Bitcoin Price Prediction using Machine Learning

and Combinatorial Fusion Analysis

Yuanhong Wu∗ , Wei Ye† , Jingyan Xu∗ , Student Member, IEEE and D. Frank Hsu∗ , Senior Life Member, IEEE
∗
Laboratory of Informatics and Data Mining,
Department of Computer and Information Science, Fordham University, New York, NY 10023, USA.
† Department of Economics, Fordham University, New York, NY 10023, USA.

Abstract—In this work, we propose to apply a new model exchanges (DEXs) like Uniswap. Once invested in Bitcoin, an
arXiv:2602.00037v1 [[Link]] 19 Jan 2026

fusion and learning paradigm, known as Combinatorial Fusion inevitable question arises: what will the price of Bitcoin be the
Analysis (CFA), to the field of Bitcoin price prediction. Price next day? Should one buy more, hold their current position,
prediction of financial product has always been a big topic
in finance, as the successful prediction of the price can yield or sell the Bitcoin in their portfolio? To address this, a robust
significant profit. Every machine learning model has its own algorithm capable of predicting Bitcoin’s price is essential. In
strength and weakness, which hinders progress toward robust- the regard, we make use of Combinatorial Fusion Analysis,
ness. CFA has been used to enhance models by leveraging rank- a recently developed computational learning and modeling
score characteristic (RSC) function and cognitive diversity in the paradigm, to predict bitcoin price.
combination of a moderate set of diverse and relatively well-
performed models. Our method utilizes both score and rank A. Machine Learning Models
combinations as well as other weighted combination techniques.
Key metrics such as RMSE and MAPE are used to evaluate There have been extensive works on Bitcoin price pre-
our methodology performance. Our proposal presents a notable diction. One category focuses on predicting the direction of
MAPE performance of 0.19%. The proposed method greatly Bitcoin price movements (e.g., up or down), which frames the
improves upon individual model performance, as well as out-
performs other Bitcoin price prediction models. problem as a classification task. Another one aims to forecast
Index Terms—Bitcoin Price Prediction, Combining Estimators, Bitcoin prices directly on a testing dataset, which treats it as
Cognitive Diversity, Combinatorial Fusion Analysis , Rank-Score a regression problem.
Characteristic Function Previous studies have explored Bitcoin price prediction,
with some of them focusing on forecasting the next day’s
I. I NTRODUCTION Bitcoin price. For instance, the performance of time series
Bitcoin, introduced by Satoshi Nakamoto, is a peer-to- models (ARIMA) and neural networks (Neural Network Au-
peer cash payment system designed to eliminate centralization toregression) was compared, and it demonstrated that time
and serve as a next-generation currency [24]. It leverages series models are more adept at capturing Bitcoin’s volatile
blockchain technology, the proof-of-work mechanism, and price changes compared to neural networks [23]. Other studies
digital signatures to address challenges like double spending have investigated traditional machine learning algorithms, such
and the Byzantine Generals Problem. However, due to the as Random Forest (RF), Support Vector Machine (SVM),
volatility of Bitcoin’s price, it is not well-suited as a currency, Logistic Regression, and XGBoost, as well as other deep
particularly as a medium of exchange. Its volatility may occur learning methods.
during the path towards equilibrium point. However, as the Random Forest has also been shown to perform well when
future is never certain, the market may not price bitcoin Bitcoin prices are below $60,000, but struggle when prices
correctly [3]. exceed this threshold [4]. In addition, statistical methods
With the approval of spot Bitcoin exchange-traded products sometimes outperform more complex machine learning models
by the SEC on January 14, 2024 and the 2024 U.S. presidential for daily Bitcoin price prediction. For instance, a statistical
election race projection, Bitcoin prices reached a historical method has achieved an accuracy rate of 66% for this task
high of over $94,000. According to CoinMarketCap, Bitcoin’s [6]. In another work, researchers finds that LSTM outper-
market capitalization stands at $1.82 trillion, with a 24-hour forms CNN for Bitcoin price prediction, with a MAPE of
trading volume exceeding $66.43 billion. Bitcoin is now acces- 0.196%, which highlights LSTM’s ability to capture temporal
sible not only to large investment firms such as Jump Trading dependencies in time-series data [28]. Different deep learning
and BlackRock but also to the general public, as the public models have unique strengths for different tasks: LSTM excels
can purchase Bitcoin through centralized exchanges (CEXs) in regression problems, while Deep Neural Networks (DNNs)
like Coinbase and Binance or being swapped via decentralized are better suited for classification problems [16].
