Forecasting Bitcoin Prices: A Comparative Study of

Machine Learning and Deep Learning Algorithms

Hamed Alizadegan
Qazvin Branch, Islamic Azad University
Arian Radmehr
Islamic Azad University, South Tehran Branch
Mohsen Asghari Ilani

University of Tehran https://orcid.org/0000-0003-0189-6956

Research Article

Keywords: Time series forecasting, LSTM, Bi-LSTM, Deep learning, ML, LGBM, XGBoost, BTC,
cryptocurrency, RNN.

Posted Date: May 9th, 2024

DOI: https://doi.org/10.21203/rs.3.rs-4390390/v1

License:   This work is licensed under a Creative Commons Attribution 4.0 International License.
Read Full License

Additional Declarations: The authors declare no competing interests.

Page 1/19
Abstract
The cryptocurrency market, particularly Bitcoin, has witnessed significant volatility, making accurate
price prediction a challenging yet crucial task. This research explores the application of four powerful
machine learning algorithms), Light Gradient Boosting Machine (LightGBM , Long Short Term Memory
(LSTM), Bidirectional Long Short Term Memory (BiLSTM) and Extreme Gradient Boosting (XGBoost), for
forecasting Bitcoin prices. The study focuses on evaluating the predictive performance using Mean
Absolute Error (MAE) and Root Mean Squared Error (RMSE) as the evaluation metrics. The LSTM and Bi-
LSTM, a type of recurrent neural network (RNN), are known for that ability to capture long-term
dependencies in time series data. On the other hand, LightGBM and XGBoost, a gradient boosting
framework, excels in handling large datasets efficiently and delivering accurate predictions. By
employing these algorithms, this research aims to enhance the accuracy of Bitcoin price predictions
compared to traditional methods. The experimental setup involves training and validating the models on
historical Bitcoin price data. The MAE and RMSE metrics are utilized to assess the models' predictive
accuracy, providing a comprehensive evaluation of their performance. The comparative analysis of
machine learning models sheds light on their strengths and weaknesses in the context of cryptocurrency
price prediction. The results showcase the importance of employing advanced machine learning
techniques in forecasting financial time series, highlighting the potential for improved decision-making in
cryptocurrency trading and investment strategies.

Introduction
The cryptocurrency market has emerged as a dynamic and intricate financial landscape, characterized
by its inherent volatility and unpredictability. Among the myriad of digital assets, Bitcoin stands as a
prominent and influential player, captivating the attention of investors, traders, and researchers alikeon
tethenical(Ortu et al., 2022). The unprecedented surge in interest surrounding Bitcoin has led to an
increasing need for accurate and reliable methods of price prediction to inform investment decisions
and risk management strategies. The cryptocurrency market represents a complex and dynamic
financial ecosystem characterized by its inherent volatility and unpredictability. Within this landscape,
Bitcoin emerges as a prominent and influential digital asset, attracting the attention of investors, traders,
and researchers. The surge in interest surrounding Bitcoin necessitates accurate methods for price
prediction to guide investment decisions and risk management strategies. Cryptocurrencies rely on
Blockchain, a digital ledger system, for transactions. While Blockchain ensures privacy and security, its
inherent limitations in data formatting hinder efficient search queries. Cryptocurrencies, independent of
traditional banks, offer secure peer-to-peer transactions, but their Proof-of-Work mechanism demands
significant energy consumption, posing sustainability challenges. Bitcoin operates on a decentralized
peer-to-peer network, enabling currency creation and transaction management without central authority
control. BTC remains the most recognized cryptocurrency, with Ethereum, Litecoin, and Ripple as notable
alternatives. The competition among cryptocurrencies drives technological advancements within the
industry. However, its high volatility challenges its role as a store of value and transaction medium.
Efforts to predict Bitcoin prices often involve analyzing correlations with other commodities, yet past
Page 2/19
studies indicate weak correlations with assets like gold and stock market indices. Another research
direction involves leveraging AI algorithms and computational power to forecast Bitcoin prices (Rathore
et al., 2023). Machine learning, particularly LSTM and Bi-LSTM models, has gained traction in various
financial markets(Patel et al., 2020). This paper aims to explore deep learning and machine learning
applications in time series forecasting by employing LSTM, Bi-LSTM signals, LGBTM an XGBoost in
algorithmic investment strategies for Bitcoin(Ren et al., 2022). Our study proposes deep learning and
machine learning models for price prediction, focusing on machine learning techniques in algorithmic
investment strategies applied to BTC cryptocurrencies. deep learning and machine learning algorithms
facilitate cryptocurrency price predictions despite their nonlinear behavior. We evaluate prediction
performance using metrics like RMSE and MAE.

