Topic : Bitcoin Price Prediction Using Machine Learning Algorithms

ABSTRACT
After the boom and bust of cryptocurrencies’ prices in recent years, Bitcoin has been
increasingly regarded as an investment asset. Because of its highly volatile nature, there is a
need for good predictions on which to base investment decisions. Although existing studies
have leveraged machine learning for more accurate Bitcoin price prediction, few have
focused on the feasibility of applying different modeling techniques to samples with
different data structures and dimensional features. To predict Bitcoin price at different
frequencies using machine learning techniques, we first classify Bitcoin price by daily price
and high-frequency price. A set of high-dimension features including property and network,
trading and market, attention and gold spot price are used for Bitcoin daily price prediction,
while the basic trading features acquired from a cryptocurrency exchange are used for 5-
minute interval price prediction. Statistical methods including Logistic Regression and
Linear Discriminant Analysis for Bitcoin daily price prediction with high-dimensional
features achieve an accuracy of 66%, outperforming more complicated machine learning
algorithms. Compared with benchmark results for daily price prediction, we achieve a better
performance, with the highest accuracies of the statistical methods and machine learning
algorithms of 66% and 65.3%, respectively. Machine learning models including Random
Forest, XGBoost, Quadratic Discriminant Analysis, Support Vector Machine and Long
Short-term Memory for Bitcoin 5-minute interval price prediction are superior to statistical
methods, with accuracy reaching 67.2%. Our investigation of Bitcoin price prediction can be
considered a pilot study of the importance of the sample dimension in machine learning
techniques.

CHAPTER ONE
INTRODUCTION

Bitcoin is a crypto currency used worldwide for digital payment or simply for investment
purposes. Bitcoin is decentralized i.e. it is not owned by anyone. Transactions made by
bitcoins are easy as they are not tied to any country. Investment can be done through various
marketplaces known as “bitcoin exchanges.” These allow people to sell/buy bitcoins using
different currencies. The largest bitcoin exchange is Mt Gox. Bitcoins are stored in a digital
wallet which is basically like a virtual bank account. The record of all the transactions, the
timestamp data is stored in a place called Blockchain. Each block contains a pointer to a
previous block of data. The data on blockchain is encrypted. During transactions the users
name is not revealed, but only their wallet ID is made public.
Predicting the future is no easy task. Many have tried and many have failed. But many of us
would want to know what will happen next and would go to great lengths to figure that out.
Imagine the possibilities of knowing what will happen in the future! Imagine what you
would have done back in 2012 when Bitcoin was less than $15 knowing that it would
surpass $18,000! Many people may regret not buying Bitcoin back then but how were they
supposed to know in the first place? This is the dilemma we now face in regards to
Cryptocurrency. We do not want to miss out on the next jump in price but we do not know
when that will or will not happen. So how can we potentially solve this dilemma? Maybe
machine learning can tell us the answer.
Machine learning models can likely give us the insight we need to learn about the future of
Cryptocurrency. It will not tell us the future but it might tell us the general trend and
direction to expect the prices to move. Let’s try and use these machine learning models to
our advantage and predict the future of Bitcoin by coding them out in Python!

Existing System

After the boom and bust of crypto currencies’ prices in recent years, Bitcoin has been
increasingly regarded as an investment asset. Because of its highly volatile nature, there is a
need for good predictions on which to base investment decisions. Although existing studies
have leveraged machine learning for more accurate Bitcoin price prediction, few have
focused on the feasibility of applying different modeling techniques to samples with
different data structures and dimensional features. To predict Bitcoin price at different
frequencies using machine learning techniques, we first classify Bitcoin price by daily price
and high- frequency price. A set of high-dimension features including property and network,
trading and market, attention and gold spot price are used for Bitcoin daily price prediction,
while the basic trading features acquired from a cryptocurrency exchange are used for 5-
minute interval price prediction.

Proposed System

Statistical methods including Logistic Regression and Linear Discriminant Analysis for
Bitcoin daily price prediction with high-dimensional features achieve an accuracy of 66%,
outperforming more complicated machine learning algorithms. Compared with benchmark
results for daily price prediction, we achieve a better performance, with the highest
accuracies of the statistical methods and machine learning algorithms of 66% and 65.3%,
respectively. Machine learning models such as Long
Short-term Memory for Bitcoin 5- minute interval price prediction are superior to statistical
methods, with accuracy reaching 67.2%. Our investigation of Bitcoin price prediction can
be considered a pilot study of the importance of the sample dimension in machine learning
techniques.

