Height-end Multi-layer Perceptron and Machine Learning

Methods of Forecasting Bitcoin Price Time Series

This paper was downloaded from TechRxiv (https://www.techrxiv.org).

LICENSE

CC BY 4.0

SUBMISSION DATE / POSTED DATE

04-09-2022 / 07-09-2022

CITATION

Ziweritin, Stanley (2022): Height-end Multi-layer Perceptron and Machine Learning Methods of Forecasting
Bitcoin Price Time Series. TechRxiv. Preprint. https://doi.org/10.36227/techrxiv.20891200.v1

DOI

10.36227/techrxiv.20891200.v1
 1

Height-end Multi-layer Perceptron and Machine Learning

Methods of Forecasting Bitcoin Price Time Series
Stanley Ziweritin*1, Emmanuel Nwaeze2, Joshua Oluwasegun Jatto3

 and investing a lot due to the many future benefits and it could
Abstract—Bitcoin trading is now the center of attraction probably become our future currency[10],[11],[12]. The
and people are investing heavily in it. Rapidly fluctuating digital asset can still help investors make a lot of money to
prices have made few investors' millionaires and others build monumental wealth in the crypto-currency market. This
crushing and severe losses. Thousands of investors are is still possible provided you have some prior knowledge of
now very broke as Bitcoin drains their wallets. It has been future fluctuating patterns and know when to buy and sell and
nightmares for several investors who have seen their have cash to afford the risk[12],[13],[14],[15]. Virtual
money disappear before their eyes in hopes that the currency is a file stored using an application called a digital
market will recover soon. Existing methods are inefficient wallet on a smart phone or computer system and each
for predicting future data in real time. Uncertain price transaction made is stored in a list called blockchain[16],
movement can be predicted in advance to avoid future [17],[18]. Bitcoin miners receive their rewards through
losses for potential investors with greater accuracy. This verified transactions added to the blockchain list[19]. It is a
research focuses on creating high-end multi-layer digital currency or cash that uses the peer-to-peer Internet to
perceptron with other machine learning techniques to instantly confirm transactions between different users[20].
predict Bitcoin price time series. LOESS smoothing Digital money can be moved and converted into cash through
technique was applied to clearly show features of the use of third-party exchange brokers, including ATMs at a
underlying data patterns and trends before model fitting specific market value or peer-to-peer when you sell your
using an average rolling mean of 3. The adma's Bitcoin[21],[22],[23]. Bitcoin was created by Nakamoto[24] to
optimization technique was employed with Relu and be worth over $40 billion as it is traded daily. It offers low
sigmoid activation functions to update weights of the MLP transaction fees compared to traditional payment systems
neural network. It combined the best features of AdaGrad controlled by decentralized authorities and not issued by
and RMSProp; which improves the learning rate. We individuals, banks or governments. It is one of the most
adjusted the input weights with output errors to help important assets in speculation and investment. Rapidly
compute errors in the previous layer. And created KNN fluctuating bitcoin prices have made few investors millionaires
and SVM classifiers with pipeline concept for better and inflicted severe and devastating losses on others.
performance; validated and compared the performance on Thousands of investors are now very broke as Bitcoin
different fine-tuned hyper-parameter values. The SVM drains their wallets due to fluctuating investment values; in the
produced 93% accuracy, KNN gave 90%, and MLP hope that the Bitcoin market will recover soon. It has been a
recorded an accuracy rate of 96% as the best. nightmare for some investors who have seen their
money disappear before their eyes. Furthermore, existing
Index Terms—Forecast,K-nearest neighbor, multi- methods address the uncertainty and low accuracy applied to
layer perceptron, support vector machine real-time data, caused by the problem of over-matching of
models and underlying data. Price fluctuation periods can be
predicted ahead of time to avoid losses for potential
I. INTRODUCTION
investors using Bitcoon price time series smoothed using

T he Multi-layer perceptron is one of the best-known

machine learning techniques used by professionals to
solve classification and regression tasks[1],[2],[3],[4]. The
moving average. We use our machine learning experience to
build a model that can predict short, medium and long term
returns using time series in the Bitcoin market, ranging from
deep MLP neural network can also be used to predict future 2014 to 2022. This research focuses on creating high-end
Bitcoin price patterns using time series data [5],[6],[7],[8],[9]. MLP and machine learning methods for predicting Bitcoin
The virtual currency called bitcoin is gradually gaining price time series. We intend to develop a model that can
popularity and everyone including: rich and poor are tweeting predict the daily price movement of Bitcoin with
greater accuracy using price and payment networks over
 S. Ziweritin is with Akanu Ibiam Federal Polytechnic, Unwna-Afikpo, eight(8) years ranging from 2014 to 2022. The K-Nearest
Ebonyi State, Nigerian E-mail: stanlo4godsluv@yahoo.com
Neighbor (KNN) and Support Vector Machine (SVM) are the
 E. Nwaeze is with Alex Ekwueme Federal University Ndufu-Alike, Ebonyi
State, Nigerian. E-mail: nwaezeema@yahoo.com selected machine learning algorithms used to predict frequent
 J. O. Jatto is with Our Kings International School, Port Harcourt, Rivers and uncertain movements in cryptocurrency investment prices
State, Nigeria. E-mail: jattohjt@gmail.com using seasonal data or time series. Python Spyder IDE is used
.
 2