Related Works

In recent years, there has been a surge in research endeavors aimed at harnessing machine learning
techniques to predict cryptocurrency prices, particularly focusing on Bitcoin (BTC), Ethereum (ETH), and
Litecoin (LTC). A variety of models have been explored for this purpose, ranging from traditional
statistical methods to advanced deep learning architectures. LSTM (Long Short-Term Memory) and GRU
(Gated Recurrent Unit) models have emerged as popular choices due to their ability to capture temporal
dependencies in sequential data. For instance, Kim et al. (Kim et al., 2021) compares LSTM and GRU
models for BTC, ETH, and LTC price predictions, revealing that GRU outperforms LSTM during downward
trends in BTC and ETH, while LSTM fares better during upward trends. Moreover, GRU exhibits superior
performance for LTC in specific scenarios. Similarly, Seabe et al. (Seabe et al., 2023) introduces LSTM,
GRU, and Bi-LSTM models for price predictions, with Bi-LSTM demonstrating the highest accuracy
among the models considered.

Furthermore, researchers have explored the application of LSTM specifically in forecasting

cryptocurrency prices against traditional financial indicators. Michańków et al. (Michańków et al., 2022)
employs LSTM networks to forecast BTC and S&P500 index values, highlighting the importance of
innovative loss functions and meticulous data preprocessing in enhancing forecasting accuracy.
Additionally, Wu et al. (Wu et al., 2018a) proposes LSTM models for daily Bitcoin price forecasting,
introducing an LSTM with AR(2) model that outperforms conventional LSTM approaches. These studies
contribute novel frameworks adaptable to various cryptocurrencies and financial time-series data, as
shown in Fig. 1 (a-d). Moreover, machine learning-based approaches beyond LSTM have also shown
promise. Chen et al. (Chen, 2023) evaluates regression models' performance for minute-interval Bitcoin
price predictions, achieving high accuracy with low Mean Squared Error (MSE) and high R-squared (R2)
values.

These studies collectively underscore the growing interest and advancements in leveraging machine
learning for cryptocurrency price prediction, offering valuable insights for investors, traders, and
researchers alike.

Page 3/19
The collective findings of recent studies underscore the pivotal role of employing cutting-edge deep
learning algorithms and advanced machine learning techniques in accurately predicting cryptocurrency
prices. These insights are not only invaluable for researchers seeking to understand market dynamics
but also for investors and traders navigating the volatile landscape of the cryptocurrency market.
Motivated by this burgeoning field of research and the practical implications for investment strategies,
our study aims to leverage cryptocurrency datasets, augmented from reputable sources such as
Binance, to predict Bitcoin prices. By drawing upon the latest advancements showcased in recent
publications, we aim to enhance the reliability and confidentiality of our results. Moreover, our approach
focuses on comparing the efficacy of four powerful algorithms, LSTM Bi-LSTM, LGBM and XGBoost,
renowned for their predictive capabilities in financial forecasting. One of the key advantages of our
methodology lies in its accessibility and ease of use, making it accessible not only to seasoned
researchers but also to investors seeking actionable insights. By providing a streamlined prediction
procedure applicable across various cryptocurrencies, our approach eliminates the need for extensive
manual analysis, thereby saving valuable time and resources. This cost-effective solution is tailored to
real-world datasets, ensuring its relevance and applicability in dynamic market conditions. our study
aims to empower investors and traders with a robust predictive framework that can navigate the
complexities of cryptocurrency markets with confidence. By harnessing the power of advanced machine
learning techniques, we strive to offer a comprehensive solution that transcends traditional barriers,
facilitating informed decision-making and maximizing returns in the exciting world of cryptocurrency
investment.

