Download as pdf or txt
Download as pdf or txt
You are on page 1of 5

Machine Learning-Based Timeseries Analysis for

Cryptocurrency Price Prediction: A Systematic


Review and Research
Siva Kumar.A1 Gopirajan PV2,* Beulah Jackson3
1Assistant 2Assistant 3Professor,
Professor, Professor,
Department of Data Science and Department of Computational Department of ECE,
2023 International Conference on Networking and Communications (ICNWC) | 979-8-3503-3600-9/23/$31.00 ©2023 IEEE | DOI: 10.1109/ICNWC57852.2023.10127439

Business Systems, Intelligence, Vel Tech Rangarajan Dr. Sagunthala


School of Computing, School of Computing, R & D Institute of Science and
Technology,
SRM Institute of Science and SRM Institute of Science and
Chennai, India
Technology,Kattankulathur Campus, Technology, Kattankulathur Campus,
beulah.jack@gmail.com
Chennai – 603203, India Chennai – 603203, India
siva.aaru53@gmail.com gopivrajan@gmail.com
*Corresponding author

Abstract— A virtual currency known as cryptocurrencies able to determine the plan and proceed accordingly.
holds all business online. It's virtual money that wouldn't Timeseries models like ARIMA (Autoregressive
materialize like complicated conventional paper currency. Integrated Moving Average), Auto-Regressive (AR),
Thus, this study emphasizes a distinction between distributed
Moving Average (MA), Auto-Regressive Integrated
paper currency and cryptocurrencies, where these individuals
may access information without outside interference. Because Moving Average (ARMA), ARIMAX, and Seasonal Auto-
of its considerable market swings, such cryptocurrencies have Regressive Integrated Moving Average with eXogenous
an influence upon commerce as well as foreign diplomacy. factors (SARIMAX) methods can help in future prediction
Virtual currencies which are available in the market, such as based on the current flow of data and past information.
Bitcoin (BTC), Ethereum (ETH), Terra (LUNA), Solana This research aims to examine whether the price of
(SOL), Cardano (ADA), Tether (USDT), Binance coin (BNB), Bitcoin can be predicted like other stock market tickers.
USD coin, XRP coin, Avalanche coin (AVAX) and Lite coin This will depend on whether we can further use it as a
(LTC), etc. This study focussed on a detailed analysis of the
payment medium. There are various challenges to address
literature about Machine Learning (ML) methods used for
predictions. This proposed work also focused on when it comes to scientific research. Similarly, they do not
implementing an efficient Machine Learning (ML)-based time rate well as market drivers regarding market capitalization.
series model for predicting BTC cryptocurrency prices. Long Another top-ranked pre-mined currency has a third feature
Short-Term Memory (LSTM) forecasting theory was that Society open-sourced, non-mineable coins lack. Non-
established to accommodate the fluctuation of bitcoin prices mineable coin transactions are also supported via a
and achieve great precision. The effectiveness of the LSTM in controlled blockchain.
predicting the price of a cryptocurrency is demonstrated by
this suggested study's comparison between it and comparable II. RELATED WORKS
time-series models.
This literature review includes work on BTC [1] price
Keywords: Bitcoin, LSTM, Cryptocurrency, Machine prediction utilizing various methodologies, the necessity
learning, Price prediction. for, and the evaluation of Recurrent Neural Networks
(RNN) [2]. Machine learning and artificial intelligence
I. INTRODUCTION play a meaningful work irrespective of the domains [3],
Until the market date, the top-rated two [4]. Machine learning research on cryptocurrency
cryptocurrencies are bitcoin (BTC) and Ethereum (ETH). prediction is insufficient, particularly on deep learning
Traders trade these currencies and make a substantial models. Towards this relevant topic, more than 600 papers
profit, buying or selling that currency on the coin swap have indeed been released. [5]. Islam et al. discussed the
platform through network-based exchange channels that different types of cryptocurrencies and explained the
encrypt transactions. They have a wallet where they keep importance of blockchain technology for implementing
all their money and savings. In addition, cryptocurrency is secured cryptocurrency [6]. Artificial Neural Networks
similar to the stock market because it buys and sells shares. (ANN), Convolutional Neural Networks (CNN), RNN,
They will buy and sell currency in the same manner. and other neural network models of Machine learning
However, the price varies considerably based on the methodologies help in categorizing and predicting through
criteria we described earlier. Again for purpose of statistical approach [7]–[10]. Zhaofang Li et al. proposed
forecasting well-known cryptocurrencies like bitcoin and an idea based on Distributed Ledger Technology (DLT) in
Ethereum, several academics had explored with machine cryptocurrency for providing high security to transfer the
learning and deep learning techniques. Many digital digital currency and track the currency from its initial
currencies, such as Tron, Bitcoin, Stellar, and others, have position to its current position [11]. Though blockchain
a variety of banking uses. supports the cryptocurrency and its product, bitcoin, cyber
Trading firms want cryptocurrency price predictions criminals may induce data safety. Possible attacks could
since they must set an objective. As a result, it will float up happen online at an end. Ransomware is one such
and down. As a result of the price estimate, they will be inevitable threat to the data [12].