At the high-frequency data level, such as 5-minute time
Corresponding authors: E-mail: {ywu463; hsu}@[Link]
©2025 IEEE. Personal use of this material is permitted. Permission from intervals, GRU has shown superior predictive power. For in-
IEEE is required for all other uses. stance, a previous work achieves a Mean Squared Error (MSE)
of 0.00002 using GRU, which illustrates its effectiveness in task’s data range to start from March 11, 2020, to better
handling short-term, high-frequency Bitcoin price fluctuations capture relevant trends. As Figure 1 shows, the dataset spans
[27]. from March 11, 2020, to March 9, 2024. We partitioned the
data into a training set (magenta) and a test set (orange) with
B. Combinatorial Fusion Analysis an 80:20 ratio.
Combinatorial Fusion Analysis (CFA), introduced by Hsu,
Chung, and Kristal [9], utilizes the rank-score characteristic TABLE I
(RSC) function and models’ diversity strengths as weights to DATA S OURCE
perform score and rank combinations [12]. Data Source
The basic idea of CFA is that as all algorithms and models Bitcoin Coinmetrics
have their each respective strengths and weaknesses, superior ETH Coinmetrics
performance can be achieved through model fusion. CFA has Gold Price Yahoo Finance
Hashrate [Link]
been successfully applied in a variety of domain applications,
S&P500 Yahoo Finance
including Denial of Service (DOS) attack detection [26], target Vix Index Yahoo Finance
tracking [22], and drug discovery [18]. US bond yield Yahoo Finance
Several applications of CFA have also been explored in Dollar Strength Yahoo Finance
finance. One prominent example is its use in portfolio man- Nvidia Price Yahoo Finance
Tesla Price Yahoo Finance
agement, where CFA has been shown to enhance portfolio
management strategies in the U.S. stock market [15]. Similarly,
there has also been work to apply CFA techniques, including
rank combination, score combination, and rank-score combi-
nation, to rank U.S. equities [17]. Those results demonstrate
that CFA outperforms the iShares S&P 500 over the same
period in terms of Sharpe Ratio and cumulative return.
Current study is the first to apply the CFA method to
facilitate the prediction of Bitcoin’s next-day price. In this
approach, we obtain a scoring system for the next day’s price
from each of the five algorithms employed. The CFA method
is then utilized to combine these five scoring systems, resulting
in a single combined prediction for Bitcoin’s next-day price.
Our results demonstrate that the CFA method performs
effectively on the testing dataset, as evidenced by a lower Fig. 1. Historical Bitcoin price and data partitioning
MAPE than other models. Furthermore, our combined ap-
proach outperforms the results reported in previous studies, The features selected for prediction include ETH price
highlighting the efficacy of CFA in price prediction. and gold price, along with the hash rate, which measures
Section II provides a comprehensive overview of the miners’ computational power. These features have been proved
methodology which comprises four subsections: subsection A effective for the task of bitcoin predictions [4]. A higher hash
details the dataset processing steps; subsection B discusses rate indicates a more stable Bitcoin network. Other features
the five base models employed in this study; subsection C include the S&P 500 index and the VIX index, which measures
introduces the CFA technique; and subsection D outlines the market uncertainty. The US bond yield, considered a proxy
overall workflow of the methodology. Section III presents the for the risk-free rate, and dollar strength, which compares the
results, comparing the performance of the proposed model US dollar to a basket of other currencies, are also included.
with those of the base models and other existing research Nvidia’s price is considered due to its role as a major chip
studies. Finally, Section IV concludes the paper, summarizes provider, which correlates with miners’ demand for ASIC
the key findings, and discusses potential directions for future hardware. Finally, Tesla’s price is included as an indirect
research. measure of Elon Musk’s influence on cryptocurrencies.
Inspired by equity analyses, we also incorporate technical
II. M ETHODOLOGY
indicators such as the Exponential Moving Average (EMA)
A. Data Preparation and the Moving Average Convergence Divergence (MACD)
The dataset includes daily-level data from sources shown line.
in Table I, covering the period from August 10, 2015, to Bitcoin is traded 24/7, whereas most of the selected features
March 9, 2024. Our primary goal is to predict the next day’s are only available on weekdays. To address this discrepancy,
Bitcoin price. However, the first few years’ Bitcoin price series missing values in these features are forward-filled using the
provides limited informational gain about subsequent prices last available non-missing values. To ensure consistency, we
because the most significant Bitcoin price gains occurred dur- only use daily Bitcoin price data to align with the daily
ing the COVID-19 period. Therefore, we narrow our prediction frequency of the selected features. Since hash rate is updated
every three days, missing values are similarly forward-filled, considered commonly used base models and the effectiveness
as hash rate changes are minimal between adjacent days. of model fusion. By integrating these five models, we aim
To prepare the data for the models, we normalize each of to leverage their complementary strengths, enhancing both
the features to the range [0, 1]. accuracy and robustness in our predictive framework.