Methodology
Data collection
The data used in this study comprises Bitcoin price data obtained from the Binance exchange relative to
the Tether (USDT) cryptocurrency. The dataset spans from October 1, 2019, to September 30, 2023,
capturing daily price observations over this period. The choice of Binance exchange for data collection is
motivated by its status as one of the leading cryptocurrency exchanges globally, known for its liquidity
and extensive trading volume. Additionally, the use of Tether (USDT) as the reference currency provides
stability and consistency in price comparisons due to its peg to the US dollar. The daily timeframe was
selected to provide sufficient granularity for analysis while avoiding excessive noise associated with
shorter time intervals. By aggregating price data on a daily basis, we aim to capture overarching trends
and patterns in Bitcoin price movements over the specified timeframe.

Datasets
the features and target variable are depicted in Fig. 2 The depicted features are associated with their
respective Fisher's scores. Fisher's score is a statistical measure used for feature selection in machine
learning and predictive modeling tasks. It quantifies the discriminatory power of each feature by
assessing its ability to differentiate between different classes or categories within the dataset.

Page 4/19
Specifically, Fisher's score evaluates the separation between the means of different classes relative to
the variance within each class. Features with higher Fisher's scores indicate greater discriminatory
power, suggesting that they are more informative and relevant for predicting the target variable. The term
"Close" refers to the closing price of Bitcoin, which serves as our target variable. Our objective involves
predicting future closing prices based on a range of relevant features extracted from the dataset. In our
feature set, we have included several date-related features extracted from the timestamp of each data
point. These features provide additional temporal context to the dataset and help capture potential
seasonal or periodic patterns in Bitcoin's price movements. Here's a brief explanation of each date-
related feature:

Month: This feature represents the month component of the timestamp, indicating the calendar
month in which the data point was recorded.
Year: The year feature denotes the calendar year associated with each data point, providing a
broader temporal context to the dataset.
Day of Year: This feature represents the ordinal day of the year (ranging from 1 to 365 or 366),
capturing the annual progression of time.
Day of Month: Similar to the day of year, this feature indicates the specific day within the calendar
month, ranging from 1 to 31.
Day of Week: This feature denotes the day of the week (e.g., Monday, Tuesday, etc.) associated with
each data point, allowing for the capture of potential weekly patterns in Bitcoin's price dynamics.

Sin(DayofYear) and Cos(DayofYear): These features encode the cyclical nature of the day of the year
using sine and cosine transformations, respectively. By representing the day of the year as a continuous
cycle rather than a linear progression, these features help capture seasonal patterns and periodic
fluctuations in Bitcoin prices more effectively.

Incorporating these date-related features into our predictive models enhances their ability to capture
temporal dependencies and seasonal variations in Bitcoin's price movements, thereby improving the
accuracy of our forecasting models.

The graphical representation in Fig. 3 depicts the trend of Bitcoin's closing prices over the analyzed time
period. As observed, the Bitcoin price exhibits notable fluctuations and volatility, characterized by
periodic peaks and troughs. This dynamic price behavior is indicative of the inherent volatility and
speculative nature of the cryptocurrency market.

In the presented graphical representation in Fig. 4, we observe the trading volume associated with
Bitcoin transactions over the specified duration. The trading volume reflects the total number of Bitcoin
units traded within a given timeframe, typically measured in terms of the number of coins exchanged or
the equivalent value in fiat currency. The depicted volume graph showcases fluctuations in trading
activity, with periods of heightened trading volume often coinciding with significant price movements or

Page 5/19
market events. This dynamic interplay between trading volume and price movements underscores the
importance of market liquidity and investor sentiment in shaping the dynamics of the Bitcoin market.

The scatter plot depicted in Fig. 5 depicts the relationship between Bitcoin's closing price and trading
volume. Each data point on the scatter plot represents a specific observation, with the x-axis
corresponding to the closing price of Bitcoin and the y-axis representing the corresponding trading
volume. The scatter plot allows for visual examination of any potential correlations or patterns between
price movements and trading activity. By analyzing the distribution of data points, we can discern trends,
identify outliers, and assess the strength of the price-volume relationship.