to build, train, test and forecast the fluctuating data patterns of consisting of decision tree, KNN regression and LR to predict
investment values and overcome the identified problem. three different types of future cryptocurrency price models
(XRP, Bitcoin and Ethereum) using time series data. The
This paper is divided into different sections as follows: bivariate time series method was used using the daily closing
Section 1 provides the introduction, Section 2 provides a brief prices of cryptocurrencies as a continuous variable and
overview of previous approaches to the field of study and Morgan Stanley Capital International(MSCI) for all world
research gaps of the proposed model; Section 3 presents the country indices as the predictive variable. The decision tree
materials and methods used to develop the model; Section 4 regression technique outperformed LR in terms of accuracy
focuses on the results and a detailed discussion of the results; and MAE, MSE, RMSE, and r-squared value. The model
Section 5 presents the conclusion of the article. could not learn to extract non-economic features in the
II. RELATED WORKS training and testing phases. Jaquart et al.,[33] proposed a
Saad[25] developed a support vector regression model machine learning model capable of predicting Bitcoin prices
with radial kernel functions to estimate the prices of XRP and in the short term using CNN, RF and gradient boost
Ethereum crypto-currencies. RBF's Gamma hyper-parameter classifications with time series data. The proposed model was
and polynomial kernel values have been set to 0.0001 and trained to learn and predict the feature set of the
0.00005, respectively. The R-squared(R2) of RBF provided target variable, resulting in negative returns on low-precision
78% higher performance than linear and polynomial kernel trades. Mehta & Sasikala[34] used RNN and RF techniques to
functions which resulted in 48% and 43% in exceeding the predict Bitcoin's short-term future investment prices using
variance of the dataset, respectively. The decision limit was historical data from 2015 to 2018. The prediction accuracy of
affected for smaller Gamma (γ) hyper-parameter values and RNN yielded 92% more and an error rate of 0.098, which
the model was heavily penalized for errors in misclassification performed better than the RF gave 84% with an error rate of
of the data. Agarwal and Bisht [26]; proposed the use of 0.98 after training and testing. Aygün and Kabakçı[35]
machine learning and data mining techniques to predict combined ARIMA, RNN, and CNN ML techniques to predict
the prices of market capitalization bitcoins with different the prices of Bitcoin's using a univariate time series. The
functions and attributes. The linear regression technique has window size of CNN and RNN was set to 32 with the concept
been adopted, trained and tested to efficiently calculate of trial and error method. RELU and sigmoid activation
the intercept and variance in predicting trends and future functions were used in the hidden layers. The RNN
values. The model produced prediction accuracy measures outperformed ARIMA's statistical and CNN techniques in
of 40%; it was bad and below average. Kenney[27] combined predicting bitcoin prices, but did not perform well with the
several machine learning techniques to create and predict daily opening prices, average prices, market capitalization and
chaotic Bitcoin price behaviors over time in the short, volume features included in the multivariate time series.
medium, and long term. The projected floating prices in the Pabuccu et al.,[36] proposed ANN, Naive Bayes, RF, LR and
short term were correct, but ineffective with long term prices. SVM classifiers to predict future fluctuating Bitcoin price
The accuracy of the prediction was below requirements. Jen- trends. Sigmoid, tangent, and logistic transfer functions were
Peng and Genesis[28] combined SVM, RF and deep learning used at the hidden and output layers of the ANN with multiple
techniques to estimate the return of Bitcoin's price trend. The configurations to determine the best values of hyper-
prediction accuracy of the proposed techniques varies from parameters or network settings related to the learning rate,
60% to 70% with polynomial, radial, Epanecknikov, multi- number of iterations or epochs, hidden layer neurons and
quadratic and point kernel functions. Patil[29] adopted logistic momentum. The ANN performed best and SVM provided
linear regression (LR) in machine learning and artificial the lowest metric with the discrete dataset, while the RF
neural network(ANN) techniques capable of predicting results produced the highest prediction accuracy and NB the
annual Bitcoin investment prices. The LR achieved 99.87% lowest with the continuous dataset. Jane et al.,[37] proposed
accuracy in the training phase with an error rate of 0.08 a one-day numerical solution using MLPNN, RF, ANFIS and
compared to ANN, which recorded 99.97% as the best at support vector regression to predict the prices of bitcoin
predicting Bitcoin time series data. Wu[30] Combined CNN, features using a surrogate time series (SGTS). The wavelength
KNN, and LR to predict Bitcoin's investment values based on transform function was adopted to decompose the original
inherent relationships. CNN achieved an accuracy rate of 95% time series into linear and non-linear components with
with a better multiple correlation coefficient of the R2 value projections representing feature prices. The methods of
in predicting Bitcoin price movements over time, with training deferential evolution and meta-heuristic optimization have
data being the best among LR and ANN techniques. CNN's R2 been adopted to obtain a more accurate and reliable result.
value was also better Sriramya & Reddy[31]; developed a MLPs performed better than other techniques, but did not
hybrid model consisting of linear regression(LR), K-means, work well with multivariate time series data. Dutta
SVM, Naive Bayes, CNN and RNN techniques to predict et al.,[38] proposed Gated recurrent enclosed unit and LSTM
patterns of future Bitcoin daily prices. The LASSO neural network models to predict the investment values of
outperformed the other algorithms followed by LR and NB Bitcoin feature prices which outperformed other ML
requires longer training times to predict future models due to techniques. A single layer GRU with recurring dropout was
the limited number of training and test datasets and required a introduced and trained to predict short and long term price
large dataset to improve accuracy of the forecast. trends based on a collective list of Bitcoin features. The
Alahmari[32], developed a series of nonlinear ML techniques recurring dropout enhanced the performance of GRU
 3