B. Base Models
C. Method of Combination
Five base models are used in our study. They are: SVM,
Random Forest, XGBoost, CNN, and LSTM. CFA presents a fresh approach to model fusion and en-
1) SVM: Support Vector Machines (SVMs) are a class of semble learning. It capitalizes on computational models that
supervised learning models widely used for classification and generate score and rank functions, which are derived by
regression tasks [1]. The central idea of SVM is to identify the respective scoring systems. The diversity between these
a hyperplane in a high-dimensional space that best separates systems can be calculated based upon the distance between the
the data into distinct classes. This hyperplane is chosen to proposed Rank-Score characteristic functions using Cognitive
maximize the margin, defined as the distance between the Diversity.
closest data points (support vectors) of each class and the 1) Score Function, Rank Function, and Rank-Score Func-
hyperplane itself. tion: Data and information fusion involves combining multiple
2) Random Forest: Random Forest is an ensemble learning scoring systems from a variety of sources such as sensor data
method that builds a multitude of decision trees during training or decision level systems. CFA can be performed on both
and aggregates their outputs to make robust predictions [2]. the attribute and decision level. It sees each system A as a
It operates by constructing each tree on a randomly selected scoring system, which has a score function sA , and a derived
subset of the data, both in terms of features and training rank function rA , and a rank-score characteristic function fA .
samples. Given a set of objects D = {d1 , d2 , . . . , dn }, scores are in the
3) XGBoost: eXtreme Gradient Boosting (XGBoost), is set of real number R, and they are normalized to the range of
a scalable and fast implementation of the gradient boosting 0 to 1. A rank function is generated from the scores in the set,
framework widely used in supervised learning tasks [5]. It with the higher scores corresponding to lower rank numbers.
builds an ensemble of decision trees sequentially, where each For an ML/AI model A, there are score function sA , rank
tree corrects the errors of the previous ones, optimizing with function rA , and RSC function which maps its relationship
second-order gradients for improved performance. in the duality of both Euclidean and rank spaces. The RSC
−1
4) CNN: Convolutional Neural Networks (CNNs) are neu- function is defined as fA (i) = sA (rA (i)). [9, 10, 12, 14]
ral networks designed for grid-like data, such as images, The RSC function, defined by Hsu, Shapiro, and Taksa,
mimicking how neurons in the brain respond to visual stimuli was used in information retrieval [13]. The RSC functions
[21]. They consist of layers like convolutional layers, which of different systems have been used to measure the dissimi-
extract features using sliding kernels, pooling layers that larity/diversity between these systems.
reduce spatial dimensions, and fully connected layers. 2) Cognitive Diversity: Cognitive diversity (CD) is a mea-
5) LSTM: Long Short-Term Memory networks (LSTMs) is surement to quantify the dissimilarity of a pair of models by
a specialized type of recurrent neural network (RNN) designed computing the area between the RSC functions of the said
to handle sequence data and capture long-term dependencies models [11]. It is dataset independent. High CD means that the
[8]. Each LSTM cell uses a gating mechanism—comprising scoring systems have relatively different ranking and scoring
input, forget, and output gates—to regulate the flow of infor- behaviors. The underlying idea is that lower diversity would
mation, making it effective for tasks like time-series predic- yield less optimal results because the output would essentially
tion. be not much different from the base models. High cognitive
We use the 10-fold cross validation and random search diversity is beneficial for combination methods because it
to optimize the models including SVM, Random Forest and allows the systems to complement each other, correcting
XGBoost. The CNN model consists of two 1D convolutional individual errors and improving overall prediction accuracy.
