CRYPTOCURRENCY MARKET DYNAMICS: A MACHINE

LEARNING-BASED APPROACH FOR PRICE PREDICTION

Dr. D.Ganesh Mr. Vemula Chandu Harshavardhan
Associate Professor of CSE, PG Scholar, Department of CA,
Mohan Babu University Mohan Babu University
(Erstwhile Sree Vidyanikethan Engineering (Erstwhile Sree Vidyanikethan Engineering
College (Autonomous), Tirupati, College (Autonomous), Tirupati,
Andhra Pradesh, India Andhra Pradesh, India
[email protected] [email protected]

ABSTRACT trend. The market valuation of Bitcoin increased

dramatically from approximately USD18 billion in
Now-a-days Crypto currencies are digital currency they
early 2017 to almost USD600 billion by the year's end.
are suitable for usage as long-term investments or for all
kinds of transactions. Investor interest in it is high
From its humble beginnings as a retail investment,
because of its fundamental values of security, Bitcoin has grown into a model for other digital
transparency, and decentralization. Developing currencies like Ethereum, XRP, and Litecoin, among
successful investing strategies requires an accurate price many others[3]. The most valuable crypto currency in
prediction of the crypto currency due to its unique the world, bit coin is traded on more than 40 exchanges
qualities and price volatility. The price of Bitcoin, the throughout the globe that accept more than 30
most popular crypto currency, is predicted using this different currencies. Its market value is currently $9
model. A multitude of factors will be considered in order billion USD, and more than 250,000 transactions
to forecast the accurate prices. The change point occur there every day. Because of its youthful age and
detection technique is used to provide robust prediction
performance in the unseen price range. Furthermore,
thus high price fluctuation that is far greater than that
unique records that exist in crypto currencies and are of other digital currencies Bitcoin presents a unique
posted on the blockchain, or "on-chain data," are potential for price prediction as a currency. It differs
gathered and used as input variables to forecast values. from conventional fiat currencies that is open-ended;
Experiments with real world Bitcoin price data and with fiat currencies, there isn’t comprehensive data on
various methods have proven the proposed system’s cash transactions or the amount of money in
effectiveness in Bitcoin price prediction. Cryptocurrency circulation. Long-term financial market prediction,
price trading is a huge demand of exchange right now. like that of stock market, has
It’s a great advantage for investors and day-to-day
been studied extensively.
traders. In this work we will use latest Machine Learning
algorithms for finding out the Price Prediction of Crypto Conventional time series prediction techniques, such
currency. These algorithms help in determining price the exponential smoothing models developed by Holt-
prediction. For accurate findings, these techniques are Winters[4], rely on linear assumptions and work best
compared against each other. with data that can be segmented into trend,
seasonal, and noise components. Regarding contracts
Keywords: Cryptocurrency, Bitcoin, Blockchain, On- like sales forecasting, where seasonal impacts are
chain-data, Point detection technique, Decentralization evident, such type of methodology works better.