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE


Akanksha Jalan et al. have compared the leading III. PROPOSED SYSTEM
cryptocurrencies, such as BTC, ETH, and LTC, through
A. Data collection
surveys that happened on social media. They have
portrayed the importance of having trust in the people who Its an essential and initial phase for adequate algorithm
are doing international digital currency transformation preparation. This proposed system was developed from the
[13]. Moritz Wendla et al. discussed the work and stake free-to-use dataset in "Kaggle." The "Cryptocurrency
proof of the cryptocurrencies available in the market. They Historical Prices" dataset available in Comma-Separated
have also concerned about the algorithms used for Value ('.csv') file format was used primarily along with the
cryptocurrency trading [14]. They have concluded the other data collected from https://coinmarketcap.com/
significance of policy-making in digital money and safety. (open to use) website[25]. And also, more dataset was
Though cryptocurrency has many advantages in terms of prepared from web-mining methods[26]. Web crawlers
security and ease of use, Benjamin Johnson et al. have were also developed exclusively for this study to collect
discussed the additive behavior of digital money. information from multiple websites. All the datasets
Cryptocurrencies are easily used in gambling, leading to involved in this study were stored in the cloud repository
the loss of mental health and happiness [15]. Saeed Sazzad for easier access. The experiment's goal will be to produce
Jeris et al. discussed the trend of cryptocurrency and how the result with more accuracy. Using the same test data, we
it is associated with the livestock market. Though bitcoin will test and analyze them to determine their predictive
was launched in 2009, it became so popular in 2013. Thus, performance. Dataset was prepared for almost 12 different
the author discussed the progress of crypto-assets over the cryptocurrencies. Out of which, the bitcoin dataset was
stock market. made a high priority. Figure 1 depicts the detailed flow of
Alex De Vries discussed the impacts of the the anticipated study. Datasets collected from various
merger of BTC and ETH that happened in September sources were combined as a single dataset and pre-
2022. The author described the power supply usage before processed to remove redundant data. Then those data were
and after the merger and how the power demand was transformed into the required format (‘.csv’).
reduced. Sidharth Quamara and Awadhesh Kumar Singh
discussed the security constraints and methods associated
with cryptocurrencies. They have clearly illustrated the
fault-tolerant mechanism and Proof of Work (PoW)
methodology applied in cryptocurrencies [16].
Yao Yue et al. suggested the ideology of
economic impact after the cryptocurrency launch. They
have surveyed the effects in China [17]. Yi-Shuai Ren et
al. reviewed the application of ML in cryptocurrency[18].
They have stressed the significance of implementing ML
for price prediction. They have also concluded that
Random Forest (RF) is the most used ML methodology in
classification and price prediction.
Patrick Jaquart et al. analyzed various ML models Fig 1. Cryptocurrency price prediction system architecture
for online trading and price prediction. They have also
achieved a prediction accuracy of about sixty percent [19]. The table content in figure 2 shows the sample dataset
Gian Pietro Bellocca et al. discussed the implementation of collected online. For each column in the dataset shown in
ML methods into practice to anticipate trading and the figure, a variable is assigned to segregate the values
cryptocurrency cost momentum. They have introduced such as high, low, open, close, and Marketcap. These
candlestick, trend, Volatility, and volume approach in their variables were assigned a multi-dimensional array value
ML model to predict the outcomes [20]. Yaqi Wang et al. for further processing. As a next step, those array values
support the prediction accuracy of the ML models in their were fed to the algorithm for training the model. Results
study. They have implemented various ML models and obtained for MA, RA, ARIMA, and LSTM were validated
tested the behavior of the prediction result. They with the real-time data available. Their results were
concluded that lacking training sets would lead to poor compared to figure out the optimal ML model for price
outcomes [21]. The influence of data mining, transaction prediction. The results were stored in the cloud repository
network analysis, and continuous market watch are critical for future use.
factors for a good prediction model [22].
Haijun Lin et al. proposed an LSTM model for B. Data cleaning, pre-processing, and transformation
predicting heart rate. They compared the efficiencies of
Cleaning and extracting only the required information is
LSTM with RF, ARIMA, and a few other methods and
vital as the data is collected from multiple sources. This
concluded that LSTM was more efficient [23]. LSTM was
study uses a formal data cleaning procedure such as
very successful for Human Activity Recognition model
redundant data removal, adjusting structural errors, and
proposed by Hao Wu et al. [24].
missing data organization. The K Nearest Neighbour
(KNN) [27] method was used for the missing data. KNN
provides the average values calculated from the existing
neighboring values. This cleaned data were compiled into
a single '.csv' file for feeding the algorithm. This master
file contains around 79000 rows of information pertaining
to bitcoin.