compared to other ML techniques with Big Data. Yassin The study strategy consists of pre-processing steps, SVM,
et al.,[39] proposed MLP-based nonlinear auto-regressive KNN, MLP model building, and model training, prediction
model with exogenous inputs to predict future bitcoin prices. and evaluation.
The PSO-based technique was used to increase the number of 3.1 Preprocessing: involves dividing or grouping proposed
input, hidden and output levels of the exogenous technique. system data into specified attributes, labels (input classes and
The Levenberg-Marquardt technique was adopted in training targets), training, and validation in a way that improves model
the MLPs and modified after 4-iterations or epoch that led to performance. We use the moving average smoothing
an over-fitting of the model. Sin and Wang[40] introduced technique to smooth out point segments with little change over
MLPNN with GA to predict the price trend of bitcoin time. It filters noisy data from randomly fluctuating points
sales with a set of 200 features. The MLPNN was employed to from the original set. A model trained with smoothed
predict the target class as a binary classification, producing an
estimates of sample data yield results with better accuracy
output of zero(0) or one(1). GA was employed to reduce than unsmoothed data, which can be done using equation 1.
generalization errors and target output generated using the A + A + …+An
Moving Average(MA) = 1 2 1
n
moving average of the MLPNN ensemble and re-trained
where A1 Bitcoin prices and "n" is the total period. It uses
to predict a day in advance on Bitcoin's prices. The MLPNN
historical data with current observations to show the trend
model produced a prediction accuracy of 53%. Kalpanasonika
of the underlying data and help provide the best fitting model.
et al.[41]; developed MLPNN with 13-neorons and trained it
We set the sliding window to small width because larger
over a period of 256 iterations to predict Bitcoin price
windows can result in overly smoothed time series which can
movements. This resulted in 99% prediction accuracy with
effectively negate the effect of the Bitcoin seasonal data trend.
specific features of network settings, but could not
LOCalized regression (LOESS) and LOCally weighted
perform well with other crypto-currencies like Litecoin,
regression(LOWESS) are used in the neighborhood points
Etherium and Ripple, etc. Ho et al.,[42] used the SLTM neural
called from the Python package statistics models and have
network with linear regression technique to predict the
been used to control the degree of smoothing with the
investment values of Bitcoin's fetaures. The Graphic user
argument called fraction which controls the accuracy of
interface was created to help user read in the 4-input
neighborhood data points in the fitting indicates the model.
features(open, high, low and close prices) and predict the next
3.2 Algorithms employed: High-end non-linear SVM, KNN
target value of bitcoin. The time series data was fitted, trained,
and MLP algorithms are used to predict the outcome of the
and tested on a limited number of items. The SLTM neural
target variable using Bitcoin's time series data. SVM is a
network produced better prediction accuracy than the LR
statistical learning technique that uses machine learning
technique, but produced low accuracy with other Bitcoin
methods in its predictions to solve regression or classification
features. Das[43] used wavelet transformers, auto-encoders,
tasks[44]. By default, the SVM produces high resistance to
and SLTM deep learning models to forecast crypto-currency
model over-fitting, but this depends on the smoothing
market values. A seq-to-seq encoder model with a tangent
parameter C. The SVM handles a regression task using a
trigger function was created to convert input words into
quadratic optimizer. It accepts input data into a high-
corresponding vectors with layers used to predict fluctuating
dimensional feature transformation using a Gaussian kernel
prices of Bitcoin's.
and performs linear regression on the transformed space. Turn
III. MATERIALS AND METHODS the above process into a nonlinear regression activity in low-
Dataset: Experimental data was obtained from a public library dimensional space, as shown in equations 1 and 2.
𝑁
made available at The dataset G = {(𝑥𝑖 , 𝑦𝑖 )}𝑖=1 2
https://www.investing.com/crypto/bitcoin/historical-dat which Given an unknown function called g(x), we can determine a
has attributes: open price, high, low and closes as shown function called f that approximates g(x) given knowledge of
in Table 1. The datset was divided into 80%(2206) training and the data set G, as shown in equation 2 below.
20%(552) testing set with attributes. f(x) = ∑i ωi ∅i (x) + b 3
Table1: Bitcoin Time dataset where ∅i(x) is the feature data while ωi and b are the estimated
Dates Open --- Volume Weight_Price variables obtained from the dataset. We proposed to use the
0 9/17/2014 465.864 - - - 21056800 460.95 loss function insensitive to Ɛ, where the error below is not
1 9/18/2014 456.86 --- 34483200 437.82 penalized. The SVM nonlinear linear regression class was
2 9/19/2014 424.103 --- 37919700 407.82 adopted to map data in a high-dimensional space using the
3 9/20/2014 394.673 --- 36863600 404.19 concept of pipelined soft voting to reduce white noise. Data
4 9/21/2014 408.085 --- 26580100 403.13 points are allowed to fit within the inner margin with
-- --- --- --- --- --- a penalty such that white noise does not occur in the training
2754 4/2/2022 46285.5 45868.95 46241.31 dataset and is controlled using the gamma (γ) parameter.
2755 4/3/2022 45859.13 --- 46453.57 46315.07 Algorithm 1: Support vector machine(SVM)
2756 4/4/2022 46445.27 --- 46622.68 46273.71 Step Processes involved
2757 4/5/2022 46624.51 --- 45555.99 46207.86 1 Start
2758 4/6/2022 45491.38 --- 45413.17 45288.27 2 Find candidate_SV with closest pair from classification
(SV=>support vector)
3 If there are violating points:
4 Find violating_points
 4