Problem Statement
The highly volatile nature of Bitcoin prices presents a significant challenge for investors
seeking to make informed decisions regarding their investments. While machine learning
techniques have shown promise in predicting financial markets, few studies have focused
specifically on Bitcoin price prediction, particularly considering different data structures and
dimensional features. Therefore, there is a need to explore the feasibility of applying various
modeling techniques to Bitcoin price prediction across different sample dimensions and
frequencies. The problem at hand is to develop a robust Bitcoin price prediction system
integrated with a user-friendly graphical user interface (GUI) using machine learning
algorithms..

Aim and Objectives

The aim of this project is to develop machine learning algorithms and a graphical user
interface (GUI) for accurately predicting the price of Bitcoin, leveraging technical indicators
as inputs. By harnessing the power of machine learning and integrating user-friendly
visualization through a GUI.

Objectives

1. Identification of ML Algorithms: Investigate and identify machine learning algorithms

suitable for the prediction of Bitcoin prices based on historical data and technical indicators.
This objective aims to explore a range of algorithms to determine the most effective ones for
price prediction.

2. Comparison of ML Algorithms: Compare the predictions generated by different machine

learning algorithms using various evaluation metric scores, such as Mean Absolute Error
(MAE), Mean Squared Error (MSE), and Root Mean Squared Error (RMSE). This objective
seeks to assess the performance and accuracy of each algorithm in predicting Bitcoin prices.

3. Impact Analysis of Technical Indicators: Analyze the influence of individual technical

indicators on the prediction performance of each machine learning model. By examining the
significance and effect of various technical indicators, this objective aims to identify key
factors driving Bitcoin price movements and their impact on prediction accuracy.

Scope and Limitations

Scope

This project aims to identify and compare machine learning algorithms for Bitcoin price
prediction using technical indicators as input features. It will explore the effectiveness of
various algorithms in forecasting Bitcoin prices and analyze the influence of different
technical indicators on prediction accuracy.

Limitations

● The study will focus solely on technical indicators as input features for machine
learning models, potentially limiting the scope of analysis.
● The accuracy of Bitcoin price predictions may be affected by factors beyond the
scope of this study, such as market sentiment and regulatory changes.
● The availability and quality of historical price data may vary, potentially impacting the
performance of machine learning algorithms.

Certainly! Here's the significance of your study:

---

Significance of the Study:

The "Bitcoin Price Prediction System" project holds several significant implications:
1. Enhanced Decision-Making: By leveraging machine learning algorithms and technical
indicators, the system provides traders and investors with valuable insights into the future
price movements of Bitcoin. This enables them to make more informed decisions regarding
buying, selling, or holding Bitcoin assets, ultimately improving their trading strategies and
investment portfolios.

2. Risk Management: Accurate price predictions facilitate better risk management strategies,
allowing individuals and organizations to mitigate potential losses and maximize returns in
the highly volatile cryptocurrency market. Understanding the impact of technical indicators
on prediction accuracy enables users to adjust their risk exposure accordingly.

3. Technology Adoption: The project demonstrates the application of cutting-edge

technologies, including machine learning and graphical user interface (GUI) development, in
the domain of financial forecasting. By showcasing the integration of advanced algorithms
with user-friendly interfaces, it promotes the adoption of innovative solutions for addressing
complex financial challenges.

4. Educational Value: The project contributes to the academic and research communities by
exploring the effectiveness of various machine learning algorithms for cryptocurrency price
prediction. Findings from the comparative analysis of algorithms and technical indicators
provide valuable insights into the factors influencing Bitcoin price dynamics, paving the way
for further studies in this area.

Research Questions

RQ1: Which ML algorithms are better suited to predict the prices of Bitcoin?

RQ2: Which among the selected ML algorithms performs the best in predicting the
price of Bitcoin when using technical indicators RSI, EMA, SMA as input, and which
among these indicators affects the prediction performance of each ML algorithm the
most?

Operational Definition of Terms

1. Bitcoin Price Prediction System: A software system developed using machine learning
algorithms and a graphical user interface (GUI) that forecasts the future prices of Bitcoin
based on historical data and technical indicators.

2. Machine Learning Algorithms: Computational algorithms used to analyze historical

Bitcoin price data and identify patterns or trends that can be used to predict future price
movements. Examples include linear regression, decision trees, random forests, support
vector machines (SVM), and neural networks.

3. Graphical User Interface (GUI): A user interface that allows users to interact with the
Bitcoin price prediction system through visual elements such as buttons, menus, and charts.
The GUI provides an intuitive way for users to input parameters, view predictions, and
interpret results.