Fig 3. Close value graph

This graph contains the comparison of density VS close for


the dataset used. Close value refers to the last achieved
price for the trade that happened. This graph shows a high
peak achieved during the initial twenty thousand data and
values decreasing for the remaining increase in data.

Fig 2. Algorithm implementation workspace

C. Model generation
The prepared dataset was inputted to the LSTM, MR, AR
and ARIMA models using 'sklearn' and 'pandas' packages
in Python. Figure 2 shows the implementation process of
the proposed work.

IV. RESULTS AND DISCUSSION


Table 1 shows the % accuracy comparison of the AR, MR,
Fig 4. High-value graph
ARIMA, and LSTM ML models used in this study for
various training and test set combinations. Totally, 78922 Figure 4 shows the high-value graph drawn for the given
rows of the dataset were considered to validate this input dataset. Here x-axis takes a high value, and the y-axis
proposed model. From the table, it is evident that 80:20 is takes the count value. This shows the constant variation of
the optimal ratio to be followed for training and test set. data between 0 to 1. Variations were noted with multiple
The LSTM model achieved higher % accuracy of about colors. The proposed model was designed as a user
71% for the 80:20 ratio. interface where the user can see the day, month, and year
predictions. For day-wise information, the model predicts
TABLE 1. ACCURACY COMPARISON OF ML MODELS
the hourly information, which will be very useful for those
Iterati Datas Trainin Testing ML model Accuracy % who want to see the progress of money translational value
on et g data data A M ARI LST
Numb rows percent percent R R MA M
while buying bitcoin.
er age age
1 78922 70 30 5 59 61 65
1
2 78922 75 25 5 58 59 68
3
3 78922 80 20 5 61 64 71
4
4 78922 85 15 5 55 62 67
2
5 78922 90 10 5 57 63 70
5
Fig 5. Day-wise and month-wise prediction

Relatively, for 75:25, LSTM achieved 68% for 75:25 and Figure 5 shows the day and month trends as well as the
70% for 90:10. Next to LSTM, the ARIMA model got an prediction of bitcoin cost. Figure 5.A shows the single-day
accuracy % of 64, AR model got an accuracy of 55% for prediction, whereas 5.B shows the five days prediction,
the 90:10 ratio. MR model got a 61% accuracy ratio of 5.C shows the one-month prediction, and 5.D shows the
80:20. Figure 3 shows the close value graph obtained for six-monthly prediction. It can be noted that the day-wise
this proposed LSTM method. predictions show negative or falling values while the one-
month prediction shows positive values. This variation in
values is due to the market's currency value fluctuation.
Similarly, Figure 6 shows the year-wise prediction. 6.A
and 6.B shows the one-year prediction. It can be seen
clearly that there is a fall of bitcoin values from 30L to 15L
during the period May 2022 to October 2022. 6.C and 6.D
shows the five years and above data. There is a shortfall in
2022 year-end due to the market trend can be noted in these
graphs. This figure 5 and 6 were generated for the LSTM
model. These values were validated with the existing
dataset, and it was found that the LSTM prediction model
could achieve 71% accuracy compared with the other
models.

Fig 8. Analysis of ML predictive performance in %.

