Stock Prediction Based on Traditional Statistical Models, Machine
Learning Models and Fusion Models
Yu Du
School of Data Science, Capital University of Economics and Business, 100000, Beijing, China
Keywords: Stock Prediction, Statistical Models, Machine Learning Models, Fusion Models.
Abstract: This study aims to evaluate how well machine learning (ML) algorithms and classic time series analysis
methods can forecast stock market trends. Accurate forecasts of stock prices can greatly aid professionals and
investors in making strategic decisions owing to the unpredictable nature of the stock market. This research
aims to create a composite model that combines the accuracy of traditional statistical models, which are good
at making short-term predictions, with the capabilities of machine learning models that can handle large
amounts of complex and nonlinear data. The goal is to enhance the precision of long-term stock price forecasts.
This research aims to assess the strengths and weaknesses of four distinct models: Autoregressive Integrated
Moving Average(ARIMA), Generalized Autoregressive Conditional Heteroskedasticity(GARCH), Long
Short-Term Memory (LSTM), and Random Forest(RF) through training and evaluation with historical stock
market data. Additionally, a comparison between these distinct models and an integrated model will be
conducted as part of the investigation to develop a more reliable tool for informing investment decisions.
1. INTRODUCTION
The stock market serves as a barometer for a nation's
economic and fiscal dynamics (Lu et al., 2021).
Consequently, investors are highly concerned about
the future trend of stock values (Li, Pan and Huang,
2019). Nevertheless, numerous variables influence
stock prices, including alterations in national policies,
fluctuations in the local and global economic
landscape, and shifts in the international situation
(Sim, Kim and Ahn, 2019). Consequently,
forecasting stock prices is a formidable undertaking.
Producing accurate predictions of stock prices can
greatly reduce the level of risk for investors. These
forecasts allow investors to integrate projected stock
values into their investing strategy, so increasing the
potential for enhanced investment returns (Lu et al.,
2021).
In recent times, a wide range of approaches and
frameworks have emerged for forecasting stock
values. The methodologies can be broadly
categorized into two types: classic statistical models
and machine learning-based models. When it comes
to predicting time series, particularly for short-term
predictions, the ARIMA model is widely regarded as
more resilient and effective compared to the most
used artificial neural network techniques. Other
statistical models are generalized autoregressive
conditional heteroskedasticity (GARCH) regression
and exponential smoothing (Ariyo et al., 2014).
However, statistical models ignore the effects of
external factors other than the time factor and are all
based on the premise that there will be no sudden
changes in the market in the future, so statistical
models alone are not sufficient for some special cases.
Machine learning advancements have facilitated the
utilization of ML methods such as random forests and
LSTM networks to decipher complex nonlinear
patterns in financial datasets. The LSTM network, a
variant of a recurrent neural network, is highly
proficient in many applications due to its ability to
accurately differentiate between recent and past data
points by assigning distinct weights and selectively
eliminating irrelevant information that is not crucial
for future predictions. Unlike other types of recurrent
neural networks that primarily handle short-term data
sequences, this specific model excels at managing
longer input sequences, making it more suitable for
applications that require retaining substantial history
information. As a result, it is highly effective at
predicting stock prices when applied to nonlinear
datasets with a huge volume of data (Sunny et al.,
2020 & Nelson et al., 2017). Conversely, the
160
Du, Y.
Stock Prediction Based on Traditional Statistical Models, Machine Learning Models and Fusion Models.
DOI: 10.5220/0013007100004601
Paper published under CC license (CC BY-NC-ND 4.0)
In Proceedings of the 1st International Conference on Innovations in Applied Mathematics, Physics and Astronomy (IAMPA 2024), pages 160-168
ISBN: 978-989-758-722-1
Proceedings Copyright © 2024 by SCITEPRESS – Science and Technology Publications, Lda.