5 Compute candidate_SV= candidate_SV + nodes and the sigmoid trigger functions on the second hidden
voilating_points) layer set to 8 nodes.
6 If there is any αp< 0 due to the addition of c to S
that gives negative:
Step 2: Model compilation: we used a stochastic(Adam's)
7 Candidate_SV = candidate_SV gradient descending algorithm optimizer with the loss function
8 Repeat module to prune all data points to evaluate and search for different weights to improve the
9 end_if training model. Adam's is the best training algorithm selected
10 end_if for MLPs where data is entered one by one, neurons are
activated; data is processed and emitted through a process
The K-nearest neighbor: KNN algorithm assumes that called forward pass. The activation functions map the
similar points exist in close proximity or near to each other. weighted input to the output and adjust the threshold values of
KNN captures the idea of similarity sometimes called activated neurons to amplify the output signal. The error value
distance, proximity or closeness. We proposed a new weight is calculated to compare the target with the expected output
function called the Gaussian kernel to solve the model over- and redistribute it across the network. Network weights have
fitting problem and calculate the weight value of the Gaussian been updated for calculation error and processed for all
kernel weight function KNN as shown below in equation 4. training examples. Cross-entropy was used as the basis and
2
d(x(i) ,query) loss for the binary classification set on Keras with metrics
w i = exp (− ) 4
σ2 adjusted for accuracy.
Where w is the weight variable, 𝜎 is a parameter and query is Step 3: model fitting: The MLP is trained to perform forward
the estimation point. The dataset was split into k-folds with and back passes and iteration divided into batches. Epoch
training and test sets in the cross-validation test. The decision provides row passes of training data and the batch_size sent to
boundary of KNN is relaxed around K training data samples take the length of training data. The model learns to map
of ϰr, r=1,..,K as the closest point to the distance x *, and rank rows from input to output classifications to understand the
by majority vote among the neighboring k. We create a KNN correlation between independent and target variables[50].
model based on the training dataset. Given tabular set of Step 4: Evaluuate: We evaluated the MLPs using testing set
data(D) = (X,Y). The implementation processes of KNN with a function to pass the same input and output at training
contains the following steps state. The function returns loss, validation loss, validation
Algorithm 2: KNN accuracy and accuracy computed using:
Steps Processes involved accuracy=model.evalute(X_test, y_test) in Python where
1 Start “X_test” is input and “y_test” as target.
2 Load Dataset
3 initialize K to your chosen number of neighbors Algorithm 3: MLPs
4 FUNCTION KNN(f, X_train, K) Step Processes involved
5 L<-{(Xt, Yt), For i in range(n)} as training set of 1 Start
observations X with class Yi 2 Begin Epoch
6 Calculate d(Xt, X) for i in range(n) 3 For each training sample, d, do:
Select set of KNN training points to the query 4 Propagate input and forward through the network:
points(𝑌𝑡 , 𝑌𝑡−1 , 𝑌𝑡−2 , … , 𝑌𝑡−𝑛 ) 5 compute network output for d's input values
7 Sort ordered collection of distances and indices 6 Propagate errors backward through the network:
8 Pick first K entries from the sorted collection 7 for each network output unit
9 Get labels of the selected K entries 8 δj = Oj . (1-Oj)*(t1 - O)
9 for each hidden unit j
10 Forecast points with distance weight-weight voting
10 δj = Oj . (1-Oj)* ∑ δk . Wjk
11 Return class category
11 Update weights wj by back-propagation error using
Muti-layer perceptron neural network: MLP is a perceptron learning rule
with multiple layers that has input, output and hidden 12 Wij(New)=∆Wij + Wij(old) Where ∆Wij = π. δj. Oi
13 End Epoch
layers[45],[46],[47]. The input, hidden and output layers of
MLPs are completely interconnected. The data is sent to the In the back propagation we used output error to adjust the
model via channels weighed by the nodes[48]. The weights weights of inputs at the output layer and compute the error at
between inputs and hidden layers can be computed using the the previous layer. We repeated this process of back-
activation function(sigmoid, Relu and tanh). The forward pass propagating errors through a number of layers using the
computes outputs corresponding to the input and backward 'RELU' and sigmoid activation function.
pass propagates the error backwards through the network 3.3 Forecasting: is a technique that accepts historical data as
starting at the output units. The MLP neural network structure input to make an estimate of the direction and behavior of
has 6 inputs (date, high, low, closed, custom volume), 18 future data trends[51],[52]. It can extract short-term and long-
hidden levels and output level. The MLP structure includes: term recurring data patterns[53]. We can rewrite the time
Step 1: Model definition: We defined a sequential model to series of future virtual currency prices based on the past as:
create a layer-to-layer stacking, as required by a multilayer y=f(x) 5
perceptive neural network, and added layers one by one with Where 'y' is the observation over time as future observations
the appropriate number of input features[49]. The Relu and 'x' the input of past and present observations which can
activation function was used on the first hidden layer set to 10 virtually be transformed into the following equation:
 5

𝑌𝑡+1 = 𝑓(𝑌𝑡 , 𝑌𝑡−1 , 𝑌𝑡−2 , … , 𝑌𝑡−𝑛 ) 6

But we don't know Yt+1, the data of tomorrow reframed to
have the full dataset starting from yesterday for the testing
situation
𝑌𝑡 = 𝑓(𝑌𝑡−1 , 𝑌𝑡−2 , 𝑌𝑡−3 , … , 𝑌𝑡−𝑛−1 ) 7
Where yt is the target as output of future prices while
𝑦𝑡−1 , 𝑦𝑡−2 , 𝑦𝑡−3 , … .. are the past prices for the training
situation using a regression function.

3.4 Evaluation metrics used

This is the final step used to measure the model's performance
in terms of accuracy, standard deviation, mean score, RMSE
and receivers operating characteristic(ROC) curve and area
under the curve (AUC) to determine which is best.
IV. RESULTS AND DISCUSSIONS
The proposed system results are presented and discussed in
detail using suitable ML tools. The design and implementation
was done with some improvements to provide better and more
accurate Bitcoin prices.