4. Technical Indicators: Quantitative data points derived from historical Bitcoin price and
volume data, as well as other market variables, used to analyze and forecast future price
movements. Examples include moving averages, relative strength index (RSI), stochastic
oscillators, and MACD (Moving Average Convergence Divergence).

5. Prediction Accuracy: A measure of how close the predicted Bitcoin prices generated by
the machine learning algorithms are to the actual observed prices. Accuracy is typically
assessed using evaluation metrics such as Mean Absolute Error (MAE), Mean Squared Error
(MSE), Root Mean Squared Error (RMSE), or prediction intervals.

6. Risk Management: Strategies and techniques employed to identify, assess, and mitigate
risks associated with Bitcoin trading and investment. In the context of the project, risk
management involves using accurate price predictions to make informed decisions about
portfolio allocation, position sizing, and trade execution.

7. Validation Metrics: Quantitative measures used to assess the performance of machine

learning algorithms in predicting Bitcoin prices. These metrics, such as MAE, MSE, and
RMSE, provide objective criteria for comparing the effectiveness of different algorithms and
selecting the most suitable ones for the prediction task.

8. Feature Engineering: The process of selecting, transforming, or creating new features

from the raw Bitcoin price and technical indicator data to improve the performance of
machine learning models. Feature engineering aims to capture relevant information and
patterns that contribute to more accurate price predictions.

CHAPTER TWO

LITERATURE REVIEW

Numerous comparative studies exist for predicting the price of cryptocurrencies, employing
various methods. Some researchers have utilized deep learning (DL) models, while others
have focused solely on machine learning (ML) techniques. Additionally, studies have
incorporated different cryptocurrencies such as Bitcoin, Ethereum, and Litecoin, and have
explored both short-term and long-term prediction strategies. The following papers provide
insight into the current research landscape:

Phaladisailoed et al. 2019 employed 1-minute interval trade data from January 1, 2012, to
January 8, 2018, to predict Bitcoin values using LSTM, GRU, Theil-Sen, and Huber regression
models. The study revealed that GRU and LSTM outperformed other regression models,
with GRU achieving the highest accuracy. However, the evaluation solely relied on R2
metrics, limiting the assessment of nonlinear relations between variables.
Seabe et al. 2013 utilized three types of RNN, including LSTM, GRU, and Bi-Directional LSTM,
to predict the prices of Bitcoin, Ethereum, and Litecoin. The Bi-LSTM model demonstrated
superior prediction accuracy, measured by Mean Absolute Percentage Error (MAPE) and
RMSE. However, the study solely compared deep learning algorithms and evaluated models
using RMSE and MAPE metrics.

Vanderbilt et al. 2021 tested three cryptocurrencies—Bitcoin, Litecoin, and Ripple—using

Simple RNN, LSTM, and GRU models. Historical price data from coinmarketcap.com were
utilized, with additional experiments incorporating Google Trends data. The study concluded
that all three procedures yielded comparable accuracy levels, with no significant
improvement observed with the inclusion of Google Trends data. Only RMSE statistics were
used for evaluation.

Kim et al. 2020 evaluated LSTM and GRU deep learning algorithms on Bitcoin, Ethereum,
and Litecoin datasets based on kurtosis and skewness criteria. Both models were trained
and tested with varying epochs, and accuracy was measured using RMSE and MAE. The
study concluded that GRU was advantageous for downward stabilization trends, while LSTM
was suitable for upward stabilization trends.

Iqbal et al. 2013 compared ARIMA, FB Prophet, and XGBoosting techniques for predicting
Bitcoin prices using RMSE, MAE, and R2 parameters. ARIMA exhibited the best performance
among the three methods. The study focused on supervised machine learning algorithms
and utilized historical price data from the Kaggle dataset.

Kumar et al. 2009 employed MLP and LSTM deep learning techniques to predict Ethereum
cryptocurrency trends. Daily and hourly data from August 2015 to August 2018 were
utilized, with evaluation based on RMSE, MAE, and other metrics. LSTM demonstrated
superior accuracy and robustness for long-term predictions.

Derbentsev et al. 2008 focused on short-term price prediction using the updated Binary
Auto Regressive Tree (BART) model, which outperformed ARIMA. The study utilized log
return series data from January 2017 to March 2019 and considered SMA, EMA, and RSI
values as target variables.

Chen et al. 2023 employed ML techniques to predict Bitcoin prices using regular and high-
frequency price data. XGB and SDA algorithms demonstrated higher accuracy compared to
benchmark results. The study utilized aggregated daily price data from CoinMarketCap.com
and 5-minute interval trading data from Binance, considering features such as block size and
hash rate.