substantial level of noise and frequent fluctuations in
crucial characteristics within the stock market render
stock prediction intricate and inefficient. Random
forests have the ability to conduct feature analysis,
which quantifies the significance of each input
feature. Utilizing Random Forest (RF) for feature
extraction can enhance the precision of stock price
forecasts (Ma, Han and Fu, 2019).
This work presents a novel strategy that integrates
statistical and machine learning method to overcome
the shortcomings of the previously mentioned
models, aiming to enhance the precision of predictive
analysis. A fusion model is created by training an
LSTM model using the outputs of ARIMA, GARCH,
and Random Forest models as features. ARIMA, a
conventional statistical model, is ideal for predicting
short-term outcomes, whereas machine learning
models like LSTM are better suited for analyzing
extensive datasets with nonlinear patterns. Therefore,
the objective of this thesis is to investigate whether
the results of the fusion of the two types of models
outperform the separate models for their respective
predictions. An LSTM model trained with the outputs
of ARIMA, GARCH and Random Forest models as
features will be used as the fusion model. To
investigate the merits and demerits of these methods
as well as the possibility of a hybrid approach to
predicting stock prices, this study will evaluate each
of these models individually and compare the
prediction results of the individual models with their
fusion model in order to explore the most accurate
stock prediction model.
The major contributions of this paper are as
follows:
1. Development of a novel fusion model
combining ARIMA, GARCH, LSTM, and Random
Forest to enhance stock price prediction accuracy.
2. An extensive evaluation and comparative
analysis of traditional statistical models, machine
learning models, and the proposed fusion model
based on historical stock market data.
3. Demonstration of the effectiveness of feature
engineering and integrated learning phases in
improving prediction performance.
4. Provision of a more reliable and precise tool
for stock market analysts and investors to make
informed decisions in volatile financial markets.
The manuscript is structured as follows: Section 2
provides a review of related work in the fields of
statistical modeling and machine learning methods
for stock price prediction. Section 3 details the
methodology, including the construction and
development of the ARIMA, GARCH, LSTM, and
Random Forest models, as well as the fusion model.
Section 4 discusses the experimental procedure, data
pre-processing, and evaluation metrics used in this
study. Section 5 presents the results and comparisons
of the models. Finally, Section 6 concludes the paper
and suggests directions for future research.
2. RELATED WORK
 Statistical Modeling in Stock Price Prediction
Model Study
ARIMA modeling is widely regarded as an
exceptionally efficacious forecasting methodology
within the domain of stock forecasting. As its
predictions are derived from the values of the input
variables and the error term, ARIMA forecasting does
not necessitate the presupposition of any underlying
model or associated equations. However,
sophisticated nonlinear real-world problems may
introduce some bias into the ARIMA model due to
the fact that it is a linear regression model. However,
it is generally observed that linear models outperform
complex structural models when it comes to short-
term forecasting (Ma, 2020). A method for
forecasting the price of garlic was introduced by Yan
W. et al. (Wang et al., 2022). This method utilized a
combination of GARCH family models and LSTM.
By constructing a GARCH family model, they
acquired data on volatility characteristics, including
volatility aggregation, of garlic price series. The
LSTM network was employed to examine the
complex nonlinear interactions between sequences of
garlic prices and their intrinsic volatility, aiming to
forecast subsequent garlic price trends. Resulting
from the independence of the machine learning
models, the fusion model proves to be effective. The
study by Yan W. et al. and the anticipated stock price
prediction model in this study share some similarities
(Wang et al., 2022); this provides the inspiration for
the concept of model fusion in this article.
 Research into machine learning methods for
forecasting stock prices
LSTM is crafted as a variant of recurrent neural
networks, especially skilled in handling and
predicting major events in time series data marked by
substantial intervals and periods. Jin, Z. et al. noticed
the advantages of the LSTM model in analyzing the
relationship between time series and adapted the
LSTM model by using the attention mechanism to
predict the closing price with greater precision (Jin et
al., 2020). Park H et al. introduced a new stock
prediction framework called LSTM-Forest, which
combines LSTM and Random Forest to address the
Stock Prediction Based on Traditional Statistical Models, Machine Learning Models and Fusion Models
161