Figure 3: Smoothen Bitcoin Price Time series

Figure 3 is the graph of smoothing Bitcoin time series random
variability providing lots of short and small term fluctuations
to show underlying features. It's been applied to ignore or
smoothen out small and short term fluctuations in the times
using the average rolling mean of 3 with 5% and 15% fraction
value across range of data. The 15% smoothed curve produced
a better times series than 5% with the best underlying data
features.
Figure 1: Bitcoin Price trend over the Years
Figure 1 is the graph used to provide an overview of opening,
lowest and closing price behavour(trend) of Bitcoins over time
using historical data that ranges from 2015 to the years 2022.

Figure 4: Autocorrelation
Figure 4 is the graph showing the spike corresponding to each
lag ranging from 0, 10, 20, 30, 40 to 50 and the height depicts
values of the autocorrelation function. The lag zero
autocorrelation is equal to 1 for each term and itself. To every
Figure 2: The Random White Noise spike recorded above the 0.5 line is considered to be
Figure 2 shows the presence of white noise in the time series statistically significant measured to be different from 0 and
having zero mean, equal variance, zero correlation and highly correlated to each other. Meaning that; when the price
variables which are independent. This cannot be used to make of virtual currency increases, it will tend to continue
reasonable prediction about random Bitcoon price movement increasing and when the price decreases, it will tend to
which required time series smoothing to build a better model. continue decreasing.
 6

Figure 8: The accuracy plot of support vectors

Figure 5: Partial Autocorrelation Figure 8 is the plot of SVM accuracy rate against different
Figure 5 is the partial autocorrelation graph of minimum daily fine-tuned support vectors. The prediction accuracy increased
Bitcoin price returns that resulted in white noise and recorded across different growing number of support vectors that ranges
near zero for all registered lags. This reveals that lags are from 1, 2, 3, 4, 5 to 6. The value mapped to SVM6 produced
required in building the regression analysis to have a good the highest prediction accuracy, followed by svm5, svm4, and
model. The historical and present time series(lags) are helpful svm1 gave the least. The prediction accuracy increases along
in predicting the future trend. different support vectors as we increase the degree of hyper-
parameter value.

Figure 9: The accuracy plot of KNN

Figure 6: The Year-Wise trend of Bitcoin Figure 9 depicts the KNN accuracy rate across neighborhood
Figure 6 depicts the Bitcoin yearly Box plot and there is values from 1, 2, 3, to 6 as obtained from the dataset. There is
change variability in the time series data from 2014 to 2022. a significant increase in prediction accuracy against the k-
The year 2021 produced the highest, followed by 2022, 2020 neighbor values as we increase n_Neighbor values of the KNN
and decreases from 2019, 2018 to 2014. technique. The k=6 produced 92% as the highest, followed by
k=5, k=4 to k=1 having the least prediction accuracy.

Figure 7: The Month-Wise trend of Bitcoin Figure 10: The graph of SVM and KNN Mean score
Figure 7 is the Box plot of Bitcoin month-wise trend recorded Figure 10 shows the KNN and SVM variation in mean score
a cyclic pattern obtained ranging from September to value across different k-neighbors. The means score value of
December and from January to August. SVM at k=1(0.87), 2(0.89), 3(0.90) are measured to be closer
to 1 and higher than KNN. Also the KNN mean score value at
 7

k=6(0.94), 7(0.94) outperformed the SVM model in predicting Figure 13 shows the Elbow plot of KNN RMSE value against
the target variable. k-clusters at training stage. There is a significant decrease in
RMSE value across different fine-tuned hyperparameters at K-
clusters ranging from 0.0, 2.5 to 17.5.

Figure 11: The standard deviation of KNN and SVM

Figure 11 shows the graph of KNN and SVM variation in
standard deviation across different support vectors. The
standard deviation of SVM at k=1(0.00), 2(0.00), 3(0.02),
4(0.03), 5(0.02) and 6(0.02) are all measured to be closer to 0
and smaller than the KNN. The KNN standard deviation at
k=1(0.0.05), 2(0.04), 3(0.03), 4(0.04), 5(0.03), 6(0.05) and
7(0.03) outperformed by the SVM technique.
Figure 14. The MLPs neural network structure
The MLP design has 5-input nodes, 18-hidden layers(layer-1
containing 10-nodes and layer-2 with 8-nodes) and 1-
output(high or low) visualized using Graphviz library in
Python. Each neuron compute an activation function with two
hidden layers sufficient enough to solve any problem. The
increase in hidden layers does not really affect the model
accuracy but depends on the complexity of the problem at
hand. The input data was passed to the hidden layers through
weighted channels and processed by the neuron with added
bias. The trained MLP learn correlations and dependencies
between target and independent variables. The learning
process proceeds by way of presenting network with training
set consist of input and response patterns. The validation error,
accuracy and loss value of NN was computed using the "relu",
Figure 12: The SVM plot of RMSE against episolon
and "sigmoid" activation functions at the first and second
Figure 12 depicts the RMSE plot of SVM against epsilon
hidden layers. The error values are then used to alter the
hyper-parameter value. The RMSE value decreases from 0.0
connection strength between layers in other to achieve a better
to 2.5 and spiked up and decreases down in a cyclic trend as
network response.
we finetune the epsilon hypaparameter value from 2.5 to 17.5.