1. INTRODUCTION
Can financial, crypto, and macroeconomic shocks
affect Bitcoin's price? Is the idea of an efficient market
something that Bitcoin should adhere to? Does the
volatility of the price of Bitcoin being influenced by
other crypto currencies [1] In 2009, Bitcoin was born
as initial cryptocurrency, drawing in new investors
with its promise of big returns. The revenue of bitcoin,
as reported by Coinbase, have increased by over
twelvefold between 2016 and 2017, hitting U.S.
dollars20,000 from dollars900 for purchasing of one Figure1.Bitcoin
Bitcoin token [2]. This clearly demonstrates this
The Remaining article is set up as per. Portion 2 Considering fees, realized an average gain of 117%
Provides the Literature survey of Latest research on total. reliable in forecasting significant meme coin
crypto currencies. Section 3 Provides the Proposed price changes. Every execution of the trading
Methodology and Algorithm for detecting the Fake algorithm uses all of the available capital.
News. Section 4 gives the details and processing. And
Section 6 provides the outcomes and last section Huiru Xu, Binwei Chen, Changqing Luo, and Lurun
concludes the work. Pan (2022)[8] developed a model to predict the price
2. Related Works of Bitcoin with deep learning techniques and
multiscale data. Original Bitcoin price series broken
There are several uncertainty and high fluctuation in down into components with varying frequencies: high,
the crypto currency market, which has attracted a medium, and low. utilizing deep learning techniques
multiple attention recently. Investors, traders, and to improve model performance. components are
other players in the financial sector have relied heavily reduced and intrinsic mode functions (IMFs) are
on accurate price predictions of crypto currencies in rebuilt. combination of disassembled parts by means
this context. A new line of inquiry into the reliability of the sample entropy technique.G. Kim, J.
of crypto currency price forecasts is the G. Choi, S. Lim, and D.-H. Shin (2022)[9] The
implementation of deep neural learning and machine suggested system is a model for predicting bitcoin
learning algorithms. Here, we survey the relevant prices based on deep learning. In order to forecast
literature, drawing attention to the popular methods Bitcoin (BTC) prices, it employs on-chain data as
and difficulties in the field of crypto currency price
input variables. For consistent prediction performance,
forecasting.
the system use of a change point detection technique.
A system utilizing LSTM and random forest It utilizes the multiple long short-term memory (SAM-
regression was proposed by Chen, Junwei (2023)[5]. LSTM) self-attention-based paradigm. Experiments
It seeks to forecast Bitcoin's price for the upcoming have shown that this method is effective at predicting
day. Compared to LSTM, random forest regression the price of bitcoin.
offers better prediction errors (RMSE and MAPE). B. Amiri and F. Feizian, (2023) [10] suggested a
Bitcoin prices are influenced by various factors, model that uses sentiment analysis to forecast
including the price of ETH, oil, and US stock market cryptocurrency prices in the future. Machine learning
indices. Since 2018, the price of ETH and the JP225 methods used to data analysis of time series. Weights
index of the Japanese stock market have been depending on influence factor are applied to tweet
significant factors. The most accurate model for sentiment scores. hybrid model that combines
prediction is the one that uses a single lag for its weighted sentiment scores with text analysis elements.
explanatory variables. outperformed other models in terms of Cardano, EOS,
To be capable of forecast Bit coin fluctuations, and Ethereum price prediction. For Bitcoin and Ripple
Dimitria dou, Athan asia, and Andros Gregoriou predictions, weighted sentiment scores based on
(2023) [6] suggest a machine-learning architecture. influence factor fared better than other models[11-16].
Exchange rates, interest rates, and macroeconomic
factors are among the twenty-four variables in the 3. Methodology:
dataset. The random forest technique, logistic Machine learning is a critical department of
regression, and support vector machine are the three manufactured insights (AI). It may be divided into
designs which are employed. With a 66% accuracy three categories: sponsored education, unsupervised,
rate, compared to all other examples, the one based on and reinforced learning, depending on whether a goal
logistic regression scores better. The analysis variable is present. To be able to predict future Bitcoin
recommends that the Bitcoin market reject weak form prices, a relapse method using administered learning is
efficiency. Changing Luo Paolo Trunfio (2023) [7] employed this investigation. According to machine
suggested process for recommending the timing of learning's cohesive execution theory, after a preset
cryptocurrency purchases and sales. integrates deep computation, a learner is formed. To obtain a high-
learning, text analytics, and statistical methods. takes precision learner, the learner is prepared again using
into account the connection involving price changes training data and the review procedure is repeated. The
and posts on social media. considers the causal last step is to hand over the test results to the
relationship between prices and social media users'
mood. any transaction fees, had an average gain of
194% total.
student, who can then use them for their own work.
Both the unpredictable woodland recurrence and the
LSTM demonstration preparation are implemented in
this paper using the open-source Python machine
learning library. Arbitrary Woodland Relapse uses sky
learn as its library and LSTM uses keras for research.
The Pandas package handles the information
collection and pre-processing.