issue of overfitting in prediction models (Park, Kim
and Kim, 2022). RF can handle very many features to
avoid the overfitting problem and LSTM outperforms
the decision tree in terms of temporal patterns, so
their method can be used with ensure that no useful
information is lost to mitigate overfitting and
consequently enhance the efficacy of predictions.
According to a study by Y. Ma et al., the majority of
researchers who employed LSTM neural networks to
forecast stock prices trained the networks using
unprocessed stock data (Ma, Han and Fu, 2019). This
approach resulted in the training model absorbing a
substantial amount of noise and ultimately
diminished the model's predictive performance.
Principal Component Analysis (PCA) and Random
Forest were employed to identify critical input
features, thereby enhancing the stock price prediction
performance. The outcomes demonstrate that the
stock prediction model constructed using Random
Forest and LSTM yields more accurate predictions.
3. METHODOLOGY
3.1. Model Construction
 ARIMA
In order to normalize non-stationary time series data,
the ARIMA model integrates the concepts of
differencing (I) and autoregressive (AR) and moving
average (MA) models.Three primary parameters
define the ARIMA model: 𝑝, 𝑑, 𝑞
𝑝: the number of autoregressive terms. The
autoregressive component is the effect of past values
on current values in the model.
𝑑 : Represents the number of non-seasonal
differencing operations applied to stabilize the time
series.
𝑞: Number of moving average terms. The moving
average part models the effect of the forecast error
term.
The model structure of ARIMA can be
represented by the following equation:
(1 −𝐵)

𝑌

= 𝛿+ 𝜙

𝑌

+ 𝜙

𝑌

+ ⋯+ 𝜙

𝑌

−𝜃

𝑒

−𝜃

𝑒

−⋯−𝜃

𝑒

+ 𝜖

(1)
In this formulation, 𝑌

signifies the value of the
time series at time 𝑡; 𝐵 acts as the backward shift
operator, 𝐵𝑌

= 𝑌

;
(
1 −𝐵
)

represents the
differencing operator to make the time series
stationary; 𝛿 is a constant term; The autoregressive
components are indicated by 𝜙

, 𝜙

,…,𝜙

, and the
parameters of the moving average section are denoted
as 𝜃

, 𝜃

,…,𝜃

; 𝜖

denotes the error at time 𝑡, which
adheres to a normal distribution with a mean of zero
and a variance of 𝜎

; 𝑒

represents the forecast errors
at time 𝑡.
 GARCH
Tim Bollerslev proposed the GARCH model in 1986.
This model is a statistical framework used for
analyzing time series data and is an extension of the
ARCH model, which was initially developed by
Robert F. Engle in 1982, which aimed to model the
conditional variance fluctuations over time or their
aggregation. The GARCH model, in essence,
leverages historical fluctuations to predict
autoregressive changes in volatility. It is effective in
addressing issues like heteroskedasticity, volatility
aggregation, leverage effect, asymmetric effects, and
can closely capture the volatility dynamics of time
series data (Wang et al., 2022).
A standard GARCH(𝑝, 𝑞) model can be described
using the following formula:
𝑦

= 𝜇

+ 𝜎

𝜂

𝜀

= 𝜎

𝜂

(2)
𝜎


= 𝛼

+
∑


𝛼

𝜖


+
∑


𝛽

𝜎


(3)
The variables denoted as follows: 𝜂

denotes the
variables that are identically and independently
distributed, 𝜂

∼𝑁
(
0,1
)
; 𝜎


outlines the conditional
variance at time 𝑡; 𝛼

represents the constant; 𝛼

represents the coefficients of the ARCH terms; 𝛽

details the coefficients of the GARCH terms; and 𝜀

captures the residual at time 𝑡.
 LSTM
An enhanced iteration of the RNN method, LSTM
was initially introduced by Hochreiter and
Schmidhuber in 1997. LSTM introduces the
mechanism of "gates", which can effectively control
the forgetting and remembering of information, so
that the network can still maintain a stable gradient
flow in long sequences, and thus capture long-
distance data dependencies.
A standard LSTM unit is composed of four key
components: the forget gate, the input gate, the cell
state, and the output gate (Park, Kim and Kim 2022
& Hochreiter et al., 1997). The architecture of a
traditional LSTM is depicted in Figure 1. The
activation value of the Forget Gate is 𝑓