Figure 15: The learning curve of MLPs

Figure 13: KNN plot of RMSE against K clusters Figure 15 is the ANN learning curve with training and testing
set. The model performance is comparable in terms of
 8

accuracy on training and testing dataset. There is no Figure 18 depict the up and down movement of Bitcoin future
consistency in the model training and testing data patterns price returns in the cause of time. There is a little deviation or
across different iterations(Epoch). variation recorded in future price fluctuations between SVM,
KNN and MLP from late 2022 to 2024.

Figure 16: The learning curve of MLPs Figure 19: The Violin plot of SVM, KNN and MLP accuracy
Figure 16 is the learning curve which decreases across Figure 19 shows the violin plot of MLP, SVM and KNN
different Epoch(iterations) with minimal gap existing between prediction accuracy. The SVM produced 93%, KNN gave
the loss values. The gap between validation and training set is 90% as the least and MLP recorded 96% accuracy as the best.
the gap of generalization which shows a learning curve with
good fit. The point of training and validation loss decreases V. CONCLUSION
towards a point of model stability.
The selected MLP's and ML techniques will help investors to
have a prior knowledge of feature price returns for the purpose
of decision making. The SVM and KNN prediction accuracy
reduces as we increase training samples in forecasting daily
price returns. The performance metrics of MLP model
increase along with more training dataset. The experimental
results MLP proved to be highly efficient and accurate in
forecasting future patterns of target variable. We therefore;
evidently conclude that the MLP performed better with good
fit.
REFERENCES
[1] M. A. Razi, and K. A. Athappilly, "Comparative Predictive
Analysis of Neural Networks (NNs), Nonlinear Regression
Figure 17: Forecasting Bitcoin Prices over time and Classification and Regression Tree (CART) Models,"
Figure 17 is the future price prediction graph using past and Expert Systems with Applications, vol.29, no.1, pp.65-74,
present data with respect to time. The historical time series 2005
price plot of MLP ranges from 2015 to early 2022 represented [2] S. P. N. Ureshbhai, P. Bhattacharya, and S. Tanwar, "KaRuNa: A
with black color and predicted future values represented with blockchain based Sentiment Analysis Framework for Fraud
red(MLP), blue(SVM) and green(KNN) color. Cryptocurrency Schemes," in Proc. IEEE Int. Conf.
Commun. Workshops (ICC Workshops), pp.1-6, Jun. 2020.
[3] T. E. Koker, and D. Koutmos, "Cryptocurrency Trading using
Machine Learning," J. Risk Financial Manage, vol. 13, no.
8, p. 178, Aug. 2020.
[4] N. A. Bakar, & S. Rosbi, Autoregressive Integrated Moving
Average (ARIMA) Model for Forecasting Cryptocurrency
Exchange Rate in High Volatility Environment: A New
Insight of Bitcoin Transaction, International Journal of
Advanced Engineering Research and Science, vol.4, no.11,
pp.130-137, 2017.
[5] D. L. K. Chuen, L. Guo, and Y. Wang, "Cryptocurrency: A new
Investment Opportunity?" J. Alternative Investments, vol.
20, no. 3, pp. 16-40, 2017.
[6] Y. B. Kim, J. G. Kim, J. W. Kim, J. H. Im, T. H. Kim, S. J. Kang,
and C. H. Kim, "Predicting Fluctuations in Cryptocurrency
Transactions Based on User Comments and Replies," PloS
Figure 18: Future Bitcoin prices One, vol. 11, no. 8, article no. e0161197, 2016.
 9