We utilized a heatmap visualization in Fig. 6 to explore the correlations among various features extracted
from the cryptocurrency market dataset obtained from Binance. The heatmap provides a visual
representation of the pairwise correlation coefficients between different variables, offering insights into
the interrelationships within the dataset. The heatmap analysis revealed significant correlations between
certain market indicators and Bitcoin price movements. We observed a strong positive correlation
between Bitcoin's price and trading volume, indicating that higher trading activity tends to coincide with
price increases. Furthermore, the heatmap allowed us to identify potential multicollinearity issues among
predictor variables, which could affect the robustness of predictive models. By examining the correlation
matrix, we were able to discern redundant or highly correlated features, enabling us to refine the feature
selection process and improve the model's predictive performance.

In this experiment, we adopted a standard train-test split approach, where 90% of the dataset was
reserved for training the predictive models, and the remaining 10% was set aside for testing purposes.
The division of the dataset into training and testing subsets ensures that the models are evaluated on
unseen data, thereby providing a reliable assessment of their generalization performance. As illustrated
in Fig. 7, the training data (depicted in blue) encompasses 90% of the total dataset and is utilized to train
the machine learning and deep learning models. This portion of the dataset includes historical Bitcoin
price data, along with relevant market indicators and trading volume, spanning a period of four years.
Conversely, the testing data (depicted in orange) constitutes the remaining 10% of the dataset and is
held out for model evaluation. This independent subset of the data allows us to assess the models'
predictive performance on unseen instances, thereby gauging their ability to generalize to new data and
make accurate forecasts. By delineating the training and testing data in this manner, we ensure a
rigorous evaluation of the predictive models' efficacy in forecasting Bitcoin prices. Additionally, this
approach mitigates the risk of overfitting, as the models are assessed on data that they have not been
exposed to during the training phase.

ML Model
Machine learning, a subset of artificial intelligence (AI), empowers computers to learn from data and
make predictions or decisions without explicit programming. In the context of the referenced article,
machine learning algorithms are harnessed to analyze historical Bitcoin price data, uncovering patterns

Page 6/19
or relationships that inform predictions of future price movements. By learning from past data, these
algorithms forecast future data points, providing investors and researchers with insights into potential
trends in the cryptocurrency market. The study employs diverse machine learning models, including deep
learning, ensemble methods, and time series forecasting algorithms, each offering distinct advantages
and capabilities for analyzing intricate financial data.
LSTM (Long Short-Term Memory)
Long Short-Term Memory (LSTM) represents a specialized recurrent neural network (RNN) architecture
meticulously crafted to grasp and model long-term dependencies inherent in sequential
data(Christoforou et al., 2020). Unlike conventional RNNs, LSTM networks integrate dedicated memory
cells and gating mechanisms, addressing the prevalent issue of vanishing gradients that often obstructs
deep neural network training on lengthy sequences. Comprising multiple LSTM units, each unit
encompasses three pivotal components: the input gate, the forget gate, and the output gate. The input
gate orchestrates information influx into the memory cell, determining the relevance of data to update
and retain. Conversely, the forget gate governs the retention or removal of irrelevant data from previous
time steps, enhancing network adaptability. The output gate regulates the information flow from the
memory cell to the output, discerning pertinent data for prediction. Central to LSTM architecture is the
memory cell, acting as a cornerstone component by upholding a persistent internal state capable of
storing information across extensive periods. This cell state undergoes updates via a blend of element-
wise operations and activation functions, facilitating selective information retention or discarding based
on relevance (Nasirtafreshi, 2022). During training, LSTM networks undergo optimization via
backpropagation through time (BPTT), a variant of the backpropagation algorithm tailored for sequential
data. The network refines its parameters (weights and biases) by minimizing a predefined loss function,
commonly quantified in terms of prediction error such as Mean Squared Error (MSE) or Mean Absolute
Error (MAE).