These studies contribute to the understanding of cryptocurrency price prediction methods

and highlight the effectiveness of various algorithms and techniques. However, there is a
need for further research to explore the potential of incorporating technical indicators and
improving prediction accuracy.
 Summary of related work results are noted in Table
Author Technique Outcome Evaluation Dataset

RMSE:
ARIMA,
Bitcoin ARIMA:
Phaladis Kaggle
FBProph 322.4
ailoed et, XGB
price FBProphet:22
prediction 9.5 XGB: 369
 R2:

XBG: 4.855%

Retur
Forecasting ns:
returns for OLS:
3703 Equal- Vari
Seabe OLS, XGB
cryptocurrenc weighted: ous
ies, creating 6.297% sour
equal- Capital- ces
weighted weighted:
portfolio 1.852%

XGB:

Equal-
weighted:
7.105%

Capital-
weighted:
2.165%

Ensemble
Close price Accuracy:
learning
of Ensemble Coin-
methods, K-
Vanderb cryptocurren learning: marketc
NN model,
ilt cy and their 92.4% ap,
gradient
correspondi Gradient Cci30.co
boosted trees,
ng index boosted: m
neural net
90%
model.

Average:
Kim Decision trees Bitcoin Investin
Geometric:
g
2.77
 Arithmetic:
price
prediction 4.20

Cryptocurren Accuracy: Blockch

LSTM,
Iqbal cy price Bitcoin: ain,
prediction: 0.9957 Ethersca
Regres- sion
Bitcoin,Ethere Ethereum: n
trees
um
0.9999

Accuracy
: LR:
66.0%
LSTM, QDA, Coin-
Bitcoin
LDA: 63.9%
Kumar SVM, RF, marketc
XGB, LR, QDA: 55.1% ap
price
LDA. SVM: 65.3% Binance
prediction

RF: 51.0%

XGB: 48.3%

LSTM: 57.0%

Cryptocurrenc RMSE of
ARIMA,
y forecasting: BART: Yahoo
Derbent
Regression Bitcoin,Ethere finance
sev tree, BART 14 days: 4%
um, Ripple
21 days: 6%

30 days: 8%
 CHAPTER THREE

Methodology

Theoretical background highlighting some topics related to the project work is given below.
The description contains several topics which are worth discussing and also highlight some of
their limitations that encourage finding solutions, as well as highlighting some of their
advantages, which are the reasons these topics and their features are used in this project.

3.1 Prediction Models

3.1.1 Deep Neural Networks

The first prediction model was based on a deep neural network (DNN). A DNN usually
consists of an input layer, several hidden layers, and an output layer, as shown in Figure 3.1.

Figure 3.1 : An example of a fully-connected deep neural network (DNN) model.

In the DNN prediction model, every node in one layer is connected to every node in the
next layer, i.e., fully-connected. That is,each node in the hidden layer takes input vectors
x1, . . ., xn from the previous layer, where n is the number of nodes in the previous layer
and the size of each Vector is m. Its output h is then defined as

h=σh(w1 x1+· · · +wn xn+b)

where the weight vectors w1, . . ., wn and the bias b are parameters to be learned using
training data. In addition, is an element-wise multiplication (Hadamard product) and σh is
a nonlinear function, such as the logistic sigmoid, defined by the activation function

σ(x) = 1
1+exp(−x)

and the rectified linear unit (ReLU) , defined as g(x) = max(0,x). σh is applied element-
wise, and thus the size of his also m. In the output layer, the set of input vectors are
flattened into a single vector and then translated into a single value by using a dot product
between the flattened vector and a weight vector. For a regression problem, the final value
y is used as the predicted Bitcoin price. For a classification problem y is further translated
into a value between 0 and 1 using a sigmoid function and then rounded off. In particular, 1
means the rise of the price and 0 means the fall or no change.
If we represent the whole DNN model as a function f , the parameters of the model are
optimized using the back-propagation algorithm and training data so that the following
mean squared error can be minimized:

Where f(Xi) is a predicted value based on an input sequence Xi and Yitrue is the true
Bitcoin price at the day after the day corresponding to Xi[m] for regression. For
classification, Yitrueis 1 if the true price is higher than the price at Xi[m]and otherwise 0.