; the
activation value of the Input Gate is 𝑖

, which
IAMPA 2024 - International Conference on Innovations in Applied Mathematics, Physics and Astronomy
162

determines the number of new candidate values added
to the cell state from 𝑔

; the activation value of the
Output Gate is 𝑜

; the cell state at the current time
step is 𝑐

, and the output of the current time step is
ℎ

.
Figure 1. LSTM (Park, Kim and Kim, 2022).
 RF
The RF approach is extensively utilized for machine
learning tasks in classification and regression. It is a
technique called ensemble learning that improves the
reliability and precision of a prediction model by
combining the results of numerous decision trees to
make a final conclusion. This unique framework
enables the mitigation of flaws in individual
classifiers and the amalgamation of the strengths of
each classifier, hence enhancing prediction accuracy
and managing overfitting.
The fundamental concept of random forest
encompasses two key elements: the random selection
of samples and the random selection of attributes
from the data. Bootstrap Sampling is used to
randomly choose data, whereas the random selection
of features is achieved by determining which features
are utilized as the training set during the division of
each node in the decision tree.
The most important parameters for implementing
the random forest model include:
Augmenting the quantity of decision trees boosts
the stability of the model, but at the expense of
increased computing requirements. The quantity of
features considered during the splitting of each node
in a decision tree influences the range of features
sampled by the model, impacting both the bias and
variance of the model's outcomes (Genuer et al., 2010
& Breiman, 2001).
3.2. Model Development
Machine Learning Models: LSTM models excel in
capturing long-term correlations within the time
series data of stock prices and adeptly managing
nonlinear patterns. (Jin, Yang and Liu, 2020). On the
other hand, RF models provide feature importance
assessment to identify the key factors affecting stock
prices and are able to avoid overfitting problems. The
combination of these two models enhances the
generalization ability of the model by effectively
handling nonlinear patterns and high dimensional
data. Therefore, the fusion model of Random Forest
and LSTM may have better prediction results (Park,
Kim and Kim 2022 & Ma, Han and Fu, 2019).
Time Series Models: ARIMA models are used to
capture linear trends in time series, for example, to
analyze and forecast specific seasonal characteristics
and cyclical movements or trends of the stock market.
Conversely, GARCH models are predominantly
employed for the purpose of predicting the volatility
of stock returns in the forthcoming periods. The
convergence of GARCH and ARIMA models
improves the accuracy of forecasts, specifically with
regard to identifying patterns of volatility within the
dataset.
3.2.1. Model Fusion Strategies
The GARCH model may determine the volatility
characteristics of the series for the LSTM and
ARIMA models. By combining these three models,
more accurate prediction results can be obtained
(Wang et al., 2022). The combination of the RF
model and the LSTM model is capable of effectively
handling nonlinear patterns and high-dimensional
data, while also improving the model's generalization
capacity (Bollerslev, 1986 & Engle, 1982).
3.2.2. Feature Engineering Stage
First, the historical stock price data are modeled and
forecasted using ARIMA and GARCH models,
respectively. The ARIMA model outputs the
predicted future values and the GARCH model
outputs the predicted volatility estimates.
Second, the historical stock price data are fitted
with a random forest model in order to evaluate the
significance of various features, including those
produced by ARIMA and GARCH and other stock
metrics.
3.2.3. Integration Learning Phase
First, integrating ARIMA and GARCH: The forecasts
from the ARIMA and GARCH models are combined
to form a comprehensive set of time series features.
Subsequently, the features identified by the
random forest and the combined outputs from
ARIMA and GARCH models are utilized as input
data to enhance the LSTM model's training process.
Stock Prediction Based on Traditional Statistical Models, Machine Learning Models and Fusion Models
163

