Journal of Robotics and Control (JRC)

Volume 5, Issue 5, 2024

ISSN: 2715-5072, DOI: 10.18196/jrc.v5i5.22460 1322

Stock Price Forecasting with Multivariate Time

Series Long Short-Term Memory: A Deep Learning
Approach
Furizal 1, Asdelina Ritonga 2, Alfian Ma’arif 3*, Iswanto Suwarno 4
1
Department of Informatics Engineering, Universitas Islam Riau, Pekanbaru, Indonesia
2
Master Program of Management, Universitas Islam Riau, Pekanbaru, Indonesia
3
Department of Electrical Engineering, Universitas Ahmad Dahlan, Yogyakarta, Indonesia
4
Department of Electrical Engineering, Universitas Muhammadiyah Yogyakarta, Yogyakarta, Indonesia
Email: 1 [email protected], 2 [email protected], 3 [email protected], 4 [email protected]
*Corresponding Author

Abstract—Stocks with their inherent complexity and capital appreciation or dividend payments, but they also carry
dynamic nature influenced by a multitude of external and risks associated with market volatility and company
internal factors, play a crucial role in investment analysis and performance [5], [6].
trend prediction. As financial instruments representing
ownership in a company, stocks not only reflect the company's In the dynamic and complex world of finance, predicting
performance but are also affected by external factors such as stock prices is one of the primary challenges for investors,
economic conditions, political climates, and social changes. In a analysts, and market participants [7], [8], [9], [10].
rapidly changing environment, investors and analysts Accurately anticipating stock price movements can provide a
continuously develop models and algorithms to aid in making
significant competitive advantage. However, the limitations
timely and effective investment decisions. This study applies a
Sequential model to predict stock data using a LSTM neural
of human ability to predict accurately and precisely cannot be
network. The model consists of a single hidden LSTM layer with denied, whether theoretically or based on experience. Manual
200 units. The LSTM layer, the core element of this model, prediction processes often require deep analysis of various
enables it to capture temporal patterns and long-term factors, such as market conditions, company performance,
relationships within the data. The training and testing data were economic trends, and other elements influencing stock prices
divided into 80% for training and 20% for testing. The Adam [11], [12]. In practice, however, humans have limitations in
optimizer was chosen to optimize the model's learning process, efficiently managing and analyzing large and complex
with a learning rate of 0.001. Dropout techniques were applied datasets, often leading to inaccurate predictions.
to reduce overfitting, with a dropout rate of 0.4, along with
Additionally, humans are susceptible to cognitive, emotional,
batch normalization and ReLU activation functions to enhance
model performance. Additionally, callback mechanisms,
and interpretive biases, which can impact the overall quality
including ReduceLROnPlateau and EarlyStopping, were used of their predictions. This can result in errors in risk
to optimize the training process and prevent overfitting. The assessment and investment opportunities, affecting the
model was evaluated using MAE and MSE metrics on training, overall performance of an investment portfolio [13], [14].
testing, and future prediction data. The results indicate that the
model achieved high accuracy, with an MAE of 0.0142 on the Furthermore, the process of technical prediction often
test data. However, future predictions showed higher MAE requires significant time and human effort, especially when
values, suggesting room for improvement in long-term dealing with complex data and large volumes [15]. This can
forecasting. The model's ability to accurately predict future hinder quick responses to rapid market changes, leading to
stock closing prices can assist investors in making informed missed valuable investment opportunities or suboptimal
investment decisions. decisions. In a competitive and fast-moving market
environment, the ability to respond swiftly and efficiently to
Keywords—LSTM; Time Series; Deep Learning; Stock Price; market changes is crucial for achieving optimal results. The
Forecasting.
limitations of human prediction capabilities can result in
I. INTRODUCTION significant consequences, including substantial financial
losses, market instability, and reduced overall investment
Stocks are financial instruments representing partial performance. Hence, there is a need to adopt more
ownership in a company [1], [2]. When an individual sophisticated and automated approaches in stock price
purchases a company's stock, they effectively buy a small prediction, such as using Deep Learning technology. One
portion of that company. In this context, shareholders have
algorithm that can be utilized in forecasting activities is Long
rights to a portion of the company's profits and the ability to
Short-Term Memory (LSTM) [16], [17], [18], [19].
participate in key company decisions through voting at
shareholder meetings. Stocks are often traded on stock LSTM enables stock price forecasting based on historical
exchanges, where their prices fluctuate based on market stock data, which is inherently time series data [20], [21],
supply and demand dynamics [3], [4]. As investment [22]. LSTM offers the ability to capture complex temporal
instruments, stocks offer the potential for profit through patterns in time series data [23], even when there are long-