layers, a fully connected dense layer, and an output layer. Cognitive diversity is computed using the following for-
The LSTM model is structured with a 100-unit LSTM layer mula:
followed by a 50-unit LSTM layer, incorporating Dropout
r Pn
2
i=1 (fA (i) − fB (i))
for regularization and ending with a dense layer leading to CD(A, B) = d(fA , fB ) = (1)
n
a single-neuron output. The CNN and LSTM models are
compiled using the Adam optimizer and mean squared error For a set of t scoring systems, A1 , A2 , . . . , At , the diversity
(MSE) loss function. strength ds(Aj ) of system Aj is defined as the arithmetic
We selected our base models based on diversity and em- average of cognitive diversity between A Pj and other scoring
d(Aj ,Ak )
pirical considerations. These five models come from different systems [18]. It is written as ds(Aj ) = k̸=j t−1
domains, have distinct architectures, and utilize varied com- The performance strength p(Aj ) is determined by the per-
putational mechanisms, ensuring a broad range of learning formance of the scoring system Aj . The performance strength
capabilities. Given the large number of CFA papers, we also is assessed based on the designated metrics such as AUROC,
accuracy, or precision, depending on the task and the dataset
at hand.
3) Combination: When combining the t scoring systems
A1 , A2 , ..., At , three types of combination strategies are
considered [9, 12, 18]:
1) average combination (AC)
2) weighted combination by diversity strength (WCDS)
3) weighted combination by performance (WCP).
For average score combination (SC) or rank combination
(RC), the score function for the score combination sSC and
of the rank combination
Pt sRC are as follows
Pt [12, 18]:
j=1 sAj (di ) j=1 rAj (di )
sSC (di ) = t , sRC (di ) = t
In the case of weighted combination, whether by diversity
strength or performance, the score function for score combi-
nation and rank combination are calculatedPast [12, 1
18]:
t
j=1 wj rAj (di )
P
j=1wj ×sAj (di )
sSC (di ) = Pt , sRC (di ) = Pt 1 Fig. 2. Phase I of methodology workflow
j=1 wj j=1 wj

such that wj ∈ {pj , dsj }

D. Methodology Workflow original test set. The standard deviation reflects the variability
in a model’s predictions, indicating the ability of how different
Figure 2 illustrates the first phase of our methodology.
a given model can predict after training on a training set. We
The dataset starts from March 11, 2020, to March 9, 2024,
suppose, while the predicted prices may vary depending on the
encapsulating the entirety of the COVID-19 pandemic. To
training set, the standard deviation remains relatively stable
prepare for analysis, we divided the dataset into training and
for a given model. Consequently, we use the same standard
test sets using an 80:20 split. For time series data, employing a
deviation from the original test set for different days when
sliding window approach is crucial due to the sequential nature
building a normal distribution.
of the data, where past values influence future trends. This
method ensures that the model learns temporal dependencies
effectively, preserving the chronological order of information.
In this study, the sliding window begins on the dataset’s first
day and remains fixed for all subsequent predictions, enabling
the model to incorporate the entire historical trajectory of
the pandemic. Consequently, predicting a single day’s Bitcoin
price leverages all preceding days’ data. As we progress
toward predicting more recent dates, the training set expands
to include additional days. Following this preparation, the five
models described in Section II(B) are applied to forecast daily
Bitcoin prices, as shown in Figure 2.
Similar to prior research efforts in Bitcoin price prediction,
we train machine learning models on a designated training set
and evaluate their performance on a separate test set. How-
ever, our approach distinguishes itself by generating a price
distribution for each day’s prediction, rather than predicting Fig. 3. Phase II of methodology workflow
a single deterministic value. From a statistical perspective,
traditional machine learning models focus on estimating the After generating five normal distributions for each day’s
most probable price for a given day, which corresponds to the prediction, we truncate these distributions at two standard
price with the highest likelihood of occurrence. Our method deviations from the mean, capturing approximately 95% of
aims to uncover the range of potential prices that could the possible price range. The smallest and largest prices of
plausibly occur. By identifying these distributions for each the five price ranges define the final range. If the lower bound
model, we leverage CFA to integrate the predictions from five of the final interval is negative, we set the minimum price to
distinct models, enabling a more robust and comprehensive zero, as shown in Phase II’s table, since negative prices are
approach to price forecasting. not meaningful in this context. It is important to note that the
As illustrated in Phase II of Figure 3, each daily price precise lower bound is not critical because the final predicted
prediction from the five base models is expanded into a normal price is determined by the value with the highest probability
distribution. The predicted price serves as the mean of the within the interval after combination. The probabilities derived
distribution, while the standard deviation is derived from the from the five normal distributions are treated as the scores
for each model, while the prices within the intervals serve as
the data items. This process establishes five distinct scoring
systems, which are then utilized in the subsequent model
combination phase.