3.1.2 Recurrent Neural Networks and Long Short – Term Memory

For time series data, it is well known that a recurrent neural network (RNN) model and a
long short- term memory (LSTM) model are effective. Although a DNN model can be
used to predict time series data, it does not exploit the temporal relationship between the
data points in the input sequence. To exploit the temporal relationship, both RNN and
LSTM models use an internal state (memory). Figure 2 shows example RNN and LSTM
models for the Bitcoin price prediction.
Figure 3.2. Recurrent neural network (RNN)model (left) and long short-term memory (LSTM)
model (right).

Figure 2. Recurrent neural network (RNN)model (left) and long short-term memory (LSTM) model
(right).

To illustrate, suppose that we use three days data to predict the price. Then, the input layer of an
RNN model consists of three nodes each of which takes the whole data of a single day, i.e., a vector
of 18 features. Given xtas an input, each hidden state htis then computed as follows:

ht=tanh(Whxt+Uhht−1+bh)

where Wh , Uh, and bh are parameters to be learned and tanh is a

hyperbolic tangent function. The main difference from DNNs is that RNNs
use not only the input xt, but also the previous hidden state ht−1, to
compute the current hidden state ht, thus exploiting the temporal
dependency between the input sequence data. Assuming that we use three
days sequence data, the final output y for regression is computed as
follows,

y=wyh3+by

where wy and by are also parameters to be learned. One drawback of RNNs is that when the length
of the input sequence is large, the long-term information is rarely considered due to the so-called
vanishing gradient problem. LSTM avoids the vanishing gradient problem mainly by using several
gate structures, namely input, output, and forget gates, together with a cell unit (the memory part of
an LSTM unit), and additive connections between cell states. For an LSTM unit at time t, its hidden
state ht, i.e., the output vector of the LSTM unit, is computed as follows,
Forget Gate:
ft=σ(Wfxt+Ufht−1+bf)f_t = \sigma(W_f x_t + U_f h_{t-1} +
b_f)ft=σ(Wfxt+Ufht−1+bf)

● ftf_tft: Forget gate activation for the current time step.

● σ\sigmaσ: Sigmoid activation function.
● WfW_fWf: Weight matrix for the input xtx_txt.
● UfU_fUf: Weight matrix for the previous hidden state ht−1h_{t-
1}ht−1.
● bfb_fbf: Bias term for the forget gate.

Input Gate:
it=σ(Wixt+Uiht−1+bi)i_t = \sigma(W_i x_t + U_i h_{t-1} +
b_i)it=σ(Wixt+Uiht−1+bi)

● iti_tit: Input gate activation.

● WiW_iWi: Weight matrix for the input xtx_txt.
● UiU_iUi: Weight matrix for the previous hidden state ht−1h_{t-
1}ht−1.
● bib_ibi: Bias term for the input gate.

Output Gate:
ot=σ(Woxt+Uoht−1+bo)o_t = \sigma(W_o x_t + U_o h_{t-
1} + b_o)ot=σ(Woxt+Uoht−1+bo)

● oto_tot: Output gate activation.

● WoW_oWo: Weight matrix for the input xtx_txt.
● UoU_oUo: Weight matrix for the previous hidden state ht−1h_{t-
1}ht−1.
● bob_obo: Bias term for the output gate.

Cell State Update:

gt=tanh⁡(Wgxt+Ught−1+bg)g_t = \tanh(W_g x_t +
U_g h_{t-1} + b_g)gt=tanh(Wgxt+Ught−1+bg)

● gtg_tgt: Candidate values for the cell state.

● WgW_gWg: Weight matrix for the input xtx_txt.
● UgU_gUg: Weight matrix for the previous hidden state ht−1h_{t-
1}ht−1.
● bgb_gbg: Bias term for the candidate values.
 Cell State:
ct=ft⋅ct−1+it⋅gtc_t = f_t \cdot c_{t-1} + i_t \
cdot g_tct=ft⋅ct−1+it⋅gt

● ctc_tct: Cell state at time step ttt.

● ct−1c_{t-1}ct−1: Previous cell state.
● ftf_tft: Forget gate activation, controlling how much of the previous cell state to keep.
● iti_tit: Input gate activation, determining how much of the new candidate values to add to the
cell state.
● gtg_tgt: Candidate values for the cell state.

Hidden State:
ht=ot⋅tanh⁡(ct)h_t = o_t \cdot \tanh(c_t)ht=ot
⋅tanh(ct)

● hth_tht: Hidden state at time step ttt.

● oto_tot: Output gate activation, controlling the output of the LSTM cell.
● ctc_tct: Current cell state, transformed by the tanh function.

where W, U, and b are parameters to be learned and is a sigmoid function. ft , it , and ot,
respectively, correspond to the forget, input, and output gates, and ctis the cell state vector.