[7] P. Jay, V. Kalariya, P. Parmar, S. Tanwar, N. Kumar, and M. [25] A. A. Saad, "Predicting the Price of Cryptocurrency Using
Alazab, "Stochastic Neural Networks for Cryptocurrency Support Vector Regression Methods, Journal of Mechanics
Price Prediction,'' IEEE Access, vol. 8, pp. 82804-82818, of Continua and Mathematical Sciences," vol.15, no.4,
2020. pp.313-322, 2020.
[8] S. Tandon, S. Tripathi, P. Saraswat, and C. Dabas, "Bitcoin Price [26] A. Bisht, and P. Agarwal, "Analysis of Bitcoin using Linear
Forecasting using LSTM and 10-Fold Cross Validation,'' in Regression and Data Mining Techniques," International
Proc. Int. Conf. Signal Process. Commun. (ICSC), pp. 323- Journal of Advanced Research in Computer and
328, Mar. 2019. Communication Engineering, vol.6, no.11, pp.43-46, 2017.
[9] H. Sebastiao, "Forecasting and Trading Cryptocurrencies with [27] M. Kenney, "Investigating the Predicting of a Chaotic Time-
Machine Learning Under Changing Market Conditions,'' Series Data Using Reservoir Computing Deep-Learning and
Financial Innov., vol.7, p.3, Jan. 2021. Machine-Learning on the Short; \Medium and Long Term
[10] J. Barkatullah, and T. Hanke, "Goldstrike 1: CoinTerra's First- Pricing of Butcoin and Ethereum," Dissertation,
Generation Cryptocurrency Mining Processor for Bitcoin," Technological University Dublin, pp.1-78, 2020, doi
IEEE Micro, vol.35, no.2, pp.68-76, 2015. 10.2142//42ks-q868.
[11] M. H. Yuneline, "Analysis of Cryptocurrency's Characteristics in [28] H. Jen-Peng and S. D. Genesis, "Forecasting Bitcoin Returns: A
Four Perspectives," J. Asian Bus. Econ. Stud., vol. 26, no. 2, Data Mining Approach, Review of Integrative Business and
pp.206-219, Dec. 2019. Economics Research," vol.10, no.1, pp.51-68, 2020.
[12] M. Sivaram, E. L. Lydia, I. V. Pustokhina, D. A. Pustokhin, M. [29] P. Patil, "Bitcoin Price Prediction Using Machine Learning and
Elhoseny, G. P. Joshi, and K. Shankar, "An Optimal Least Neural Network," International Journal of Advances in
Square Support Vector Machine Based Earnings Prediction Engineering and Management(IJAEM), vol.4, no.1, pp.847-
of Blockchain Financial Products," IEEE Access, vol.8, 852, 2022.
pp.21-30, 2020. [30] Z. Wu, Predictions of Crypto-currency Prices Based on Inherent
[13] C. Leopoldo, S. Grassi, and F. Ravazzolo, "Forecasting Interrelationships, Advances in Economics, Business and
Cryptocurrencies Under Model and Parameter Instability," Management Research: International Conference on
International Journal of Forecasting vol.35, no.2, pp.485- Financial Innovation and Economic Development (ICFIED),
501, 2019. vol.648, PP.1877- 188, 2022.
[14] H. Jang, and J. Lee, "An Empirical Study on Modeling and [31] P. Sriramya, and L. S. Reddy, "A Research on Bitcoin Price
Prediction of Bitcoin Prices With Bayesian Neural Networks Prediction Using Machine Learning Algorithms,"
Based on Blockchain Information," IEEE Access, 6, International Journal of Scientific and Technology Research,
PP.5427-5437, 2018. vol.9, no.04, pp.1600-1604, 2020.
[15] S. Ha, and B. Moon, (2018). Finding Attractive Technical [32] S. A. Alahmari, "Using Nonlinear Machine Learning Algorithms
Patterns in Cryptocurrency Markets. Memetic Computing, to Predict The Price of Cryptocurrencies," International
vol.10, no.3, pp.301-306, 2018. Journal of Future Generation Communication and
[16] J. S. Heo, D. H. Kwon, J. B. Kim, Y. H. Han, and C. H. An, Networking, vol.13, no.1, pp.745-752, 2020.
"Prediction of Crypto-Currency Price Trend using Gradient [33] P. Jaquart, D. Dann, and C. Weinhardt, "Short-Term Bitcoin
Boosting," KIPS Transactions on Software and Data Market Prediction Via Machine Learning," The Journal of
Engineering, vol.7, no.10, pp.387-396, 2018. Finance and Data Science, vol.7, pp.45-66, 2021.
[17] T. Awoke, M. Rout, L. Mohanty, S. C. Satapathy, [34] P. Mehta, and E. Sasikala, "Prediction of Bitcoin Using
"Bitcoin Price Prediction and Analysis using Deep Recurrent Neural Network, International Journal of Recent
Learning Models," Springer: Communication Software Technology and Engineering(IJRETE)," vol.8, no.6,
and Networks, Pp.631–640, 2021. pp.1303- 1307, 2020.
[35]. B. Aygün , and E. G. Kabakçı, "Comparison of Statistical and
[18] W. Yiying, and Z. Yeze, "Cryptocurrency Price Analysis with
Machine Learning Algorithms for Forecasting Daily Bitcoin
Artificial Intelligence, in Proc. 5th Int. Conf. Inf. Manage.
Returns," European Journal of Science and Technology,
(ICIM)," pp.97–101, Mar. 2019.
no.21, pp.444-454 Jan. 2021.
[19] H. S. Narman, and A. D. Uulu, "Impacts of Positive and
[36]. H. Pabuccu, S. Ongan and A. Ongan, "Forecasting the
Negative comments of Social Media users to
Movements of Bitcoin Prices: An Application of Machine
Cryptocurrency," in Proc. Int. Conf. Comput., Netw.
Learning Algorithms," AIMS Quantitative Finance and
Commun. (ICNC), pp.187–192, Feb. 2020.
Economics, vol.4, no.4, pp.679–692, Nov. 2020.
[20] T. Rothman, and C. Yakar, "Empirical Analysis Towards the
[37]. R. K. Jane, I. Ghosh, and D. Das, "A diferential evolution-based
Effect of Social Media on Cryptocurrency Price and
regression framework for forecasting Bitcoin price,"
Volume," Eur. Sci. J. ESJ, vol. 15, no. 31, p. 52, Nov. 2019.
Springer Nature: Annals of Operations Research, pp.1-26,
[21] S. Garg, "Autoregressive Integrated Moving Average Model
March 2021.
based Prediction of Bitcoin Close Price," in Proc.
[38]. A. Dutta, S. Kumar, and M. Basu, A Gated Recurrent Unit
International Conference Smart System, Inventive Technol.
Approach to Bitcoin Price Prediction, MDPI: Journal of Risk
(ICSSIT), 473–478, 2018.
and Financial Management, vol.13, no.23, pp.1-16, Feb.
[22] A. Begum, A. Tareq, M. Sultana, M. Sohel, T. Rahman, and A.
2020.
Sarwar, "Blockchain Attacks Analysis and A Model to Solve
[39]. I. M. Yassin, A. Zabidi and Z. I. Rizman, "Non-linear
Double Spending Attack," Int. J. Mach. Learn. Comput.,
Autoregressive with Exogeneous Input(NARX) Bitcoin
vol.10, no.2, pp. 352-357, 2020.
Price Prediction Model Using PSO-Optimization Parameters
[23] T. Zheng, A. A. Girgis, and E. B. Makram, "A hybrid wavelet
and Moving Average technical Indicators," Journal of
Kalman Filter Method for Load Forecasting, Electric Power
Fundamental and Applied Sciences, 9(3S), pp.791-808, Sept.
Systems Research, vol.54, PP.11-17, 2000.
2017
[24] S. Nakamoto, "Bitcoin: a peer-to-peer electronic cash system, "
[40]. W. Sin, and L. Wang, "Bitcoin Price Prediction Using
Available via DIALOG. http://bitcoin.org/bitcoin.pdf.
Ensembles of Neural Networks," 2017 13th International
Accessed 21 May. 2022.
 10