3.1. Random Forest:

One type of complex regression tree is the Random
Forest. Its comprehensibility is strong, but the
preparation tests limit the expected results. Regression
tree rules state that the parent group should be divided
into subgroups using a reference to a specific variable.
The categorization is based on finding the least normal
of all the residuals divided by squared that gather, with
Condition appearing beneath. Fig 3:The random forest regression framework and parameters

The standard DNN algorithm and the RNN algorithm

are not the same. When input is substituted into the
model, the model's parameters are altered in addition
to producing an output value. The purpose of RNN
algorism is to keep the model's prior input data
information. In order to compensate for the short
memory issue of RNN, this work adopts the LSTM
model. The three activation processes that enable data
modifications regarding RNN architecture and the
storage model are recall entrance, entry gate, and
output gate.

This experiment's LSTM model structure centers on

the property that a further level of the LSTM
framework may be built by employing the drop out
layer that has been suggested in previous studies and
substituting the model's output value. Regarding the
Figure2. Parameter settings
dropout layer's parameter settings, I experimented
Figure 2 illustrates the parameter settings. It indicates with [less than = 10%, peak = 50%] for every layer in
that the random forest has 500 sub trees for regression, dropout. Additionally, the investigations concluded
with an optimal depth of 10 for a single mini that the dropout value may be predicted more
regression tree. The deepest point [least = 3, maximum accurately when done in ascension order as opposed to
= 20] & the greatest amount of sub-regression trees reverse order. After balancing the risk of over learning
[least = 200, maximum = 1000] of the interval were against accuracy, the settings of every single layer of
tested by me. The information utilized to train with the modules of [32, 64, 128, 256, 512] with the total
mistake in predictions stop to fluctuate when the total amount of layers for each LSTM [The minimum = 2,
amount of sub-regression trees approaches 500 or the highest = 6] are examined. Each layer's activation
depth limit exceeds 10, according to my subsequent function is set to "ReLU," which is more effective than
testing. "sigmoid" and "tanh.". Figure. 4 below shows the
LSTM structure and particular value. The last 10% of
the sample set is made up of the validation
information.
 Figure 5: Proposed Methodology

Figure 4: Parameters and framework of LSTM 4. Data and Pre processing:

Bagging Technique & Long-Short Term Memory From March 31, 2015, to April 1, 2023, daily data
algorithms are used to predict the rate of BTC on the were used as sample data. Bitinfocharts.com,
other day with accuracy and efficiency. Compile coinmatrics.io, investing.com, Coinmarketcap.com,
historical Bitcoin price data with relevant criteria such and Yahoo Finance were the sources of the study's
as difficulty, mean block size, mining rate, location, data. For this experiment, Bitcoin’s cost in US dollars
and reasonable transfer size, and divide it into sets for
is the target variable. To forecast the Bit coin’s price
testing and training. Extrapolate relevant data features on the future,47 variables in all these had been split
that may create an effect on the price of bit coin and
onto eight categories are employed as variables for
scale or normalize them as needed. Learn the Bagging
explanation. These variables include cost variables for
Technique using the data set used for training, in
bit coin, technical aspects of the crypto currency,
addition to the previous features and Bitcoin prices.
commodities, additional digital assets, foreign
Then, use the same training dataset to train an LSTM
exchange, stock index, attention from the general
network. In terms of locating sequential dependencies
public, and week's simulated variables.
in time-series data, such stock prices, LSTM is quite
useful. After training two models, apply the Bagging 4.1. Pre processing:
Technique to forecast using the attributes, and after Between March 31, 2015, and April 1, 2023, or seven
that, employ the test datasets to see how well the natural years in total, data were gathered for the data
hybrid model works. Statistics such as MAPE stands research sample. Nonetheless, given that Bitcoin is
for Mean Absolute Percentage Error, Mean Square unique in that it experienced two price bubbles at the
Error (MSE), and others might be utilized depending close of 2017 and 2022, and previous research has
on your preferences and the characteristics of your shown that a single sample's permitted span of time is
data. Finally, enter the most recent information into four years in total. In accordance with the two factors,
every approach to forecast the price of Bit coins the each piece of info is divided into two phases described
next day. Then, integrate the projections from each above in an attempt to create the model's price
model to get the final prediction. Figure 5 depicts the predictions more accurate: First, from the tail end of
overall experiment flow. March 2015 all the way to the tail end of September
2018, and then, from October 1, 2018, to the beginning
of October 1, 2018, there is going to be Phase 2. Make
predictions after independently analyzing two
subsamples and training models for corresponding
time periods. During machine learning, training
examples are used to instruct initial samples, which are
then substituted with test samples for evaluation.
Most of those samples (75–90%) are used for training
purposes. Like observed in Figure 5, the study's testing
and training samples were precisely divided.
Validation data makes up the final 10% of the original
set.