Thanks to the cell state, the long-term dependencies between the data points in the input sequence is
maintained well and LSTMs can be applied to long sequence data. In the experiment, only LSTM-
based prediction models were considered.

3.2 Convolutional Neural Networks

A convolutional neural network (CNN), has many applications in image analysis and classification
problems. Recently, it is shown to be also effective for sequence data analysis, and thus we also
developed a CNN-based prediction model. Normally, a CNN consists of a series of convolution
layers, ReLU layers, pooling layers, and fully-connected layers, where a convolution layer
convolves the input with a dot product. In this work, we developed a simple CNN model consisting
of a single 2D convolution layer (Conv2D) as shown in Figure 7. (A more sophisticated CNN model
is discussed in the next section.) The input is an m× 18 matrix, where m is the number of days to be
consulted for the prediction and 18 is the number of the features. A total of 36 2D convolution
filters of size 3 × 18 are used for convolution, where single real values are extracted from
consecutive three days data through each filter. That is, 3 × 18 feature values are translated into a
single value and thus
each filter produces a real-valued vector of size m− 2, which is then
applied to an element-wise ReLU activation function. Then, the 36
output vectors produced by the convolution filters are flattened into a
single vector of size (m− 2 )× 36, which is then translated into a single
prediction value through a fully connected layer. The number of filters
was determined experimentally. Moreover, unlike other image analysis
applications, simply adding more convolution layers together with
pooling layers did not improve the performance of CNN models for
Bitcoin price prediction, and therefore we present only a simple CNN
model.
3.3 Deep Residual Networks

Although stacking many layers allows a CNN model to encode a very

complex function from the input to the output, it is difficult to train
such a model due to the vanishing gradient problem as in RNNs. A deep
residual network (ResNet) resolves this problem by introducing a so-
called “identity shortcut connection” that skips one or more layers.
Figure 3 shows stacked nonlinear layers without/with a shortcut
connection. Assuming that H(x) is the desired underlying mapping,
residual learning lets the stacked nonlinear layers fit a residual
mapping of H(x)−x. The original mapping is then recast to F(x) + x. In
the work by the authors of, it is shown that the residual mapping is
easier to optimize than the original. By simply adding identity shortcut
connections, deep convolutional networks can easily be optimized and
produce a good result. Figure 3 shows our ResNet model used in the
experiment. It consists of an input layer, one 2D convolution layer, four
residual blocks, one average pooling layer, and an output layer. A
residual block implements a shortcut connection between two
convolution layers. For ResNet, we use convolution layers different
from the one used for CNN. More precisely, the first Conv2D for input
data uses 36 convolution filters of size 4 × 4. Then, each Conv2D in the
four residual blocks uses filters of size 2× 2. The number of filters
used in each Conv2D of the first, second, third, and last residual blocks
is 36, 72, 144, and 288, respectively. Although it is not shown in the
figure, we apply batch normalization and the ReLU activation function
after each Conv2D in each residual block. After applying 2D average
pooling of size 3 × 3, we flatten the resultant 288 vectors of size 1 × 1
into a one-dimensional vector. It is translated into single prediction
value through a fully connected layer as in the CNN model.
Figure 3.3. Residual Learning
3.4 Combinations of CNN and RNNs

As CNNs are effective for structural analysis, whereas RNNs are effective for temporal analysis,
developed a prediction model called CRNN by combining CNN and LSTM models so as to
integrate their strengths . Figure 4 shows the structure of our CRNN model. The same input data are
fed into both the CNN and LSTM models discussed in the previous sections. In particular, for the
CNN part, after applying 2D convolution, we also apply batch normalization and flattening. As for
the LSTM part, we additionally apply dropout . The results of CNN and LSTM are then combined
into a single vector, which is then translated into a single prediction value through a fully connected
layer. As a combining operator, we tested element-wise addition, subtraction, multiplication,
average, maximum and minimum operators, and a concatenation operator. In the experiment, the
subtraction operator was used for regression problems while the concatenation operator was used
for classification problems because they showed better performance.