LSTM networks excel in modeling and predicting time series data laden with long-term dependencies,
rendering them particularly adept for tasks like speech recognition, language modeling, and financial
forecasting. Leveraging LSTM's prowess in capturing and preserving information across extended
sequences empowers it to discern intricate patterns and dynamics inherent in sequential data, including
Bitcoin price fluctuations. In the realm of Bitcoin price prediction, LSTM networks analyze historical price
data to forecast future price movements. Harnessing the network's capacity to discern temporal
dependencies, LSTM models can potentially uncover recurring patterns, trends, and anomalies in Bitcoin
price dynamics, thus assisting traders and investors in decision-making endeavors. Despite their
efficacy, LSTM networks may encounter challenges such as overfitting, especially when trained on
limited or noisy datasets. Rigorous hyperparameter tuning, incorporation of regularization techniques,
and robust model validation are imperative to alleviate these challenges and ensure the reliability and
generalization capability of LSTM-based predictions.
Bidirectional Long Short-Term Memory (Bi-LSTM)

Page 7/19
Bidirectional Long Short-Term Memory (Bi-LSTM) represents an advancement over the conventional
LSTM architecture by integrating bidirectional processing. Unlike standard LSTM models, which solely
handle input sequences in one direction (i.e., from past to future), Bi-LSTM concurrently processes
sequences in both directions. This bidirectional approach empowers the model to grasp not only past
dependencies but also future context, thereby facilitating more informed predictions based on a
comprehensive understanding of the input data. In Bi-LSTM, the input sequence is fed into two distinct
LSTM layers: one processes the sequence in the forward direction, while the other processes it in the
reverse direction. Subsequently, the hidden states of these two LSTM layers are either concatenated or
merged at each time step, yielding a representation that amalgamates information from both past and
future contexts(Huang et al., 2022). This bidirectional processing mechanism enables Bi-LSTM models
to better comprehend the temporal dynamics inherent in the input data, consequently enabling them to
capture intricate patterns that might elude unidirectional models.

Light Gradient Boosting Machine (LGBM)

LightGBM stands out as a robust gradient boosting framework celebrated for its efficiency, scalability,
and exceptional performance when dealing with extensive datasets and intricate feature spaces.
Originating from Microsoft, LightGBM introduces a novel gradient-based approach to decision tree
boosting, enabling swift and accurate predictions across diverse machine learning tasks like regression,
classification, and ranking. Belonging to the gradient boosting machines family, LightGBM iteratively
constructs an ensemble of weak learners, typically decision trees, to collectively enhance predictive
capabilities. Unlike conventional gradient boosting algorithms, LightGBM adopts a leaf-wise tree growth
strategy, prioritizing nodes for splitting based on maximum reduction in the objective function's gradient.
Its efficiency in model training owes to an innovative histogram-based algorithm for computing gradients
and identifying optimal split points. By discretizing feature values into bins and constructing histograms,
LightGBM effectively reduces memory usage and computational overhead, rendering it well-suited for
large datasets with millions of samples and high-dimensional feature spaces. Notably, LightGBM
seamlessly handles categorical features, employing techniques like one-hot encoding or ordinal
encoding while preserving their semantic meaning for efficient computation during training. With an
array of hyperparameters available for controlling model complexity, regularization, and optimization
objectives, LightGBM incorporates regularization techniques such as shrinkage and feature sub-
sampling to prevent overfitting and enhance generalization. Further optimization through techniques like
grid search, random search, or Bayesian optimization aids in finding the optimal configuration for
specific datasets and tasks. Its proficiency in balancing computational efficiency and predictive accuracy
makes LightGBM a preferred choice for large-scale machine learning tasks across industries. Its
capacity to handle categorical features, coupled with memory-efficient implementation and scalability to
distributed computing environments, underscores its widespread adoption. In the realm of Bitcoin price
prediction, LightGBM proves invaluable in modeling intricate relationships among market indicators,
sentiment analysis, and external factors influencing Bitcoin prices. Leveraging its efficient training
process and robust performance, LightGBM facilitates accurate forecasts of Bitcoin price movements,
empowering traders, investors, and analysts to make well-informed decisions.