Journal Web site: http://journal.umy.ac.id/index.php/jrc Journal Email: [email protected]

Journal of Robotics and Control (JRC) ISSN: 2715-5072 1323

term dependencies or complex non-linear structures [24]. By This research contributes to the development of stock
employing long-term memory units and gating mechanisms, price prediction methodologies by integrating multivariate
LSTM can overcome challenges often faced by traditional data into the LSTM model, aiming to build on the successes
approaches in predicting stock prices. of previous studies. Furthermore, this research seeks to
provide a better understanding of LSTM performance in
The Deep Learning approach, particularly LSTM, has
predicting stock prices in complex and fluctuating markets.
garnered significant attention in academic literature and the
The findings are expected to offer valuable insights for
financial industry as a potential method for improving stock
investors and market participants, aiding them in making
price prediction accuracy. Various studies have been
more informed and timely investment decisions.
conducted to test the effectiveness of LSTM in stock price
prediction, with several showing promising results [25], [26], II. METHOD
[27], [28], [29]. However, there is still room to enhance the
use of LSTM in stock price prediction, particularly in A. Research Stage
leveraging multivariate data and addressing market volatility Before delving into the detailed steps of the research
and uncertainty. This study aims to investigate the potential process, the initial stage involves loading the dataset that will
and performance of LSTM in this context, focusing on the be the subject of analysis. This dataset will be used for both
integration of multivariate data and the evaluation of LSTM training and testing the model. The dataset comprises 4916
model performance in predicting stock prices. rows and 7 columns, including trade date, opening price,
highest price, lowest price, closing price, adjusted closing
A review of the literature from previous studies highlights
price, and trading volume. Once the dataset is loaded, the next
the importance of using LSTM in stock price prediction. step is data preprocessing. This step ensures that there are no
Several studies have shown that LSTM can overcome the null values that could affect the model's training and testing
challenges faced by traditional approaches, such as long-term
process. Missing values can be handled in various ways, such
dependencies and non-linear structures in time series data.
as filling in missing values with the column's mean or using
For instance, M K Ho et al. demonstrated that LSTM could
techniques like interpolation. In this case, the issue is
not only predict future stock prices with minimal error but
resolved using SimpleImputer from the scikit-learn library
also accurately reflect the patterns and behaviors of closing [19], [34], [35]. Additionally, feature scaling is necessary to
prices on the Malaysian Stock Exchange during the testing ensure all features have similar value ranges. This scaling is
period. Their LSTM model for predicting stock prices on the
crucial as it helps machine learning algorithms converge
Malaysian Stock Exchange achieved an RMSE of 16.8410 on
more effectively. In this case, scaling is performed using min-
test data and an MAPE of 0.8184 [29]. In another study [30],
max scaling.
the performance of LSTM on the DJIA dataset yielded an
MSE of 0.0785, MAE of 0.2360, and RMSE of 0.2802. After data preprocessing is complete, the dataset is split
into two parts: training data (80%) and testing data (20%).
In 2019, Masud Rana et al. investigated the impact of The training data is used to train the model, while the testing
activation functions and optimization on stock price
data is used to evaluate the trained model's performance.
prediction using LSTM. Their experimental results showed
After splitting the data, the next step is designing the LSTM
that LSTM models using linear activation functions with
model. This model is built using a Sequential approach,
Adamax optimization and tanh activation functions with
where the layers in the model are arranged in a sequence. The
Adam optimization provided the best predictions, with model has several hyperparameters that need to be set,
RMSEs of 0.0151 each [31]. These favorable outcomes were including the number of epochs (the number of times the
supported by Mahla Nikou et al., who compared LSTM with
model will see the entire dataset during training), batch size
several machine learning algorithms in predicting the closing
(the number of samples used in one iteration), learning rate
prices of iShares MSCI United Kingdom from January 2015
for the Adam optimizer, and dropout rate (the percentage of
to June 2018. They found that the LSTM method neurons that will be randomly deactivated) to reduce
outperformed others, with a Mean Absolute Error (MAE) of overfitting.
0.210350 [32]. In the same year, another study by Chun Yuan
Lai et al. also utilized LSTM for stock price prediction, Moreover, the activation function used in the model is
showing accurate evaluations. The study demonstrated that determined, with the ReLU (Rectified Linear Activation)
LSTM could provide excellent results, with a Mean Squared function often chosen for its efficiency in training the model
Error (MSE) of around 1.9% [33]. [36], [37], [38]. Once the model is designed, the next step is
training it using the preprocessed training data. The training
Additionally, in the context of static and dynamic LSTM process can take time depending on the model's complexity
models, Duc Huu Dat Nguyen et al. demonstrated that the
and the dataset's size. After the model is trained, the next step
dynamic LSTM model enhances LSTM performance,
is to test its performance using the testing data. The results of
achieving an MAE of 0.0169 [26]. Another researcher, Jingyi
this testing will provide an indication of how well the model
Du et al., also studied stock price prediction using the
can predict data it has never seen before. Once the model is
American Apple’s stock data with LSTM. In this case, two tested, the next step is to evaluate its performance. In this
approaches were used: Univariate Feature Input and context, metrics such as MSE and MAE are used to measure
Multivariate Feature Input. These approaches resulted in two
the model's prediction accuracy. After the model is evaluated,
different error values. The Univariate feature input achieved
it can be used to forecast future values based on new or
an MAE of 0.155, while the Multivariate feature input
unseen data. These forecasting results can be visualized by
achieved an MAE of 0.033 [25].
plotting appropriate graphs to better understand the model's