Figure 3.4. CRNN model (combination of CNN and RNN

Chapter 4

Result and Discussion

A System Requirement Specification (SRS) is basically an organization’s understanding of a

customer or potential client’s system requirements and dependencies at a particular point prior to any
actual design or development work. The information gathered during the analysis is translated into a
document that defines a set of requirements. It gives a brief description of the services that the
system should provide and also the constraints under which the system should operate. Generally,
SRS is a document that completely describes what the proposed software should do without
describing how the software will do it. It’s a two-way insurance policy that assures that both the
client and the organization understand the other’s requirements from that perspective at a given point
in time.
SRS document itself states in precise and explicit language those functions and capabilities a
software system (i.e., a software application, an ecommerce website and so on) must provide, as well
as states any required constraints by which the system must abide. SRS also functions as a blueprint
for completing a project with as little cost growth as possible. SRS is often referred to as the “parent”
document because all subsequent project management documents, such as design specifications,
statements of work, software architecture specifications, testing and validation plans, and
documentation plans, are related to it.

Requirement is a condition or capability to which the system must conform. Requirement

Management is a systematic approach towards eliciting, organizing and documenting the
requirements of the system clearly along with the applicable attributes. The elusive difficulties of
requirements are not always obvious and can come from any number of sources.
Functional Requirement defines a function of a software system and how the system must behave
when presented with specific inputs or conditions. These may include calculations, data manipulation
and processing and other specific functionality. In this system following are the functional
requirements: -

Following are the functional requirements on the system:

1. The entire control model set must be translated to C output Code.
2. Inputs must be models designed using CLAW design components along with standard
design components,
3. Multiple design models must be processed and the result must be combined to obtain a
single output file.

4.2 Non-Functional Requirement

Nonfunctional requirements are the requirements which are not directly concerned with the specific
function delivered by the system. They specify the criteria that can be used to judge the operation of
a system rather than specific behaviors. They may relate to emergent system properties such as
reliability, response time and store occupancy. Nonfunctional requirements arise through the user
needs, because of budget constraints, organizational policies, the need for interoperability with other
software and hardware systems or because of external factors such as: -
● Product Requirements
● Organizational Requirements
● User Requirements
● Basic Operational Requirements

4.3 RESULT AND PERFORMANCE ANALYSIS

4.3.1 Snapshots
4.4 SYSTEM ANALYSIS

Analysis is the process of finding the best solution to the problem. System analysis is the process by
which we learn about the existing problems, define objects and requirements and evaluate the
solutions. It is the way of thinking about the organization and the problem it involves, a set of
technologies that helps in solving these problems. Feasibility study plays an important role in system
analysis which gives the target for design and development.

4.4.1Feasibility Study

All systems are feasible when provided with unlimited resources and infinite time. But unfortunately,
this condition does not prevail in the practical world. So it is both necessary and prudent to evaluate
the feasibility of the system at the earliest possible time. Months or years of effort, thousands of
rupees and untold professional embarrassment can be averted if an ill- conceived system is
recognized early in the definition phase. Feasibility & risk analysis are related in many ways. If
project risk is great, the feasibility of producing quality software is reduced. In this case three key
considerations involved in the feasibility analysis are

● ECONOMICAL FEASIBILITY

● TECHNICAL FEASIBILITY

● SOCIAL FEASIBILITY

4.4.2 Economical Feasibility

This study is carried out to check the economic impact that the system will have on the organization.
The amount of fund that the company can pour into the research and development of the system is
limited. The expenditures must be justified. Thus, the developed system as well within the budget
and this was achieved because most of the technologies used are freely available. Only the
customized products had to be purchased.
 Technical Feasibility

This study is carried out to check the technical feasibility, that is, the technical requirements of the
system. Any system developed must not have a high demand on the available technical resources.
This will lead to high demands on the available technical resources. This will lead to high demands
being placed on the client. The developed system must have a modest requirement, as only minimal
or null changes are required for implementing this system.

Social Feasibility

The aspect of study is to check the level of acceptance of the system by the user. This includes the
process of training the user to use the system efficiently. The user must not feel threatened by the
system, instead must accept it as a necessity. The level of acceptance by the users solely depends on
the methods that are employed to educate the user about the system and to make him familiar with it.
His level of confidence must be raised so that he is also able to make some constructive criticism,
which is welcomed, as he is the final user of the system.

4.5 Analysis

Performance Analysis

For the complete functionality of the project work, the project is run with the help of a healthy
networking environment. Performance analysis is done to find out whether the proposed system. It is
essential that the process of performance analysis and definition must be conducted in parallel.