Page 8/19
eXtreme Gradient Boosting (XGBoost)
XGBoost, an abbreviation for eXtreme Gradient Boosting, represents a potent machine learning algorithm
within the ensemble learning family, specifically residing in gradient boosting frameworks. It has
garnered acclaim for its remarkable performance across a broad spectrum of predictive modeling
endeavors, spanning regression, classification, and ranking tasks. XGBoost operates by iteratively
amalgamating multiple weak learners, commonly decision trees, to construct a robust predictive model.
Each subsequent tree in the ensemble is trained to rectify the errors of its predecessors, culminating in a
final model of high accuracy and resilience. A noteworthy advantage of XGBoost lies in its adeptness at
efficiently handling intricate datasets. Leveraging gradient boosting, it optimizes the loss function by
iteratively fitting new models to the residuals of preceding ones (Zhai et al., 2020). This distinctive
approach empowers XGBoost to discern subtle patterns and relationships within the data, rendering it
particularly well-suited for tasks entailing structured data or high-dimensional feature spaces.
Model Evaluation
Model evaluation is a critical step in assessing the predictive accuracy and reliability of LSTM and
LightGBM models for Bitcoin price prediction. Various performance metrics are utilized to quantify the
discrepancy between predicted and actual prices. MAE (Mean Absolute Error) measures the average
absolute difference between the predicted and actual prices over the entire test dataset.

Mathematically, MAE is calculated as:

T
MAE = 1

t
∑
t=1
|y
t
− y |
t
′

RMSE is another commonly used metric that penalizes larger prediction errors more heavily than smaller
errors. It is calculated as the square root of the average of the squared differences between the
predicted and actual prices(Wu et al., 2018b):

T 2
RMSE = √ 1t ∑t=1 (yt − y )
′
t

Where, yt represent the true price of Bitcoin, yt′ denote the predicted price of Bitcoin, and T indicate the
specific period for prediction.

The coefficient of determination, commonly denoted as R-squared (R ), quantifies the proportion of

2

variance in the dependent variable (actual prices) that is accounted for by the independent variables
(predicted prices). Ranging between 0 and 1, higher values signify a stronger alignment of the model
with the observed data.

LSTM and LightGBM models are evaluated using the aforementioned performance metrics on the
testing dataset. The results are compared to determine which model exhibits superior predictive
accuracy and generalization capability for Bitcoin price prediction. The interpretation of evaluation
Page 9/19
results involves understanding not only the absolute performance metrics but also the practical
implications of model predictions for real-world applications. Insights gained from the evaluation
process inform decision-making processes for traders, investors, and stakeholders in the cryptocurrency
market.

Results and discussion

The performance of various machine learning models in predicting Bitcoin prices was assessed using
diverse evaluation metrics. Results from the Long Short-Term Memory (LSTM) model indicate the
following: Mean Absolute Error (MAE) on test data was 0.0106, Root Mean Squared Error (RMSE) was
0.0158, and the R-squared value was 0.6787. Conversely, the Bidirectional LSTM (Bi-LSTM) model
displayed improved performance with MAE of 0.0096, RMSE of 0.0146, and R-squared of 0.7252.
Furthermore, the LightGBM (LGBM) model surpassed both LSTM and Bi-LSTM, exhibiting significantly
lower MAE (0.0082) and RMSE (0.0106), along with a higher R-squared value (0.854). Lastly, the XGBoost
model achieved MAE of 0.0109, RMSE of 0.0136, and R-squared of 0.7587, positioning it as a
competitive alternative to LSTM-based approaches. These findings suggest that LGBM offers superior
predictive accuracy compared to LSTM, Bi-LSTM, and XGBoost models. The lower MAE and RMSE
values imply that LGBM provides more precise Bitcoin price predictions, while the higher R-squared value
indicates better overall fit to the test data. This performance superiority may be attributed to LGBM's
inherent characteristics, such as efficient handling of large datasets and capturing complex nonlinear
relationships within the data. Additionally, it is noteworthy that Bi-LSTM outperformed LSTM,
underscoring the potential advantages of bidirectional architectures in capturing temporal dependencies
in time series data. However, both LSTM and Bi-LSTM models lag behind LGBM and XGBoost in terms of
predictive accuracy and overall performance. Overall, these results underscore the significance of
selecting appropriate machine learning algorithms for cryptocurrency price prediction tasks. While LSTM
and Bi-LSTM exhibit predictive capability, LGBM emerges as a more reliable and accurate choice for
forecasting Bitcoin prices based on the evaluation metrics considered in this study.