Phase III in Figure 4 describes the process of using CFA,
where models are systematically combined to enhance predic-
tion accuracy. The use of combination techniques avoids the
existence of extreme points. We first construct combinatorial
groups comprising subsets of the five base models, ranging
from pairs to all five models together. This process results
of
26 unique model groups, calculated as 52 +


in
a total
5 5 5


3 + 4 + 5 = 26. For each group, we explore four distinct
strategies for combining models based on input type (scores or
ranks) and weighting approach (average weighting or weighted
combination using diversity strength). These strategies include
both average score and rank combinations as well as weighted Fig. 5. Phase IV of methodology workflow
score and rank combinations by diversity strength. As it is not
possible to calculate the performance between multiple prices
In Phase IV of Figure 5, we first determine the absolute
(from a normal distribution) and a single true price, perfor-
distance between the actual price and the predicted price for
mance weighting is not applied in this paper. Consequently, the
each model combination. A smaller absolute distance indicates
26 model groups, when combined with these four strategies,
a more accurate predicted price. As illustrated in Figure 5, two
yield 26 × 4 = 104 unique combination models.
scatterplots are provided—one for the average combination
method and another for the weighted combination by diversity
strength method—each corresponding to daily predictions.
Figure 5 also includes a rank-score function to represent the
diversity among the five base models. Previous research has
shown that model combinations outperform individual models
when those models exhibit both relatively high predictive
performance and large diversity. Consistent with this principle,
the five models demonstrate considerable diversity during
the first two days of the analysis. Each scatterplot’s x-axis
displays five individual models followed by 26 model groups.
The score combination method is depicted as a blue solid
line, and the rank combination method is depicted as a red
dashed line. Our aim is to identify the smallest absolute
distance for each combination method, which serves as the
final predicted price. In the figures, an orange circle and a
green circle highlight the final predicted price for the rank and
score combination methods, respectively. If the chosen final
predicted price corresponds to a model group rather than an
individual model, this outcome indicates that the combination
approach has conferred a performance improvement on that
particular day. By counting the number of days in which a
Fig. 4. Phase III of methodology workflow combination strategy shows improvement, one can assess the
effectiveness of that strategy. At the end of Phase IV, we thus
For each combination, we rank the resulting probabilities determine the final predicted price for each day using four
in descending order to identify the price with the highest distinct model combination methods.
probability as the predicted value for a given day. This process
III. R ESULTS
generates the table in Phase III, where the columns represent
the 26 model groups and the rows correspond to the prediction To forecast Bitcoin prices, we produced a price distribution
dates. Since there are four combination strategies, we obtain for each day by using the predicted price from our models
four analogous tables, each representing a specific approach. as the mean and the standard deviation derived from the test
These tables form the foundation for selecting the most set. CFA was employed to integrate these distributions and
accurate Bitcoin price prediction for each day by analyzing determine the optimal predicted price for each combination
the corresponding row entries. method. To evaluate the predictive performance of both in-
 TABLE II
P ERFORMANCE FOR INDIVIDUAL MODELS AND MODEL COMBINATIONS .

Base Models AC WCDS

A B C D E SC RC SC RC
# days∗ - - - - - 215 222 208 258
RMSE 1057.34 738.21 966.64 2331.47 1967.18 175.22 289.45 182.28 294.22
MAPE 1.86% 1.20% 1.55% 4.91% 4.68% 0.19% 0.40% 0.22% 0.44%
* indicates the number of days that show improvement out of 292 days.

dividual models and combination models, we utilized two RMSE of 738.21. Interestingly, while score-based combina-
widely accepted metrics: Root Mean Squared Error (RMSE) tions exhibited lower RMSE values than rank-based combina-
and Mean Absolute Percentage Error (MAPE). The formulas tions, rank-based combinations showed improvements on more
for these metrics are as follows. days. This discrepancy can be attributed to the characteris-
v
u N
tics of rank-based strategies, which are more likely to yield
u1 X consistent improvements across a greater number of days but
RM SE = t (yi − ŷi )2 ,
N i=1 with a slightly smaller magnitude of improvement compared
N
to score-based strategies.