Conference on Natural Computation, Fuzzy Systems and Authors Profile

Knowledge Discovery (ICNC-FSKD), 666-671, 2017. Mr. Stanley Ziweritin, is a lecturer in
[41]. K. Kalpanasonika, S. M. Sayasri, A. Vinothini, and P. H. Suga, the Department of Estate Management
Bitcoin Cost Prediction Using Deep Neural Network and Valuation, School of
Technique, International Journal of Scientific Research in Environmental Design and
Computer Science, Engineering and Information Technology, Akanu Ibiam Federal
Technology, vol.5, no.5, pp.96-101, April. 2019. Polytechnic, Unwana-Afikpo, Ebonyi
[42]. A. Ho, R. Vatambeti, S. K. Ravichandran, "Bitcoin Price State. He holds HND (Computer
Prediction Using Machine Learning and Artificial Neural Science) from Rivers (Now Kenule Benson Saro-Wiwa) State
Network Model," Indian Journal of Science and Technology, Polytechnic, Bori. PGD(Computer Science) and M.Sc(Computer
vol.14, no.27, pp.2300-2308, Aug. 2021. Science) from University of Port Harcourt(UPH) respectively.
[43]. S. Das, "Investment Using Cryptocurrency for a Machine His research interest revolves around: Artificial
Learning Based Stock Prediction Model," International Intelligence(AI)/Machine Learning(ML), Data Mining, and
Research Journal of Computer Science(IRJCS), vol.7, no.5, Visualization, Algorithms. He has published in several
pp.70-112, May. 2020. international journals. He is a registered member of Computer
Professionals Registration Council of Nigeria(MCPN).
[44] D. H. Kwon, J. B. Kim, J. S. Heo, C. M. Kim, and Y. H. Han,
"Time Series Classification of Cryptocurrency Price Trend Dr. Emmanuel Nwaeze is a senior
Based on a Recurrent LSTM Neural Network, Journal of lecturer at Alex Ekwueme Federal
Information Process System," Vol.15, No.3, pp.694~706, University, Ndufu-Alike, Nigeria. He
June 2019. holds B.Sc Mathematics(UNIJOS),
[45] N. A. Hitam, A. R. Ismail, and F. Saeed, "An Optimized M.Sc(Mathematics FUTMIN) and PhD
Support Vector Machine (SVM) Based on Particle Swarm Mathematics(UNILORIN). He has
Optimization (PSO) for Cryptocurrency Forecasting," Proc. published in several international
Comput. Sci., vol.163, pp.427–433, Jan. 2019. journals.
[46] S. Biswas, S. M. Pawar, S. Badole, N. Galande, and S. Rathod,
"Cryptocurrency Price Prediction using Neural Networks Mr. Joshua Oluwasegun Jatto, holds
and Deep Learning," in Proc. 7th Int. Conf. Adv. Comput. BSc. (Computer Science) and MSc.
Commun. Syst. (ICACCS), pp.408–413, Mar. 2021. (Computer Science) from University of
[47] A. Radityo, Q. Munajat, and I. Budi, “Prediction of Bitcoin Port Harcourt (UPH) respectively. He
Exchange Rate to American Dollar using Artificial Neural is presently working as an ICT
Network Methods,” in Proceedings of 2017 International Officer/Administrative Assistant in
Conference on Advanced Computer Science and Information Our Kings International School, Port
Systems (ICACSIS), Bali, Indonesia, pp.433-438, 2017 Harcourt, Rivers State, Nigeria. His research interest resolves
[48] I. E. Livieris, S. Stavroyiannis, E. Pintelas, T. Kotsilieris, and P. around: Software Engineering, Database Management,
Pintelas, "A Dropout Weight-Constrained Recurrent Neural Networking Security, Artificial Intelligent (AI), Expert System,
Network Model for Forecasting the Price of Major Data Communication.
Cryptocurrencies and CCi30 Index," Evolving Syst., Early
Access, 1–16, Jan.2021.
[49] M. M. Patel, Tanwar, S., Gupta, R. and N. Kumar, "A Deep
Learning-Based Cryptocurrency Price Prediction Scheme for
Financial Institutions," Journal of Information Security
Applications, vol.55, Art. no.102583, Dec.2020..
[50] A. Saxena, and A. Sukumar, "Predicting Bitcoin Price using
LSTM And Compare its Predictability with ARIMA
Model," International Journal of Pure and Applied
Mathematics," vol.119, no.17, pp.25912600, 2018.
[51] E. A. D. Oliveira, I. E. Zarate, M. D. A. Reis, and C. N. CNobre,
"The Use of Artificial Neural Networks in the Analysis and
Prediction of Stock Prices," in IEEE International
Conference on Systems, Man, and Cybernetics, 2011.
[52] G. Gomes, L. Dias, and M. Correia, "CryingJackpot: Network
Flows and Performance Counters Against Cryptojacking," in
Proc. IEEE 19th Int. Symp. Netw. Comput. Appl. (NCA),
pp.1-10, Nov.2020..
[53] I. A. Hashish, F. Forni, "Andreotti, G., Facchinetti, T. and
Darjani, S. "A hybrid model for bitcoin prices prediction
using hidden Markov models and optimized LSTM
networks," in Proc. 24th IEEE Int. Conf. Emerg. Technol.
Factory Autom. (ETFA), pp.721-728, Sept.2019.