Furizal, Stock Price Forecasting with Multivariate Time Series Long Short-Term Memory: A Deep Learning Approach
Journal of Robotics and Control (JRC) ISSN: 2715-5072 1324

trends and performance. The overall research stages are The input gate (it) is computed using the sigmoid function
illustrated in Fig. 1. (𝝈), which processes the combination of the previous hidden
state (ht-1) and the current input (xt), using the weight matrix
B. Multivariate LSTM
(Wi) connecting the input gate to the output gate, along with
Multivariate LSTM is a method in artificial intelligence the bias vector (bi) to control the flow of information. The
that extends the scope of LSTM to handle higher-dimensional value of this input gate determines how significant the new
data. LSTM is essentially an update to RNN, capable of information from the current step will be added to the LSTM
addressing the vanishing gradient problem that often occurs cell at that step. Furthermore, the value of 𝑪 ̌ 𝒕 , generated by
during backpropagation [39], [40]. On the other hand, RNN the tanh function, is computed to update the cell state. This
is a type of artificial neural network designed to process process involves the weight matrix (Wc) connecting the
sequential or time-series data. Like LSTM, RNN is also a part previous hidden state (ht-1) and the current input (xt) with the
of neural networks [41]. In LSTM, there are gating bias (bC). The value of 𝑪 ̌ 𝒕 determines what new information
mechanisms that control the memory recording process, as will be stored in the cell state at time step t, and it is regulated
well as the use of non-linear activation functions such as by the tanh activation function to ensure its value falls within
Hyperbolic Tangent (Tanh) and Sigmoid [42]. Within LSTM, the range between -1 and 1.
there are three types of gates that play a crucial role in
information management: the forget gate, the input gate, and Additionally, there is an output gate responsible for
the output gate [43], [44], [45], [46], [47]. generating the output of the LSTM at the current step. At this
stage, the sigmoid function is used to control how important
The forget gate plays an important role in determining the information is in the context of the hidden state. By
which information needs to be updated or deleted from the producing output between 0 and 1, the sigmoid function helps
LSTM cell memory [48]. This gate uses the sigmoid function determine how much new information should be included in
to generate a vector the same size as the number of cells in the hidden state for the next time step. Meanwhile, the tanh
the LSTM memory. This vector contains values between 0 function is used to normalize and regulate the value of the
and 1, representing the importance of each memory element new cell state. This helps generate divergent information,
stored in the LSTM cell. A value of 0 indicates that the allowing the network to update and add new information to
information should be forgotten, while a value of 1 indicates the cell state with a wide range of values between -1 and 1.
that the information should be retained. This process allows On the other hand, the cell state representing the long-term
the LSTM to ignore irrelevant or outdated information from memory of the LSTM cell is updated by considering the
previous steps, thereby focusing attention on more important output from the forget gate and the input gate [55], [56], [57],
information. The formula used for this gate is shown in [58]. The cell state is computed using Equation 4. The output
Equation (1) [28], [49], [50]. gate is then used to determine the value of the next hidden
𝑓𝑡 = 𝜎(𝑊𝑓 [ℎ𝑡−1 , 𝑥𝑡 ] + 𝑏𝑓 ) (1) state, which contains information from the previous input.
The formulas used in the output gate are shown in Equations
At each time step t in the LSTM network, the information (5) and (6).
processing begins by computing the value of the forget gate
(ft), which is the result of a sigmoid function processing the 𝐶𝑡 = 𝑓𝑡 ∗ 𝐶𝑡−1 + 𝑖𝑡 ∗ 𝐶̌𝑡 (4)
current input (xt) and the previous hidden state (ht-1). This Here, Ct represents the cell state information at time step
process involves the weight matrix (Wf) connecting the forget t, ft is the forget gate at time step t, it is the input gate at time
gate to the input gate, as well as the connection bias (bt) to step t, Ct-1 denotes the cell state at the previous time step, and
control the flow of information. The value of this forget gate
𝐶̌𝑡 is the value generated by the tanh function at time step t.
determines how significant the information from the previous
This process involves element-wise multiplication between
step will be either ignored or retained in the LSTM cell at the
the value of the forget gate and the cell state at the previous
current step.
time step, and the value of the input gate with the value
On the other hand, the input gate is responsible for generated by the tanh function. Thus, the cell state
updating the LSTM cell state with new information from the information at time step t is determined based on whether
current step [51], [52], [53]. This gate consists of two main information from the previous step is forgotten or retained,
parts: a sigmoid function that decides which parts of the input and the new information added from the current step through
will be updated, and a tanh function that generates a vector the tanh function.
containing values between -1 and 1, representing the new
𝑜𝑡 = 𝜎 (𝑊𝑜 [ℎ𝑡−1 , 𝑥𝑡 ] + 𝑏𝑜 ) (5)
information to be added to the LSTM cell memory. The
sigmoid function helps determine the importance of the new ℎ𝑡 = 𝑜𝑡 ∗ tanh(𝐶𝑡 ) (6)
information, while the tanh function adjusts these values to
Meanwhile, the value of the LSTM's output gate (ot) is
fit the desired range. The combination of these two functions
computed using the sigmoid function. This process involves
allows the LSTM to selectively update its cell memory with
the weight matrix (Wo) of the output gate, which connects the
relevant information from the current input. The formulas
previous hidden state (ht-1) and the current input (xt), along
used for the input gate are shown in Equations 2 and 3 [54].
with the bias vector (bo). The output gate determines how
𝑖𝑡 = 𝜎(𝑊𝑖 [ℎ𝑡−1 , 𝑥𝑡 ] + 𝑏𝑖 ) (2) significant the information from the LSTM cell will be
conveyed outward as output at that time step. Furthermore,
𝐶̌𝑡 = 𝑡𝑎ℎ𝑛(𝑊𝐶 [ℎ𝑡−1 , 𝑥𝑡 ] + 𝑏𝐶 ) (3) the hidden state (ht) at time step t is generated by multiplying
the value of the ot with the hyperbolic tangent (tanh) function