The loss curve for both training and validation datasets over epochs is illustrated in Fig. 8. This curve
provides insights into the convergence of the models during the training process. Both training and
validation loss curves are evaluated to ensure that the models are learning effectively without overfitting
or underfitting. The actual vs predicted values for the test dataset are presented in Fig. 9. This
visualization allows us to assess the predictive accuracy of each model by comparing the predicted
values with the ground truth. The closer the predicted values are to the actual values, the higher the
predictive accuracy of the model. The daily price trend is depicted in Fig. 10 showcasing the conformity
of different models' predictions with the actual price data. Models such as LGBM, Bi-LSTM, LSTM, and
XGBoost are compared to determine which one exhibits the highest level of agreement with the actual
price trends over time. In the analysis, a comparative view of key evaluation metrics, including Mean
Absolute Error (MAE), Root Mean Squared Error (RMSE), and R-Squared, across different models is
presented in Fig. 11. Additionally, Table 1 provides a tabulated summary of these evaluation metrics for
each model, allowing for a detailed comparison of their performance. Various evaluation metrics are
Page 10/19
displayed in a linear graph in Fig. 12, providing a comprehensive overview of the performance of each
model. This visualization enables us to identify trends and patterns in the performance of different
models across different evaluation metrics. Overall, the combination of visualizations and tabulated
results facilitates a comprehensive analysis of the performance of machine learning models in
predicting Bitcoin prices.

A detailed summary of the evaluation metrics associated with each model is provided in Table 1. It
includes metrics such as Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and R-Squared
(R^2) for each model, offering insights into their performance in predicting Bitcoin prices.

Table 1
Summary of Performance Metrics
Model/Metric MAE RMSE R-Squared

LSTM 0.0101 0.0156 0.68

BI-LSTM 0.0096 0.0145 0.72

LGBM 0.0081 0.0105 0.85

XGBoost 0.0108 0.0136 0.75

Conclusion
In conclusion, our study explored the effectiveness of various machine learning models in predicting
Bitcoin prices using historical data. Through rigorous experimentation and analysis, we evaluated the
performance of models such as LSTM, Bi-LSTM, LGBM, and XGBoost, considering key evaluation metrics
including MAE, RMSE, and R-Squared.The results demonstrated that machine learning models,
particularly LGBM, exhibit promising predictive capabilities in forecasting Bitcoin prices. LGBM
outperformed other models in terms of accuracy and efficiency, showcasing its potential as a robust tool
for cryptocurrency price prediction. Overall, our findings contribute to the growing body of literature on
cryptocurrency price prediction, offering practical implications for stakeholders in the cryptocurrency
market. As the cryptocurrency landscape continues to evolve, further research and development in
machine learning-based forecasting techniques are warranted to enhance our understanding and
predictive capabilities in this domain.

References
1. Chen, J., 2023. Analysis of Bitcoin Price Prediction Using Machine Learning. Journal of Risk and
Financial Management 2023, Vol. 16, Page 51 16, 51. https://doi.org/10.3390/JRFM16010051
2. Christoforou, E., Emiris, I.Z., Florakis, A., 2020. Neural Networks for Cryptocurrency Evaluation and
Price Fluctuation Forecasting. Springer Proceedings in Business and Economics 133–149.
https://doi.org/10.1007/978-3-030-37110-4_10/COVER