Figure 6. Trained collection and test period separation Fig 7: Considering an analysis of the real and anticipated prices
using a random forest analysis
5. Results
Any of the info from a particularly current time frame
The test samples from Periods one and two are must be used to train framework, based on the findings
predicted by the trained learner, yielding the outcomes on the quantity of data periods needed. This finding is
displayed in Figure 6 below. The Btc exp is shown as related to the theory of the efficient market. The asset's
the red line, while the price that the random forest future price is anticipated by the market in its current
regression learner predicted is shown as the green price, and the rate from a prior era is not relevant. Even
dashed line. Comparing the RMSE findings of though when it comes to estimating cost of the digital
different periods is pointless, even though Period 1's currency, the random forest regression approach has
RMSE is significantly lower than Period 2's because surpassed LSTM in order by a significant margin.
Period 1's average Bitcoin price is also significantly Despite the reality because that is not statistically
lower than Period 2's. The two periods' MAPE and DA superior than LSTM. This technique has not been
indicators are quite similar, and Period 2's forecast extensively discussed in previous research.
accuracy is marginally higher than Period 1's. Because
there aren't many samples in the training intervals of
Phase two with Bitcoin prices higher than USD
60,000, it's important to note that the random forest
regression algorithm performs poorly in the very
beginning of the taught interval of Period two
regarding predicting the price of the crypto currency.
5.2. Results of LSTM:
The LSTM algorithm is a technique for deep learning
that has good prediction performance for time series
data. The answer to the question of whether it is
necessary to consult data from many periods
previously when projecting the expenditure of bit
coins demonstrates a model that just needs knowledge
from earlier times yields the best prediction accuracy.
As observed through the findings of the inaugural
period (a ten-variable strategy) along with the
subsequent period in Figure 7 below, the price patterns
of both approaches are quite similar to the actual price. Fig 8: Precision and the certain no of lapses (LSTM), among in
Conversely, though the smoother the forecasting curve Relation.
gets plus how considerably it departs from the true
cost, the more data periods are incorporated to the
model.
 6. Conclusion: 7. P. K. Roy, A. Kumar, A. Singh and A. K. Sangaiah,
"Forecasting Bitcoin Prices using Deep Learning for
To foresee the value of bitcoins for the upcoming Consumer Centric Industrial Applications," in IEEE
Transactions on Consumer Electronics,2023.
moment, the following categories of several factors 8. Pintelas E, Livieris IE, Stavroyiannis S, Kotsilieris T, Pintelas
served as illustrative parameters in this paper: (1) P. “Investigating the Problem of Cryptocurrency Price
influences affecting the price of Btc; (2) the Prediction: A Deep Learning Approach”. Artificial
characteristics of Btc technology; (3) different Intelligence Applications and Innovations. 2020 May
6;584:99–110.
currency; (4) merchandise; (5) foreign currencies; (6) 9. Sebastião, H., Godinho, P. Forecasting and trading crypto
recognition by the public; and (7) week’s false currencies with machine learning under changing market
variables. It consisted of 8 groups in all, totaling 47 conditions. Financ Innov 7, 3 (2021).
entries. Bagging technique operates far superior to 10. Baur, Dirk G., and Lai Hoang. 2021. The Bitcoin gold
correlation puzzle. Journal of Behavioral and Experimental
RNN for rate analysis for forecasting. LSTM was a Finance 32: 100561.
popular and well-recognized framework of great 11. Chen, Wei, Huilin Xu, Lifen Jia, and Ying Gao. 2020a.
accuracy for finding BTC value in earlier studies. The Machine learning model for Bitcoin exchange rate prediction
method of machine learning termed as random forest using economic and technology determinants. International
Journal of Forecasting 37: 28–43.
modeling regression, it hasn't been used widely in the 12. Jaquart, Patrick, David Dann, and Christof Weinhardt. 2021.
field by other academics previously, is used in this Short-term bitcoin market prediction via machine learning.
work to provide a result that is more accurate in The Journal of Finance and Data Science 7: 45–66.
predicting than LSTM layer. Despite this, one major 13. Davanam, Ganesh, T. Pavan Kumar, and M. Sunil Kumar.
"Efficient energy management for reducing cross layer attacks
limitation of Bagging technique that it cannot forecast in cognitive radio networks." Journal of Green Engineering 11
outcomes that were absent from the training sets. For (2021): 1412-1426.
instance, random forest regression was unable to 14. Davanam, Ganesh, T. Pavan Kumar, and M. Sunil Kumar.
produce a price result higher than the previous "Novel Defense Framework for Cross-layer Attacks in
Cognitive Radio Networks." International Conference on
historical high moment when the historical high for Intelligent and Smart Computing in Data Analytics. Springer,
Bitcoin's price was broken. However, given the Singapore, 2021.
growing history of Bitcoin transactions, I believe 15. Walde, R.B., Prakash, O., Kumar, M.S., Sule, B., Nabilal,
random forest regression will outperform it once the K.V. and Pawar, U.C., 2023. Ranking Based Security Scheme
with Attack Detection and Improved Network Security.
price of Bitcoin stabilizes. A part from each model's International Journal of Intelligent Systems and Applications
reliability final analysis, research findings in addition in Engineering, 11(3s), pp.35-41
revealed the model’s correctness in prediction declines 16. Yadav, S.S., Maan, M.K., Kumar, M.S., Kumarnath, J., Pund,
when the amount increases of previous occurrences for S.S. and Rathod, M., 2023. A Secure IoT Smart Network
Model for the Contributory Broadcast Encryption for the Text
the substituted explicating factors, regardless of Policy Management Scheme. International Journal of
whether this algorithm used is LSTM or random forest Intelligent Systems and Applications in Engineering, 11(3s),
regression. The most accurate illustration is the one pp.42-48.
that only substitutes explaining factors from the prior
era. This result is in line using the customary efficient
market theory.