Furizal, Stock Price Forecasting with Multivariate Time Series Long Short-Term Memory: A Deep Learning Approach
Journal of Robotics and Control (JRC) ISSN: 2715-5072 1325

of the cell state (𝑪𝒕 ). This process ensures that the information of how close the model's predictions are to the actual
conveyed as output by the LSTM network at each time step observations [60], [61], [62]. Its mathematical expression is
aligns with the relevance and weighting provided by the given in Equation (7) [63], [64]. Meanwhile, MSE measures
output gate. The illustration of the LSTM architecture is the average of the squared differences between predicted and
shown in Fig. 2. actual values, offering an understanding of the overall
magnitude of prediction errors. Its mathematical formulation
C. Evaluation Method
is shown in Equation (8) [65].
In the realm of predictive modeling, it is crucial to have 𝑛
the right tools to evaluate model performance. MAE [59] and 1
MSE are two commonly used evaluation metrics in predictive 𝑀𝐴𝐸 = ∑ |𝑦𝑖 − 𝑦̂𝑖 | (7)
𝑛
modeling, including in the context of LSTM networks. These 𝑖=1
𝑛
metrics play a vital role in assessing the performance of 1
predictive models by measuring the difference between 𝑀𝑆𝐸 = ∑(𝑦𝑖 − 𝑦̂𝑖 )2 (8)
predicted values and actual values. 𝑛
𝑖=1

MAE represents the average of the absolute differences

between predicted and actual values, providing an indication

Fig. 1. Research Stage

Furizal, Stock Price Forecasting with Multivariate Time Series Long Short-Term Memory: A Deep Learning Approach
Journal of Robotics and Control (JRC) ISSN: 2715-5072 1326