1 X yi − ŷi In terms of MAPE, the best performance achieved among
M AP E = × 100%
N i=1 yi the model combinations is 0.19%, obtained using the average
score combination method. This result is approximately ten
RMSE and MAPE offer distinct advantages and serve com- times smaller than the MAPE values of any individual base
plementary purposes. RMSE, expressed in the same unit as the models in Table II, highlighting the effectiveness of the com-
data (e.g., U.S. dollars), is sensitive to the scale of the dataset. bination strategy. Similar to the results observed with RMSE,
For instance, if Bitcoin prices from the early 2010s—when score combination outperforms rank combination. MAPE is a
prices were relatively low—are used, RMSE will naturally particularly suitable metric for comparing model performance
be smaller. Consequently, comparing RMSE across datasets across datasets, as it provides a standardized measure that is
with differing scales is not meaningful. In contrast, MAPE is independent of scale. In Table III, we present several refer-
unitless, expressed as a percentage, and remains unaffected by ence studies that utilize daily Bitcoin prices as their dataset
the scale of the dataset. This characteristic allows MAPE to and evaluate model performance using MAPE. The MAPEs
facilitate comparisons across diverse datasets. In this study, reported in these studies range from 0.245% to 4.49%, all of
RMSE is employed to compare models within our analysis, which are higher than the MAPE achieved by our proposed
while MAPE is used to benchmark our models against state- method. Of particular interest in Table III is an advanced en-
of-the-art models presented in other research studies. The dual semble model named VMD-AGRU-RESVMD-LSTM, which
use of RMSE and MAPE ensures both intra-study and inter- integrates variational mode decomposition (VMD), the gated
study comparisons are robust and meaningful. recurrent unit (GRU), long short-term memory (LSTM) neural
In additional, we measure the effectiveness of each combi- networks, and attention mechanisms to provide accurate Bit-
nation strategy in CFA using the number of days of improve- coin price predictions. Despite its sophisticated architecture,
ment which are defined before. A higher number of improved our CFA technique delivers almost twice the performance
days indicates that a specific combination strategy is beneficial improvement over their reported result, demonstrating the effi-
across a greater portion of the test set. cacy and potential of our approach for Bitcoin price prediction.
The performance results are summarized in Table II. For Average score combination is the best performing methodol-
the metric of the number of days of improvement, all four ogy in both RMSE and MAPE metric. The underperformance
combination methods demonstrated improvements for over of rank combinations, compared with score combinations, may
200 days out of the 292-day test set, emphasizing the efficacy be attributed to the fact that we have fluctuating diversity
of the CFA approach. Rank-based strategy outperformed the strength values for each model. For example, if diversity
score-based strategy for both types of combination in terms of strength values of the five base models for most of the data
the number of improved days. Specifically, the weighted rank items are low, then the rank combinations won’t have superior
combination by diversity strength showed improvements on performance than the score combinations [10, 13].
258 days, whereas the weighted score combination showed
improvements on only 208 days, which marks a significant IV. C ONCLUSION
improvement. These results clearly indicate the robustness of In this article, we proposed a comprehensive approach to
weighted rank combination by diversity strength method. predict Bitcoin prices by leveraging diverse features and a
For the RMSE metric, all model combinations demonstrated moderate set of machine learning models. Five distinct ma-
superior performance compared to individual models. The chine learning algorithms were employed for raw Bitcoin price
lowest RMSE, achieved by average score combination, was predictions. Given the time-series nature of the price dataset,
175.22, a substantial improvement over the base models’ we included Long Short-Term Memory (LSTM) networks to
 TABLE III
C OMPARISON OF THE PROPOSED METHOD AND PREVIOUS METHODS IN B ITCOIN PRICE PREDICTION .

Previous Work Year Method Dataset Metric

Ji, Kim, and Im 2019 Deep neural network (DNN) Daily Bitcoin price data and MAPE: 3.61%
[16] blockchain information from
29 November 2011 to 31 De-
cember 2018
Ye, Wu, Chen, 2021 Stacking ensemble deep model Tweets, transaction data, tech- MAPE: 4.49%
et al. [29] of 2 base models: LSTM & nical data from September
GRU 2017 to January 2021 daily
data
Hamayel and 2021 gated recurrent unit (GRU) Daily Bitcoin price using open, MAPE: 0.245%
Owda [7] high, low, close features from
22 January 2018 to 30 June
2021
Zhang, Li, and 2022 stacking denoising auto- 9 features of daily bitcoin price MAPE: 1.6%
Yan [30] encoders using bootstrap data from 29 November 2014
aggregation (SDAE-B) to 30 March 2020
Kim, Shin, 2022 self-attention-based multiple Daily Bitcoin price on-chain MAPE: 1.33%
Choi, et al. long short-term memory data from 27 March 2018 to 16
[20] (SAM-LSTM) November 2021
Chen [4] 2023 Random forest Daily Bitcoin price having 47 MAPE: 3.29%
features, split into 8 categories
from 1 October 2018 to 1 April
2022
Jin and Li [19] 2023 hybrid prediction model, Daily Bitcoin price from 31 MAPE: 0.394%
VMD-AGRU-RESVMD- July 2017 to 30 September
LSTM, integrating variational 2020
mode decomposition (VMD),
the gated recurrent unit (GRU)
and long short-term memory
(LSTM)
Proposed 2025 Combinatorial fusion analysis Daily Bitcoin price from 11 MAPE:
method (CFA) of 5 base ML/AI models March 2020 to 09 March 2024 0.19%(SCAC)

capture temporal dependencies effectively. Moreover, a Con- improvement than score-based methods. Notably, our best-
volutional Neural Network (CNN) was utilized to introduce performing strategy, average score combination, achieved a
more diversity among base models by leveraging its ability to MAPE of 0.19%, lower than results reported in prior studies.