Page 11/19
 3. Huang, X., Li, Q., Tai, Y., Chen, Z., Liu, J., Shi, J., Liu, W., 2022. Time series forecasting for hourly
photovoltaic power using conditional generative adversarial network and Bi-LSTM. Energy 246,
123403. https://doi.org/10.1016/J.ENERGY.2022.123403
4. Kim, J., Kim, S., Wimmer, H., Liu, H., 2021. A Cryptocurrency Prediction Model Using LSTM and GRU
Algorithms. Proceedings - 2021 IEEE/ACIS 6th International Conference on Big Data, Cloud
Computing, and Data Science, BCD 2021 37–44. https://doi.org/10.1109/BCD51206.2021.9581397
5. Michańków, J., Sakowski, P., Ślepaczuk, R., 2022. LSTM in Algorithmic Investment Strategies on BTC
and S&P500 Index. Sensors 2022, Vol. 22, Page 917 22, 917. https://doi.org/10.3390/S22030917
6. Nasirtafreshi, I., 2022. Forecasting cryptocurrency prices using Recurrent Neural Network and Long
Short-term Memory. Data Knowl Eng 139, 102009. https://doi.org/10.1016/J.DATAK.2022.102009
7. Ortu, M., Uras, N., Conversano, C., Bartolucci, S., Destefanis, G., 2022. On technical trading and social
media indicators for cryptocurrency price classification through deep learning. Expert Syst Appl 198,
116804. https://doi.org/10.1016/J.ESWA.2022.116804
8. Patel, M.M., Tanwar, S., Gupta, R., Kumar, N., 2020. A Deep Learning-based Cryptocurrency Price
Prediction Scheme for Financial Institutions. Journal of Information Security and Applications 55,
102583. https://doi.org/10.1016/J.JISA.2020.102583
9. Rathore, M.M., Chaurasia, S., Shukla, D., Anand, P., 2023. Detection of Fraudulent Entities in Ethereum
Cryptocurrency: A Boosting-based Machine Learning Approach. GLOBECOM 2023 - 2023 IEEE
Global Communications Conference 6444–6449.
https://doi.org/10.1109/GLOBECOM54140.2023.10437184
10. Ren, Y.S., Ma, C.Q., Kong, X.L., Baltas, K., Zureigat, Q., 2022. Past, present, and future of the
application of machine learning in cryptocurrency research. Res Int Bus Finance 63, 101799.
https://doi.org/10.1016/J.RIBAF.2022.101799
11. Seabe, P.L., Moutsinga, C.R.B., Pindza, E., 2023. Forecasting Cryptocurrency Prices Using LSTM,
GRU, and Bi-Directional LSTM: A Deep Learning Approach. Fractal and Fractional 2023, Vol. 7, Page
203 7, 203. https://doi.org/10.3390/FRACTALFRACT7020203
12. Wu, C.H., Lu, C.C., Ma, Y.F., Lu, R.S., 2018a. A new forecasting framework for bitcoin price with LSTM.
IEEE International Conference on Data Mining Workshops, ICDMW 2018-November, 168–175.
https://doi.org/10.1109/ICDMW.2018.00032
13. Wu, C.H., Lu, C.C., Ma, Y.F., Lu, R.S., 2018b. A new forecasting framework for bitcoin price with LSTM.
IEEE International Conference on Data Mining Workshops, ICDMW 2018-November, 168–175.
https://doi.org/10.1109/ICDMW.2018.00032
14. Zhai, N., Yao, P., Zhou, X., 2020. Multivariate time series forecast in industrial process based on
XGBoost and GRU 1397–1400. https://doi.org/10.1109/ITAIC49862.2020.9338878

Figures

Page 12/19
Figure 1

(a) loss function history of LSTM Model in train & test, (b) loss function history of LSTM with
Autoregressive Model in train & test, (c) LSTM Model 1 prediction result and (d) LSTM with
Autoregressive Model prediction result (Wu et al., 2018a)

Page 13/19
Figure 2

Fisher’s Score

Figure 3

Historical Bitcoin price

Figure 4

Bitcoin trade volume

Page 14/19
Figure 5

Scatter Plot of Bitcoin Price vs. Trading Volume

Figure 6

Heatmap plot

Page 15/19
Figure 7

Train and Test plot

Figure 8

Page 16/19
loss train/validation of (a) LSTM Model, (b) Bi-LSTM Model, (c) LGBM Model, and (d) XGBoost Model

Figure 9

Actual vs Predicted targets of (a) LSTM Model, (b) Bi-LSTM Model, (c) LGBM Model, and (d) XGBoost
Model

Page 17/19
Figure 10

compare actual and Predicted targets

Figure 11

compare Model metrics

Figure 12

Page 18/19
show all metrics in one plot

Page 19/19