Fig. 2. Illustration of LSTM Architecture

Here, 𝑦𝑖 represents the actual value, 𝑦̂𝑖 represents the The attributes in Table I provide information about BCA's
predicted value, and n indicates the total number of samples. daily trading activity and stock performance. The opening
In the context of LSTM networks, MAE and MSE serve as price, highest price, lowest price, and closing price offer
loss functions to measure the model's ability to recognize insights into price movements during the trading day, while
patterns in sequential data and make accurate predictions. the adjusted closing price accounts for corporate actions that
MAE is more sensitive to outliers because it uses absolute may affect the stock price. Additionally, the volume of stocks
differences [66], [67], while MSE [68] is more sensitive to traded indicates the level of market activity and investor
large errors because it uses squared differences. interest in the stock. The data graphs for each attribute are
shown in Fig. 3.
D. Dataset
The dataset used is the stock data of Bank Central Asia Fig. 3 shows that the data for the Open, Low, High, Close,
(BCA). BCA is one of the leading banks in Indonesia offering and Adj. Close attributes tend to increase each day, despite
various financial services including retail banking and experiencing many fluctuations. This trend indicates overall
corporate banking [69], [70], [71], [72]. Established in 1957, growth in the value of BCA shares, even though there is
BCA has grown to become one of the largest private banks in significant short-term volatility. When looking at the entire
the country, known for its extensive branch network, data set for each attribute, the fluctuations appear almost
innovative products, and strong customer base [73], [74], uniform, suggesting that despite daily ups and downs, the
[75]. With a focus on digital banking and customer general pattern tends to be consistent over a certain period.
convenience, BCA continues to adapt to changes in the On the other hand, the volume attribute data shows a
financial industry landscape, maintaining its position as a different pattern. The trading volume of these shares tends to
market leader. vary each year. This may be due to various external factors
The dataset used for this analysis pertains to the that influence trading interest and activity over certain time
performance of BCA stock (BBCA.JK) and was obtained periods. The highest trading volumes were observed between
from Yahoo Finance (link: 2004 and 2008, which might reflect a period of very high
https://finance.yahoo.com/quote/BBCA.JK/history). This market activity or significant economic events that affected
historical stock data covers the period from June 8, 2004, to trading volume. For a more detailed analysis of the data
April 19, 2024, and was downloaded on April 20, 2024. It is distribution, refer to the histogram in Fig. 4.
important to note that the data is only available on business Fig. 4 is a histogram that provides an overview of the data
days, excluding holidays and weekends. This dataset consists distribution for each attribute of BCA shares. In the
of 4916 rows and 7 columns, where each row represents a histograms for the Open, Low, High, Close, and Adj. Close
specific date and each column represents a different attribute data, it is evident that the majority of the data is concentrated
related to stock performance. The attributes included in this in the first interval, which has a relatively low price range.
dataset are shown in Table I. For example, the first interval in the Open data histogram,
with a range of 175.0 - 1197.5, contains the largest number
TABLE I. ATTRIBUTES IN THE USED DATASET
of data points, totaling 1558. A similar pattern is observed in
Attribute Description the High data histogram, where the first interval includes
Date Trading day date 1531 data points, decreasing in the second and third intervals,
Open Opening stock price and having only 75 data points in the last interval. The
High Highest stock price during the trading day
Low Lowest stock price during the trading day
number of data points gradually decreases in each subsequent
Close Closing stock price interval, indicating that most of the Open, Low, High, Close,
Adj. Adjusted closing price (adjusted for dividends and and Adj. Close prices tend to be in the lower price range.
Close stock splits)
Volume Volume of stocks traded during the trading day

Furizal, Stock Price Forecasting with Multivariate Time Series Long Short-Term Memory: A Deep Learning Approach
Journal of Robotics and Control (JRC) ISSN: 2715-5072 1327

Fig. 3. BCA stock dataset for the period from June 8, 2004, to April 18, 2024, based on Yahoo Finance

For the volume attribute, the histogram shows a highly average values, around the 3000s. All these stock price
concentrated distribution in the first interval, with a range of attributes have very low minimum values, approximately 175
0.0 - 194996000.0, containing the largest number of data to 177.5, and very high maximum values, exceeding 10000.
points, which is 4389. The number of data points drops This indicates the presence of significant outliers in the
significantly in the second interval with 372 data points and dataset. The interquartile range is relatively wide, with the
continues to decrease in each subsequent interval. This first quartile ranging from 544 to 730 and the third quartile
indicates that most of the trading volume is within the low between 4888 and 5505. Similarly, the Volume attribute also
range, with very high trading volumes being rare. This shows significant outliers, with daily transaction volumes
distribution suggests that intense trading activity occurs only varying extremely from 0 to nearly 2 billion. The high
in a limited number of instances, while most trading happens average volume, around 107 million, and the large standard
at lower volumes. The data characteristics are further deviation, approximately 130 million, indicate that some
illustrated in Fig. 5. trading days have much higher volumes than others. The
presence of these outliers broadens the data distribution and
The Open, High, Low, Close, and Adj Close attribute data
can affect the analysis and interpretation of overall stock
in Fig. 5 show a wide range of values with relatively high
prices.