detect localized patterns within the data. A key innovation of This highlights the superiority and robustness of our approach,
our approach lies in extending Bitcoin price prediction beyond as it effectively combines diverse modeling perspectives into
traditional regression frameworks by generating prediction dis- a unified and high-performing framework.
tributions for each day’s price. Through Combinatorial Fusion Using sentiment analysis to do price prediction has been
Analysis, these distributions and their corresponding scoring a constant area of focus, as people’s opinions reflect policy
systems were combined to produce an optimized predicted changes and thus can sway the movement of prices [25].
price. This technique transcends the performance ceiling of Our future work should take into the account of consumer
traditional models, demonstrating that combining prediction sentiment to more accurately predict the prices. Additionally,
distributions via CFA can push the boundaries of accuracy we aim to investigate multi-layer CFA frameworks [14] to
and reliability in Bitcoin price forecasting. further refine Bitcoin price predictions and unlock even greater
Performance was evaluated using Root Mean Square Er- forecasting precision.
ror (RMSE) and Mean Absolute Percentage Error (MAPE), A potential limitation of our work is using the test set’s
with benchmark results from previous studies included for standard deviation for normal distribution generation, which
comparison. Our findings show that CFA-based model com- may cause data leakage. In future work, a separate validation
binations consistently outperform individual models, illus- set can be used to calculate the standard deviation instead.
trating the power of model fusion in improving predictive
R EFERENCES
accuracy. Furthermore, score-based combinations deliver more
precise predictions than rank-based combinations, although [1] B. E. Boser, I. M. Guyon, and V. N. Vapnik, “A training
rank-based methods still demonstrate higher likelihood for algorithm for optimal margin classifiers,” Proc. of the
 Fifth Annual Workshop on Computational Learning [17] N. Jiang, “Equity ranking with stochastic dominance
Theory, pp. 144–152, 1992. and combinatorial fusion,” Ph.D. dissertation, Fordham
[2] L. Breiman, “Random forests,” Mach. Learn., vol. 45, University, 2023.
no. 1, 2001. [18] N. Jiang, M. Quazi, C. Schweikert, D. F. Hsu, T. I.
[3] C. Catalini, A. de Gortari, and N. Shah, “Some simple Oprea, and S. Sirimulla, “Enhancing admet property
economics of stablecoins,” Annual Review of Financial models performance through combinatorial fusion anal-
Economics, vol. 14, no. 1, pp. 117–135, 2022. ysis,” ChemRxiv, 2023. DOI: 10.26434/chemrxiv-2023-
[4] J. Chen, “Analysis of bitcoin price prediction using ma- dh70x
chine learning,” J. of Risk and Financial Management, [19] C. Jin and Y. Li, “Cryptocurrency price prediction using
vol. 16, no. 1, p. 51, 2023. frequency decomposition and deep learning,” Fractal
[5] T. Chen and C. Guestrin, “Xgboost: A scalable tree and Fractional, vol. 7, no. 10, p. 708, 2023.
boosting system,” in Proceedings of the 22nd acm [20] G. Kim, D.-H. Shin, J. G. Choi, and S. Lim, “A
sigkdd international conference on knowledge discovery deep learning-based cryptocurrency price prediction
and data mining, 2016, pp. 785–794. model that uses on-chain data,” IEEE Access, vol. 10,
[6] Z. Chen, C. Li, and W. Sun, “Bitcoin price prediction pp. 56 232–56 248, 2022.
using machine learning: An approach to sample di- [21] Y. LeCun et al., “Backpropagation applied to handwrit-
mension engineering,” J. of Computational and Applied ten zip code recognition,” Neural Comput., vol. 1, no. 4,
Mathematics, vol. 365, p. 112 395, 2020. pp. 541–551, 1989.
