoska M. and Ristevski B. (2020). “Blockchain Technology and its Ap-
plication in the Finance and Economics”, 10th International Conference
on Applied Information and Internet Technologies – AIIT 2020, October,
Zrenjanin, Serbia.
[5] Li Zhang, Yongping Xie, Yang Zheng, Wei Xue, Xianrong Zheng, Xiao-
355
bo Xu (2020). “The challenges and countermeasures of blockchain in fi-
nance and economics”, John Wiley & Sons, Ltd. https://doi.org/10.1002/
sres.2710.
[6] Dejan Vujicic, Dijana Jagodic, Siniša Ranđić (2018). “Blockchain technol-
ogy, bitcoin, and Ethereum: A brief overview”, Available online: https://
doi.org/10.1109/INFOTEH.2018.8345547.
[7] Julija Basheska, Vladimir Trajkovik (2018). “Blockchain based Transfor-
mation in government: review of case studies”, ETAI 2018.
[8] Nick Szabo (1997). “The idea of smart contracts”. Nick Szabo’s Papers
and Concise Tutorials. https://fon.hum.uva.nl/rob/Courses/Information-
InSpeech/CDROM/Literature/LOTwinterschool2006/szabo.best.vwh.net/
idea.html (last accessed 06.10.2021).
[9] Aleksandar Matanović, “Osnove kriptovaluta i blokčein tehnologije”,
http://fzp.singidunum.ac.rs/demo/wp-content/uploads/Osnove-kriptovalu-
ta-i-blok%C4%8Dein-tehnologije.pdf, (last accessed 06.10.2021).
[10] Gu, J.; Sun, B.; Du, X.; Wang, J.; Zhuang, Y.; Wang, Z. (2018). “Consor-
tium blockchain-based malware detection inmobile devices”, IEEE.
[11] https://doi.org/10.1088/1742-6596/1693/1/012025.
[12] Androulaki, E.; Barger, A.; Bortnikov, V.; Cachin, C.; Christidis, K.; De
Caro, A.; Enyeart, D.; Ferris, C.; Laventman, G.; Manevich, Y.; et al. (2018).
“Hyperledger fabric: A distributed operating system for permissioned block-
chains”. In Proceedings of the Thirteenth EuroSys Conference ACM, Porto,
Portugal, 23–26 April; pp. 1–15. https://doi.org/10.1145/3190508.3190538.
[13] Hyperledger Burrow – Hyperledger. Available online: https://www.hy-
perledger.org/projects/hyperledger-burrow (last accessed 06.10.2021).
[14] Quora (2017). “How will blockchain impact accounting, auditing & fi-
nance?”, https://www.quora.com/How-will-blockchain-impact-account-
ing-auditing-finance, (last accessed 06.10.2021).
[15] Holotiuk, F., Pisani, F., & Moormann, J. (2017). The impact of blockchain
technology on business models in the payments industry. Wirtschaftsin-
formatik 2017 Proceedings. Retrieved https://wi2017.ch/images/wi2017-
0263.pdf (last accessed 06.10.2021).
[16] Crosman, P. (2017). “R3 to take on Ripple with cross-border payments
blockchain”. American Banker; New York, N.Y. Retrieved from https://
www.bitcoinisle.com/2017/10/31/r3-to-take-on-ripple-with-cross-border-
payments-blockchain/ (last accessed 06.10.2021).
[17] Aurélien Géron, Hands-On Machine Learning with Scikit-Learn and Ten-
sorFlow, Published by O’Reilly Media, 2017.
[18] Molly Galetto, Machine learning and big data analytics: the perfect mar-
riage http://www.ngdata.com/machine-learning-and-big-data-analytics-
the-perfect-marriage.
356
[19] Sotiris B Kotsiantis, Ioannis D Zaharakis, and P.E.Pintelas. Machine learn-
ing: a review of classification and combining techniques. Artificial Intel-
ligence Review, 26(3): 159–190, 2006.
[20] Pavan Vadapalli, Random Forest Classifier: Overview, How Does it Work,
Pros & Cons, https://www.upgrad.com/blog/random-forest-classifier (last
accessed 08.07.2021).
[21] Marko ˇCupić, Umjetna inteligencija, Uvod u strojno uˇcenje, 2020, http://
java.zemris.fer.hr/nastava/ui/ml/ml-20200410.pdf.
[22] Gemini, https://www.cryptodatadownload.com/data/gemini.
[23] Ross Jacobucci, Random Forests, University of Notre Dame, https://sta-
tisticalhorizons.com/wp-content/uploads/2021/11/Advanced-Machine-
Learning.pdf.
[24] Daniel Johnson, R Random Forest Tutorial with Example, 2022, https://
www.guru99.com/r-random-forest-tutorial.html (last accessed 11.04.2022).
[25] Jeffrey Craig, What is Transactions Per Second (TPS): A Comparative
Look at Networs,2021, https://phemex.com/blogs/what-is-transactions-
per-second-tps (last accessed 11.04.2022).
[26] Miroslav Minovic, Blockchain technology: usage beside crypto curren-
cies, (2017). Аvailable online: https://www.researchgate.net/publica-
tion/318722738.
[27] Madhuri Thakur, Realized volatility, https://www.wallstreetmojo.com/real-
ized-volatility/ (last accessed 08.04.2022).
[28] Aristeidou Christoforos, Study of the volatility of Bitcoin cryptocurrency
using Machine Learning methods: An implementation in R, 2020.
357