Furizal, Stock Price Forecasting with Multivariate Time Series Long Short-Term Memory: A Deep Learning Approach
Journal of Robotics and Control (JRC) ISSN: 2715-5072 1328

Fig. 4. Histogram of each attribute in the dataset

III. RESULT AND DISCUSSION number of epochs aims to give the model the opportunity to
learn more complex and general patterns in the data.
In this analysis, a Sequential model was implemented to
Additionally, the Adam optimizer was chosen to optimize the
predict data using an artificial neural network with the LSTM
model's learning process. With a learning rate of 0.001, it
method. The built model consists of a single hidden LSTM
adjusts the step size taken by the optimizer in seeking the
layer with 200 units. The LSTM layer is the core element of
minimum of the loss function.
this model, enabling it to capture temporal patterns and long-
term relationships in the data. With 200 units, the model has To reduce the risk of overfitting, dropout techniques were
substantial capacity to handle data complexity, allowing for applied to the LSTM layer. Dropout is a regularization
the formation of more abstract and rich representations of the method that randomly disables a portion of the units in a layer
available information. during training, helping to prevent the model from becoming
overly reliant on certain subsets of features or patterns in the
The training and testing data were split with an 80% to
training data. In the model, the dropout rate was set to 0.4,
20% ratio, respectively. The model was trained for 200
meaning 40% of the units in the LSTM layer will be
epochs, which represents the number of iterations or learning
randomly disabled during the training process.
cycles through the entire training dataset. Using a sufficient

Furizal, Stock Price Forecasting with Multivariate Time Series Long Short-Term Memory: A Deep Learning Approach
Journal of Robotics and Control (JRC) ISSN: 2715-5072 1329

Fig. 5. Characteristics of each attribute in the dataset

Additionally, overfitting was also avoided by applying improvement in the monitored metric (in this case, validation
batch normalization and the ReLU activation function to the loss) after a certain number of epochs (in this case, one
LSTM layer. Overfitting is a phenomenon where a machine epoch). This helps ensure that the model can better find the
learning model "memorizes" the training data too well, optimal convergence point and reduces the likelihood of
resulting in poor generalization to new or unseen data. In this getting stuck in a local minimum. On the other hand, the
context, "memorizing" means that the model has become too EarlyStopping callback is used to stop the training process if
tailored to the training data, including noise or random errors, there is no improvement in the monitored metric (validation
making the model overly specific and less applicable to new loss) after a certain number of epochs (in this case, ten
data. Batch normalization helps accelerate convergence and epochs). In other words, if the model's performance does not
prevent drastic changes in the distribution of input values improve for ten consecutive epochs, the training will
between layers. The ReLU activation function introduces automatically stop. This helps prevent the training from
non-linearity into the model, enabling it to learn more continuing too long without significant performance
complex representations from the input data. Additionally, improvements.
model evaluation was conducted by monitoring the loss and
For validation data, the validation data parameter is used
validation loss metrics, which measure how well the model
to explicitly provide separate validation data from the
maps inputs to expected outputs.
training data. This data is prepared and not used during the
Details of the model parameters and their values are model training process. In this case, the validation data
shown in Table II. provided are X_test and y_test, which are separate from the
training data and are used to test the model's performance
TABLE II. PARAMETERS OF THE MODEL WITH THE VALUES after the training process is complete. The use of separate
Parameter Value validation data ensures that the model's performance
Model Sequential evaluation is conducted objectively and is not influenced by
Hidden layer 1 LSTM with 200 units the data used during training. The training results, including
Epoch 200 the designed data and model, show train loss and validation
Number of feature 6
loss as depicted in Fig. 6.
Optimizer Adam
Learning rate 0.001 Fig. 6 provides information that model training only
Dropout 0.4
lasted until the 19th epoch out of the initially set 200 epochs.
Batch size 8
Activation ReLu This was due to the occurrence of overfitting in the model. In
Metriks Loss, val_loss the graph, it can be seen that the train loss, which represents
Output layer 1 the error rate on the training data, tends to stabilize and even
decrease as the epochs progress. However, in the validation
loss, which depicts the error rate on the validation data not
In addition to the parameters in Table II, a series of
used in training, there are fluctuations in the initial epochs,
callbacks were implemented in this model to optimize the
particularly from epoch 2 to 3, where the value increases from
training process and avoid overfitting. The callbacks used are
0.0007 to 0.00088. Similar fluctuations are observed from
ReduceLROnPlateau and EarlyStopping. ReduceLROn-
epoch 6 to 7, where the value increases from 0.00051 to
Plateau automatically reduces the learning rate if there is no
0.00052.