Technical Analysis

System is only beneficial only if it can be turned into information systems that will meet the
organization’s technical requirement. Simply stated this test of feasibility asks whether the system
will work or not when developed & installed, whether there are any major barriers to
implementation. Regarding all these issues in technical analysis there are several points to focus on:
 1. Changes to bring in the system: All changes should be in a positive direction, there will be
increased level of efficiency and better customer service.
2. Required skills: Platforms & tools used in this project are widely used. So the skilled
manpower is readily available in the industry.
3. Acceptability: The structure of the system is kept feasible enough so that there should not be
any problem from the user’s point of view.
Economical Analysis

Economic analysis is performed to evaluate the development cost weighed against the ultimate
income or benefits derived from the developed system. For running this system, we need not have any
routers which are highly economical. So, the system is economically feasible enough
4.6 SYSTEM DESIGN

Fig 5
Our project addresses leveraging appropriate machine learning techniques to engineer sample
dimensions for Bitcoin price prediction. Inspired by the principle of Occam’s razor and the
characteristics of our datasets, we tackle the problem as follows. First, the prediction sample is
divided into daily intervals with small sample size and 5-minute intervals with a big sample size.
Second, we conduct the features engineering: select high-dimension features for daily price and few
features for 5-minute interval trading data respectively. Third, we conduct simple statistical models
including Logistic Regression and Linear Discriminant Analysis and the more complicated machine
learning models including Random Forest, XGBoost, Quadratic Discriminant Analysis, Support
Vector Machine and Long Short-term Memory. Fourth, we adopt the simple statistical methods to
predicting Bitcoin daily price with high- dimensional features to avoid overfitting. Meanwhile, the
machine learning models are leveraged in high-frequency price few features. Fig. 5 shows the
overview of our research framework.
Our project makes observations in two ways. One is to extend the feature dimensions, and the other
is to evaluate different machine learning techniques for solving problems of multiple frequence
Bitcoin prices. The study makes the following contributions. (1) To the best of our knowledge, we are
at the forefront of establishing higher dimensional features for problems of Bitcoin daily price
prediction by integrating investor attention, media hype and XAU gold spot features with common
and traditional features such as network and market. (2) We address the importance of the sample
dimension by classifying Bitcoin price data by interval. The real-time 5-minute interval trading data
acquired from the top cryptocurrency exchange is high-frequency and large scale, and the aggregated
Bitcoin daily price obtained from CoinMarketCap is low-frequency and small scale. Hence, the
problem of Bitcoin price prediction is addressed from a broad perspective. (3) To find appropriately
complex models and meet the requirement of accuracy, we evaluate different machine learning
techniques using problems of multiple frequency Bitcoin price. Specifically, we lower the complexity
of algorithms for low-frequency daily price prediction with higher-dimension features and apply more
complicated models for high-frequency price prediction with a few features. The results show that
simple statistical methods outperform machine learning models for daily Bitcoin price prediction
while more complicated models should be adopted for high frequency Bitcoin price prediction. We
envision this study as a pilot for dealing with datasets with different scales and intervals, which can
shed light on other industrial prediction problems in the context of machine learning.

Chapter 5:

Conclusion and Future Scope

In this study, we explored the effectiveness of deep learning models, specifically Recurrent Neural
Networks (RNN) and Long Short-Term Memory (LSTM) networks, for predicting Bitcoin prices. The
findings indicate that LSTM models are particularly adept at capturing longer-term dependencies
within the data, showcasing their potential in financial forecasting.

However, predicting Bitcoin prices remains a challenging task due to the high variance nature of the
data, which complicates the achievement of reliable validation results. This inherent difficulty
emphasizes the need for a delicate balance between overfitting the model and ensuring sufficient
learning capacity. Techniques such as dropout have proven valuable in mitigating overfitting, yet
despite employing Bayesian optimization to fine-tune dropout selection, we could not consistently
achieve optimal validation outcomes.

While the metrics of sensitivity, specificity, and precision suggest commendable performance from the
neural network models, it is crucial to note that the ARIMA forecasting method demonstrated
significantly poorer performance when evaluated based on error metrics. This disparity underscores
the limitations of traditional statistical methods in comparison to advanced neural network approaches
in the context of Bitcoin price prediction.

Future Scope
The landscape of cryptocurrency prediction presents numerous opportunities for future research.
Further improvements could be made by integrating additional data sources, such as social media
sentiment analysis and macroeconomic indicators, to enhance the predictive capabilities of deep
learning models. Additionally, exploring hybrid models that combine the strengths of both neural
networks and traditional time series forecasting methods could yield promising results.

Furthermore, the adoption of more advanced techniques such as attention mechanisms and
transformer models may facilitate better modeling of complex temporal relationships in the data. As
the cryptocurrency market continues to evolve, ongoing research and innovation in predictive
modeling will be essential for developing robust forecasting systems.