Finally, this specifically tailored dataset is
employed to train the LSTM model, which is then
used to forecast stock prices.
4. EXPERIMENTAL PROCEDURE
AND ANALYSIS OF RESULTS
4.1. Data and Data Pre-Processing
Data sets and experimental tools: The original data
was collected from Yahoo Finance. The study used
stock market data from Apple, Google and Amazon
for a total of 2,571 days between January 2, 2014 and
March 20, 2024. The data is in numerical format and
includes detailed data for each day of the historical
stock. All experiments were implemented in Python
3.11's Jupyter Notebook.
The acquired data will undergo a rigorous
cleaning, normalization, and transformation process
to make it suitable for time series analysis and
machine learning modeling. Since the original stock
data is very clean and complete, this study does not
need to do much data preprocessing. Furthermore, to
boost the precision of the forecasting model, this
study incorporates supplementary indicators like
Simple Moving Average (SMA), Exponential
Moving Average (EMA), Bollingband, etc., to
describe the dataset. Technical indicators offer vital
insights into market patterns, volatility, and
momentum, hence enhancing the precision and
dependability of forecasts.
4.2. Evaluation Metrics
The study assessed the model's prediction ability by
employing four distinct evaluation measures: mean
absolute error (MAE), root mean square error
(RMSE), mean absolute percentage error (MAPE),
and R square. A prediction model's performance is
considered better when the values of MAE, RMSE,
and MAPE are lower. Additionally, a value of R
square close to 1 indicates a high degree of fitting
between the model and the data. The four evaluation
indicators are calculated as follows.
MAE =
𝟏
𝒎
∑
𝒎
𝒊𝟏
|
(
𝒚
𝒊
−𝒚

𝒊
)
| (4)
𝐑𝐌𝐒𝐄=

𝟏
𝒎
∑
𝒎
𝒊𝟏
(
𝒚
𝒊
−𝒚

𝒊
)
𝟐
(5)
𝐌𝐀𝐏𝐄=
𝟏𝟎𝟎
𝒎
∑
𝒎
𝒊𝟏
|
(𝒚
𝒊
𝒚

𝒊
)
𝒚
𝒊
| (6)
𝐑
𝟐
= 𝟏−

∑
𝒎
𝒊𝟏
(
𝒚
𝒊
𝒚

𝒊
)
𝟐

/𝒎

∑
𝒎
𝒊𝟏

𝒚
𝒊
¯
𝒚

𝒊

𝟐

/𝒎
(7)
The objective of this research is to clarify the
relative advantages and disadvantages of machine
learning methods and time series models as they
pertain to the forecasting of stock prices. Given its
ability to capture both linear and non-linear patterns
in stock price fluctuations, it is anticipated that the
stacking model will outperform individual
techniques. Through the development of a more
precise and dependable instrument for forecasting
stock prices, this study possesses the capacity to
substantially advance the domain of financial
analytics and ultimately aid investors in the process
of making well-informed judgments.
4.3. Experimental Procedure
 ARIMA.
In this study, automatic ARIMA modeling was
carried out through a auto_arima function, which
autonomously searches for and chooses the most
suitable model parameters. The function performs
model comparisons based on the Akaike Information
Criterion (AIC).The AIC aims to select parameter
configurations that best fit the data while maintaining
model simplicity. As shown in Table1 after a series
of model configuration attempts, it was determined
that the ARIMA (1,1,1) model was the optimal one.
Table 1. Configurations of ARIMA.
Argument AIC
(1,1,1) 8414.499
The detailed statistical summary of the model is
shown in Table 2, from which we know that the
intercept term is insignificant, which means that the
model does not find a statistically significant non-
zero starting point or long-term trend in the data. In
addition, both parameters ar.L1 and ma. L1 are
significant. ar. The presence of L1 indicates a
statistically significant link between the current value
of the series and its prior value, ma.L1 is substantial,
indicating a statistically significant link between the
present value of the series and its stochastic shocks at
the prior time point. However, the Ljung-Box test of
the model indicates that the residuals do not exhibit
autocorrelation, implying that the model captures
most of the information in the time series.
IAMPA 2024 - International Conference on Innovations in Applied Mathematics, Physics and Astronomy
164

Figure 2. Result of autocorrelation (Picture credit: Original).
Figure 3. the predicted stock prices using the ARIMA (Picture credit: Original).
Table 2. The detailed statistical summary of ARIMA.
Metrics P-Value
intercept 0.118
ar.L1 0.014
ma.L1 0.004
sigma2 0.000
Ljung-Box(L1) 0.91
Jarque-Bera 0.00
Based on the results in table 2 and Figure 2 we
know that the residuals of the model do not exhibit
any significant autocorrelation. This outcome
suggests the model effectively captures the structure
of the data and that random fluctuations in the
residuals are unpredictable. However, the model does
not satisfy the assumption of residual normality and
the non-normality of the residuals may constrain the
model's forecasting ability, which implies that the
predictive outcomes of the ARIMA method are
unreliable, and we should be cautious about the
predictive results of this model.
The outcome of the prediction set subsequent to
training the ARIMA model with the training set is as
shown in figure 3.
Figure 3 shows the predicted stock prices using
the ARIMA model represented by the red curve, and
Stock Prediction Based on Traditional Statistical Models, Machine Learning Models and Fusion Models
165