Furizal, Stock Price Forecasting with Multivariate Time Series Long Short-Term Memory: A Deep Learning Approach
Journal of Robotics and Control (JRC) ISSN: 2715-5072 1330

the Close feature has a contribution value of -0.0134. On the

other hand, Volume is the feature with the lowest contribution
value, which is at -0.0194. The contributions of these features
will impact the desired forecasting results.
In this case, an LSTM model is used to predict the closing
price of stocks. Selecting the closing price of stocks as the
prediction target has a strong basis in financial market
analysis. The closing price reflects the last price of a stock at
the end of the trading period, and therefore, is considered a
crucial parameter in evaluating stock performance and
making investment decisions. Investors and analysts often
use the closing price as a basis for evaluating their portfolio
performance, predicting market trends, and developing
investment strategies. By predicting the closing price,
Fig. 6. Train Loss and Validation Loss during model training investors and analysts can make more timely and effective
investment decisions. Accurate predictions of closing prices
These fluctuations in the validation loss indicate the can provide signals for when to buy or sell stocks. This helps
presence of overfitting, where the model fails to generalize investors maximize profit potential and reduce the risk of
well on the validation data. Consequently, even though the losses in their stock trading. Additionally, the closing price is
train loss continues to decrease, the validation loss fluctuates also used as a basis for calculating technical indicators and
or even rises, leading to the cessation of model training at the chart analysis that are important in financial market analysis.
19th epoch due to overfitting. The selection of the 19th epoch Indicators such as moving averages, relative strength index
as the training endpoint by the EarlyStopping callback was (RSI), and Bollinger Bands are often calculated based on
based on the model's performance evaluation using the closing prices.
validation loss. Although there was potential to further reduce
The results of predicting stock closing prices generated by
the train loss, the training was stopped to prevent overfitting
the model include three main aspects: predictions on training
that could affect the model's performance on new data.
data, predictions on testing data, and future forecasting.
The model used six features (Open, High, Low, Close, Predictions on training data are used to evaluate the model's
Adj. Close, and Volume). These features served as input to performance during the training process, while predictions on
produce the corresponding output. However, it is necessary testing data are important to evaluate the model's ability to
to assess the extent of each feature's contribution within the generalize patterns from unseen data. On the other hand,
trained model. This step evaluates the relative contribution of future forecasting is conducted to estimate stock closing
each feature in the LSTM model. The weights from the prices for future periods not included in the training or testing
LSTM layer were accessed using the get_weights() method data. Future forecasting can provide insights for investors and
on the first LSTM layer in the model. These weights represent analysts to plan long-term investment strategies and
the relationship between the input (features) and the output anticipate upcoming market changes.
(predictions) in the model. After obtaining the weights, the
In this case, future forecasting attempts to predict the
contribution of each feature was calculated by averaging the
closing stock prices for the next 15 days after the last date in
values of each row of weights. The contribution level of each
the testing data. In this instance, the testing data ends on April
feature is shown in Fig. 7.
19, 2024, so future forecasting begins from April 20, 2024, to
May 4, 2024 (15 days). Predictions on training data, testing
data, and future forecasting are shown in Fig. 8. However, the
future forecasting results in Fig. 8 are not very clear, so in
Fig. 9, a smaller dataset is plotted, specifically for the year
2024 only.
Fig. 9 presents information about the model's ability to
predict future stock prices by comparing the prediction
results with actual data. In this visualization, besides the
training and testing data, the future forecasting results are also
directly compared with the actual data for the same period.
By juxtaposing the prediction data and the actual data,
differences between them can be observed more clearly.
These plots illustrate how the model's predictions tend to
follow the fluctuations in the actual data. This indicates that
Fig. 7. Feature contribution in trained LSTM model based on the average the model can capture the general trends and movements of
get_weight
actual closing stock prices, although there are some
differences that occur. Fluctuations in actual data are often
Fig. 7 shows that each feature contributes differently. In
caused by dynamic market factors that cannot always be
this context, the Adj. Close feature provides the largest
predicted with perfect accuracy.
contribution compared to the other 5 features. Following that,