[7] M. J. Hamayel and A. Y. Owda, “A novel cryptocur- [22] D. Lyons and D. F. Hsu, “Combining multiple scoring
rency price prediction model using GRU, LSTM and bi- systems for target tracking using rank–score character-
LSTM machine learning algorithms,” AI, vol. 2, no. 4, istics,” Information Fusion, vol. 10, no. 2, pp. 124–136,
pp. 477–496, 2021. 2009.
[8] S. Hochreiter and J. Schmidhuber, “Long short-term [23] Z. H. Munim, M. H. Shakil, and I. Alon, “Next-
memory,” Neural Comput., vol. 9, no. 8, pp. 1735–1780, day bitcoin price forecast,” J. of risk and financial
1997. management, vol. 12, no. 2, p. 103, 2019.
[9] D. F. Hsu, Y.-S. Chung, and B. S. Kristal, “Combinato- [24] S. Nakamoto, “Bitcoin: A peer-to-peer electronic cash
rial fusion analysis: Methods and practices of combin- system,” Satoshi Nakamoto, 2008.
ing multiple scoring systems,” in Advanced data mining [25] S. O. Ouf, M. El Hawary, A. Aboutabl, and S. Adel, “A
technologies in bioinformatics, IGI Global Scientific deep learning-based LSTM for stock price prediction
Publishing, 2006, pp. 32–62. using twitter sentiment analysis,” Int. J. of Adv. Comp.
[10] D. F. Hsu, B. S. Kristal, and C. Schweikert, “Rank-score Sci. and App. (IJACSA), vol. 15, no. 12, 2024.
characteristics (RSC) function and cognitive diversity,” [26] E. Owusu, M. Rahouti, D. F. Hsu, K. Xiong, and
Brain Informatics: Int. Conf., BI 2010, pp. 42–54, 2010. Y. Xin, “Enhancing ML-based DOS attack detection
[11] D. F. Hsu, B. S. Kristal, Y. Hao, and C. Schweikert, through combinatorial fusion analysis,” in 2023 IEEE
“Cognitive diversity: A measurement of dissimilarity Conference on Communications and Network Security
between multiple scoring systems,” J. of Interconnection (CNS), IEEE, 2023, pp. 1–6.
Network, vol. 19, no. 1, 2019. [27] T. Phaladisailoed and T. Numnonda, “Machine learning
[12] D. F. Hsu, B. S. Kristal, and C. Schweikert, “Combina- models comparison for bitcoin price prediction,” 2018
torial fusion analysis,” IEEE Computer, 2024. 10th Int. Conf. on Info. Tech. and Electrical Engineering
[13] D. F. Hsu, J. Shapiro, and I. Taksa, “Methods of (ICITEE), pp. 506–511, 2018.
data fusion in information retrieval: Rank vs. score [28] G. Singh, A. K. Sahoo, G. Yadav, A. K. Gairola, P. K.
combination,” DIMACS, 2002. Sarangi, and J. Gupta, “Analyzing the effectiveness of
[14] L. Hurley, B. S. Kristal, S. Sirimulla, C. Schweikert, and machine learning models for next day cryptocurrency
D. F. Hsu, “Multi-layer combinatorial fusion using cog- price prediction,” in 2023 4th International Confer-
nitive diversity,” IEEE Access, vol. 9, pp. 3919–3935, ence on Intelligent Technologies (CONIT), IEEE, 2024,
2021. pp. 1–5.
[15] J. Irukulapati, D. F. Hsu, and C. Schweikert, “Long- [29] Z. Ye, Y. Wu, H. Chen, Y. Pan, and Q. Jiang, “A
term portfolio management using attribute selection and stacking ensemble deep learning model for bitcoin
combinatorial fusion,” 2018 IEEE 17th Int. Conf. on price prediction using twitter comments on bitcoin,”
Cognitive Informatics & Cognitive Computing (ICCI* Mathematics, vol. 10, no. 8, p. 1307, 2022.
CC), pp. 593–599, 2018. [30] S. Zhang, M. Li, and C. Yan, “The empirical analysis
[16] S. Ji, J. Kim, and H. Im, “A comparative study of bit- of bitcoin price prediction based on deep learning
coin price prediction using deep learning,” Mathematics, integration method,” Computational Intelligence and
vol. 7, no. 10, p. 898, 2019. Neurosci., vol. 2022, no. 1, p. 1 265 837, 2022.