the real stock prices by the curve line. The red
cureve's lack of fit to the true value is evidently
demonstrated to be a straight line; this indicates that
the model is an inadequate predictor of long-term
outcomes.
 GARCH:
This experiment constructs a number of GARCH
models by employing a two-layer loop that varies the
parameter combinations (p and q). The variables p
and q in the GARCH models denote the lag order,
which are used to determine the autoregressive and
moving average terms in the model, respectively. In
addition, we evaluated the performance of each
model using AIC and selected the model parameters
with the lowest AIC values.
After experimentation it was found that the lowest
value of model AIC was found at p=1, q=1, so we
selected this parameter for experimentation. To
reflect the accuracy of the GARCH model, we
compare the volatility of the original stock data with
the predicted volatility, and the r-squared of the final
model is 0.8756.
 LSTM:
In this study, we use two LSTMs and two Dropout
layers to construct the model. Both LSTM layers have
50 units and the parameter return_sequences is set to
True, which allows the LSTM to return so time-steps
of consecutive outputs; the Dropout layer is set to
20% to reduce overfitting; and finally the model
performs the prediction of the results through the
fully-connected layer containing 25 neurons and an
output layer.
The prediction set performs in figure 4.
Figure 4 depicts the date on the x-axis and the stock
price on the y-axis. The red line depicts the
predictions made by the LSTM model, and the actual
values are denoted by the blue line. The graph
indicates that while both lines follow a similar trend,
they rarely align perfectly. Typically, the predicted
values are slightly above the actual figures.
 Random Forest:
For this experiment, a total of 100 decision trees were
established, with all other parameters being left at
their default settings. Upon training the random forest
model with the training set, the prediction
performance of the prediction set is as shown in figure
5:
Figure 4. the predicted stock prices using the LSTM (Picture credit: Original).
Figure 5. the predicted stock prices using the RF (Picture credit: Original).
IAMPA 2024 - International Conference on Innovations in Applied Mathematics, Physics and Astronomy
166

Figure 6. The predicted stock prices using the Fusion Model (Picture credit: Original).
Figure 5 demonstrates that the majority of the
model's predictions largely coincide with the real
values, indicating the model's superior capability to
capture variations in stock prices. Nevertheless, at
certain periods, significant disparities arise between
the model's forecasts and the real values. These
discrepancies could be attributed to market
uncertainties or other factors that the model fails to
account for. Consequently, it indicates that the RF
model currently lacks the capability to accurately
predict abrupt fluctuations in stock prices.
 Fusion Model:
To assess feature importance, we utilize the Random
Forest method, incorporating the ten most crucial
features along with the GARCH model's predictions
into the previous LSTM model. The consequence
obtained from the fusion model are illustrated in
Figure 6.
 Model Comparisons:
Table 3 Results comparison.
MAE RMSE MAPE R^2
ARIMA 17.4000 19.6456 Nan -0.0231
GARCH 0.0021 0.0027 14.6341 0.8756
LSTM 0.0169 10.4901 5.8656 0.7093
RF 4.3642 6.9100 2.4215 0.8734
Fusion
Model
4.2696 5.1780 2.6105 0.9297
Table3 shows the fusion model surpasses the
other models in accurately predicting the closing
price of the stock. Its MAE, RMSE, MAPE are all the
smallest, and the r-square is the closest to 1, which
signifies that, relative to other models, fusion model
has the minimum prediction error, and the predicted
value is the closest to the original stock closing price,
which is the best fitting effect.
5. CONCLUSION
This study examines the effective integration of time
series and machine learning techniques in forecasting
stock prices. It involves the development and analysis
of ARIMA, GARCH, LSTM, Random Forest, and
integrated models. It is demonstrated that while
individual models possess their own strengths in
specific contexts, hybrid models exhibit superior
accuracy and reliability in stock price forecasting.
The fusion model put forth in this study not only
consolidates the benefits of each individual model but
also enhances prediction accuracy significantly
through feature engineering design and integrated
learning phases. Specifically, Random Forest does
exceptionally well in detecting important features,
which enables the LSTM model to effectively capture
extended relationships in time series data.
Experimental findings showcase the superiority of the
fusion model over single models both theoretically
and in practical applications. Not only does it enhance
forecasting accuracy, but it also bolsters the model's
adaptability to sudden market fluctuations. This
fusion approach equips stock market analysts and
investors with a more precise and dependable
decision-making tool, aiding them in making well-
informed investment choices within the intricate and
volatile financial markets.
It is important to acknowledge that while the
fusion model exhibits remarkable forecasting
capabilities for stock prices, its construction and
training necessitate substantial computational
resources and meticulous parameter tuning. Given the
volatile and intricate characteristics of the stock
Stock Prediction Based on Traditional Statistical Models, Machine Learning Models and Fusion Models
167