REFERENCES:
1. Dimitriadou, A.; Gregoriou, A. Predicting Bitcoin Prices
Using Machine Learning. Entropy 2023, 25, 777.
2. Belcastro, L.; Carbone, D.; Cosentino, C.; Marozzo, F.; Trunfio,
P. “Enhancing Crypto currency Price Forecasting by Integrating
Machine Learning with Social Media and Market Data”.
Algorithms 2023, 16, 542.
3. Changqing Luo, Lurun Pan, Binwei Chen, Huiru Xu, "Bitcoin
Price Forecasting: An Integrated Approach Using Hybrid
LSTM-ELM Models", Mathematical Problems in Engineering,
vol. 2022, Article ID 2126518, 17 pages, 2022.
4. Laura Alessandretti, Abeer ElBahrawy, Luca Maria Aiello,
Andrea Baronchelli, "Anticipating Cryptocurrency Prices
Using Machine Learning", Complexity, vol. 2020, Article ID
8983590, 16 pages, 2023.
5. G. Kim, D. -H. Shin, J. G. Choi and S. Lim, "A Deep Learning-
Based Cryptocurrency Price Prediction Model That Uses On-
Chain Data," in IEEE Access, vol. 10, pp. 56232-56248, 2022.
6. F. Feizian and B. Amiri, "Cryptocurrency Price Prediction
Model Based on Sentiment Analysis and Social Influence," in
IEEE Access, vol. 11, pp. 142177-142195, 2023.