The bar chart in Figure 8 compares the effectiveness


percentages of the ML models employed in this
investigation. It can be clearly noted that the LSTM model
could able to provide higher accuracy than the other time
series models. This is due to the more robust RNN
Fig 6. Year-wise prediction skeleton. LSTM models provide decision-making support
for the buyers during investment.
V. CONCLUSIONS
Due to rapid changes in the innovation that underpins
them, cryptocurrencies are gaining popularity, and traders
view them as falling within the category of asset
classes. Investors and investigators should keep a close
eye on this equity portfolio in the economic world, putting
more emphasis on cryptocurrencies. We examined the
accuracy of twelve significant cryptocurrencies employing
machine-learning techniques. Numerical experiments
were conducted for four different ML algorithms: AR, MR,
Fig 7. Comparison of training and validation set prediction ARIMA, and LSTM. Algorithms demonstrate the
predictability of upward or downward price moves. The
Figure 7 compares the training set, validation set, and average prediction accuracy percentage for AR, MR,
prediction values of the proposed LSTM model. As we ARIMA, and LSTM are 53, 58, 61.8, and 71.2. The LSTM
know the current currency trend through the available model could achieve predicted accuracy above 70% by
dataset, the validation was made considering the 2013- fine-tuning the model selection process. These results
2020 years dataset. With the sixty thousand rows of the conclude that the well-trained ML models could attain
training set, the LSTM model could attain around 70% better precision when the input training is sufficient.
accuracy. The accuracy could be achieved even more when
a strong daily-basis and hourly-basis collected dataset is
available.
splits,” Technol Forecast Soc Change, vol. 148, p. 119743,
REFERENCE 2019, doi: https://doi.org/10.1016/j.techfore.2019.119743.
[7] S. Chen, C. Zhang, M. Dong, J. Le, and M. Rao,
[1] E. Atkins, “Populist ecologies of Bitcoin,” Polit Geogr, vol. 94,
“Chen_Using_Ranking-CNN_for_CVPR_2017_paper.pdf,”
p. 102535, Apr. 2022, doi: 10.1016/J.POLGEO.2021.102535.
Cvpr, pp. 5183–5192, 2017.
[2] P. P. Ray, “A survey on cognitive packet networks: Taxonomy,
[8] H. Tang, “Image Classification based on CNN: Models and
state-of-the-art, recurrent neural networks, and QoS metrics,”
Modules,” in 2022 International Conference on Big Data,
Journal of King Saud University - Computer and Information
Information and Computer Network (BDICN), 2022, pp. 693–
Sciences, vol. 34, no. 8, pp. 5663–5683, Sep. 2022, doi:
696. doi: 10.1109/BDICN55575.2022.00134.
10.1016/J.JKSUCI.2021.05.017.
[9] C. Li and Y. Tao, “Application of support vector machine with
[3] P. Rayavel, A. Sivakumar, P. v Gopirajan, S. Surenderanath,
simulated annealing algorithm in MBR membrane pollution
and P. Rathnavel, “Real time Machine Learning Approach for
prediction,” in 2017 IEEE 15th International Conference on
a Smart Door Unlocking Using Face recognition System,” IOP
Software Engineering Research, Management and
Conf Ser Mater Sci Eng, vol. 1130, no. 1, Apr. 2021, doi:
Applications (SERA), 2017, pp. 211–217. doi:
http://dx.doi.org/10.1088/1757-899X/1130/1/012080.
10.1109/SERA.2017.7965730.
[4] G. Sivaranjani, P. V. Gopirajan, C. Gowdham, A. Abitha, and
[10] C. M. dos Santos, J. F. Escobedo, É. T. Teramoto, and S. H. M.
N. V. Ravindhar, “Computational Data Analysis, Prediction
G. da Silva, “Assessment of ANN and SVM models for
and Forecast of Health Disaster: A Machine Learning
estimating normal direct irradiation (Hb),” Energy Convers
Approach,” in 8th International Conference on Smart
Manag, vol. 126, pp. 826–836, 2016, doi:
Structures and Systems, ICSSS 2022, 2022. doi:
10.1016/j.enconman.2016.08.020.
10.1109/ICSSS54381.2022.9782289.
[11] Z. Li, Q. Lu, S. Chen, Y. Liu, and X. Xu, “A Landscape of
[5] C. W. Su, M. Qin, R. Tao, and M. Umar, “Financial
Cryptocurrencies,” in 2019 IEEE International Conference on