market, it is evident that predictive models are
incapable of entirely mirroring future pricing. This
underscores the importance for investors to employ
prudence when placing their trust in such models.
Future research endeavors could focus on delving
deeper into optimizing algorithm efficiency and
refining parameter optimization models to further
augment the precision and utility of stock price
forecasting.
REFERENCES
Lu W, Li J, Wang J, et al. A CNN-BiLSTM-AM method
for stock price prediction. Neural Computing
Application, 2021, 33(10): 4741-4753.
Li J, Pan S, Huang L. A machine learning based method for
customer behavior prediction. Tehnički vjesnik, 2019,
26(6): 1670-1676.
Sim H S, Kim H I, Ahn J J. Is deep learning for image
recognition applicable to stock market prediction?.
Complexity, 2019, 2019.
Ariyo A A, Adewumi A O, Ayo C K. Stock price prediction
using the ARIMA model, 16th International
Conference on Image Processing, Computer Vision and
Machine Learning. IEEE, 2014: 106-112.
Sunny M A I, Maswood M M S, Alharbi A G. Deep
learning-based stock price prediction using LSTM and
bi-directional LSTM model, 2020 2nd NILES. IEEE,
2020: 87-92.
Nelson D M Q, Pereira A C M, De Oliveira R A. Stock
market's price movement prediction with LSTM neural
networks,2017 Joint Conference on Neural Networks.
IEEE, 2017: 1419-1426.
Ma Y, Han R, Fu X. Stock prediction based on random
forest and LSTM neural network,2019 19th
International Conference on Control, Automation and
Systems. 2019: 126-130.
Ma Q. Comparison of ARIMA, ANN and LSTM for stock
price prediction, E3S WOC. EDP Sciences, 2020, 218:
01026.
Wang Y, Liu P, Zhu K, et al. A Garlic-Price-Prediction
Approach Based on Combined LSTM and GARCH-
Family Model. Applied Sciences, 2022, 12(22): 11366.
Jin, Z., Yang, Y. & Liu, Y. Stock closing price prediction
based on sentiment analysis and LSTM. Neural Comput
& Applic 32, 9713–9729 (2020).
Park H J, Kim Y, Kim H Y. Stock market forecasting using
a multi-task approach integrating long short-term
memory and the random forest framework. Applied
Soft Computing, 2022, 114: 108106.
Ma Y, Han R, Fu X. Stock prediction based on random
forest and LSTM neural network,2019 19th
International Conference on Control, Automation and
Systems. IEEE, 2019: 126-130.
Hochreiter, Sepp, and Jürgen Schmidhuber. "Long short-
term memory." Neural computation 9.8 (1997): 1735-
1780.
Genuer R, Poggi J M, Tuleau-Malot C. Variable selection
using random forests. Pattern recognition letters, 2010,
31(14): 2225-2236.
Breiman L. Random forests. Machine learning, 2001, 45: 5-
32.
Bollerslev T. Generalized autoregressive conditional
heteroskedasticity. Journal of econometrics, 1986,
31(3): 307-327.
Engle R F. Autoregressive conditional heteroscedasticity
with estimates of the variance of United Kingdom
inflation. Econometrica: JOTS, 1982: 987-1007.
IAMPA 2024 - International Conference on Innovations in Applied Mathematics, Physics and Astronomy
168