implications of fourth industrial revolution: Can bitcoin
Blockchain and Cryptocurrency (ICBC), 2019, pp. 165–166.
improve prospects of energy investment?,” Technol Forecast
doi: 10.1109/BLOC.2019.8751469.
Soc Change, vol. 158, no. June, p. 120178, 2020, doi:
[12] M. Conti, A. Gangwal, and S. Ruj, “On the economic
10.1016/j.techfore.2020.120178.
significance of ransomware campaigns: A Bitcoin transactions
[6] A. K. M. N. Islam, M. Mäntymäki, and M. Turunen, “Why do
blockchains split? An actor-network perspective on Bitcoin
perspective,” Comput Secur, vol. 79, pp. 162–189, Nov. 2018,
doi: 10.1016/J.COSE.2018.08.008.
[13] A. Jalan, R. Matkovskyy, A. Urquhart, and L. Yarovaya, “The
role of interpersonal trust in cryptocurrency adoption,” Journal
of International Financial Markets, Institutions and Money, p.
101715, Dec. 2022, doi: 10.1016/J.INTFIN.2022.101715.
[14] M. Wendl, M. H. Doan, and R. Sassen, “The environmental
impact of cryptocurrencies using proof of work and proof of
stake consensus algorithms: A systematic review,” J Environ
Manage, vol. 326, p. 116530, Jan. 2023, doi:
10.1016/J.JENVMAN.2022.116530.
[15] B. Johnson et al., “Cryptocurrency trading and its associations
with gambling and mental health: A scoping review,” Addictive
Behaviors, vol. 136, p. 107504, Jan. 2023, doi:
10.1016/J.ADDBEH.2022.107504.
[16] S. Quamara and A. K. Singh, “A systematic survey on security
concerns in cryptocurrencies: State-of-the-art and
perspectives,” Comput Secur, vol. 113, p. 102548, Feb. 2022,
doi: 10.1016/J.COSE.2021.102548.
[17] Y. Yue, X. Li, D. Zhang, and S. Wang, “How cryptocurrency
affects economy? A network analysis using bibliometric
methods,” International Review of Financial Analysis, vol. 77,
p. 101869, Oct. 2021, doi: 10.1016/J.IRFA.2021.101869.
[18] Y.-S. Ren, C.-Q. Ma, X.-L. Kong, K. Baltas, and Q. Zureigat,
“Past, present, and future of the application of machine learning
in cryptocurrency research,” Res Int Bus Finance, vol. 63, p.
101799, Dec. 2022, doi: 10.1016/J.RIBAF.2022.101799.
[19] P. Jaquart, S. Köpke, and C. Weinhardt, “Machine Learning for
Cryptocurrency Market Prediction and Trading,” The Journal
of Finance and Data Science, Dec. 2022, doi:
10.1016/J.JFDS.2022.12.001.
[20] G. pietro Bellocca, G. Attanasio, L. Cagliero, and J. Fior,
“Leveraging the momentum effect in machine learning-based
cryptocurrency trading,” Machine Learning with Applications,
vol. 8, p. 100310, Jun. 2022, doi:
10.1016/J.MLWA.2022.100310.
[21] Y. Wang, C. Wang, A. Sensoy, S. Yao, and F. Cheng, “Can
investors’ informed trading predict cryptocurrency returns?
Evidence from machine learning,” Res Int Bus Finance, vol.
62, p. 101683, Dec. 2022, doi: 10.1016/J.RIBAF.2022.101683.
[22] J. Wu, J. Liu, Y. Zhao, and Z. Zheng, “Analysis of
cryptocurrency transactions from a network perspective: An
overview,” Journal of Network and Computer Applications,
vol. 190, p. 103139, Sep. 2021, doi:
10.1016/J.JNCA.2021.103139.
[23] H. Lin, S. Zhang, Q. Li, Y. Li, J. Li, and Y. Yang, “A new
method for heart rate prediction based on LSTM-BiLSTM-
Att,” Measurement, vol. 207, p. 112384, Feb. 2023, doi:
10.1016/J.MEASUREMENT.2022.112384.
[24] H. Wu et al., “A novel pedal musculoskeletal response based
on differential spatio-temporal LSTM for human activity
recognition,” Knowl Based Syst, vol. 261, p. 110187, Feb.
2023, doi: 10.1016/J.KNOSYS.2022.110187.
[25] Global crypto market, “Today’s Cryptocurrency Prices by
Market Cap,” https://coinmarketcap.com/.
https://coinmarketcap.com/
[26] T. Brauner et al., “Web content mining analysis of e-scooter
crash causes and implications in Germany,” Accid Anal Prev,
vol. 178, p. 106833, Dec. 2022, doi:
10.1016/J.AAP.2022.106833.
[27] A. Garg et al., “Mechanism of biochar soil pore–gas–water
interaction: gas properties of biochar-amended sandy soil at
different degrees of compaction using KNN modeling,” Acta
Geophysica, vol. 68, no. 1, pp. 207–217, 2020, doi:
10.1007/s11600-019-00387-y.

You might also like