Furizal, Stock Price Forecasting with Multivariate Time Series Long Short-Term Memory: A Deep Learning Approach
Journal of Robotics and Control (JRC) ISSN: 2715-5072 1331

Fig. 8. Comparison of predicted closing stock prices on training data, testing data, and future forecasting against actual data

Fig. 9. Comparison of predicted closing stock prices on training data, testing data, and future forecasting against actual data (year 2024)

Fluctuations in prediction data tend to be smoother data. On the other hand, MSE measures the average of the
compared to actual data. This suggests that although the squared errors between predictions and actual values [77],
model can follow the main trends, its predictions do not fully [78]. By squaring the errors, MSE gives more weight to larger
capture the volatility present in the actual data. The errors, making it more sensitive to outliers. Similar to MAE,
smoothing effect on prediction data can be interpreted as an MSE also indicates that the smaller its value, the better the
effect of model regularization or the use of parameters such model's performance [79].
as dropout that reduce overfitting, resulting in more stable
In this study, all three types of prediction data are
predictions but less sensitivity to drastic changes in actual
evaluated and compared based on the MAE and MSE values
data.
obtained. This evaluation aims to understand how the model
To assess the performance of the model's predictions on performs not only on data seen during training but also on
training data, testing data, and future forecasting, evaluation unseen testing data and future forecasting predictions. The
metrices are needed to measure the accuracy and reliability results of the evaluation and comparison of MAE and MSE
of the model. In this study, MAE and MSE are used as the for the three types of prediction data are shown in Fig. 10.
primary evaluation metrices [76]. Both of these metrics are
Fig. 10 shows the comparison of MAE and MSE values
chosen because they provide a comprehensive view of
for predictions on training data, testing data, and future
prediction errors, each from a different perspective.
forecasting. However, it is important to note that the number
MAE measures the average absolute error between of data points in each prediction category differs. The training
predictions and actual values. It provides a direct insight into data consists of 80% of the total 4916 data, totaling 3933
the magnitude of the average error without considering the rows. The testing data comprises 20% of the total data, which
direction of the error. The smaller the MAE value, the more is 983 rows. Meanwhile, the data for future forecasting only
accurate the model is in predicting values close to the actual consists of 15 rows. This difference in the number of data

Furizal, Stock Price Forecasting with Multivariate Time Series Long Short-Term Memory: A Deep Learning Approach
Journal of Robotics and Control (JRC) ISSN: 2715-5072 1332

points can affect the accuracy and stability of the model's 0.0955, indicating that predictions for future periods are more
predictions. prone to errors. Nevertheless, these results can still be
categorized as quite good. To see the performance
The lowest MAE value is obtained from predictions on
differences, a comparison with several previous studies is
the training data, with an average of 0.0074. This indicates
shown in Fig. 11.
that the model is very accurate in predicting values close to
the actual data during training. The MAE value for the testing Fig. 11 shows that LSTM exhibits variations in prediction
data is 0.0142, while for future forecasting, it is 0.2424. The accuracy, measured by MAE values. The highest MAE value
higher MAE value for future forecasting indicates that the of 0.2360 was recorded in the study by J. Qiu, B. Wang, and
predictions for future periods are less accurate compared to C. Zhou, indicating a relatively large prediction error
the training and testing data. compared to other studies. In contrast, the studies by M.
Nikou, G. Mansourfar, and J. Bagherzadeh (0.210350)
For MSE, the best value is obtained when predicting the
showed better result with more accurate predictions. The
testing data, with an average of 0.0005. This is slightly better
studies by D. H. D. Nguyen, L. P. Tran, and V. Nguyen
than the average MSE on predictions for training data,
(0.0169) and the second model by J. Du, Q. Liu, K. Chen, and
indicating that the model is able to maintain stability and
J. Wang (0.033) also demonstrated very high accuracy.
accuracy in minimizing large errors during testing. However,
future forecasting produces a higher MSE value, which is

Fig. 10. Average MAE and MSE for each prediction

Fig. 11. Performance comparison with previous studies

Furizal, Stock Price Forecasting with Multivariate Time Series Long Short-Term Memory: A Deep Learning Approach
Journal of Robotics and Control (JRC) ISSN: 2715-5072 1333

Meanwhile, the study in this research recorded the lowest and Management, vol. 7, no. 2, pp. 81–95, 2020, doi:
10.18488/journal.62.2020.72.81.95.
MAE value of 0.0142 on the testing data, indicating the
highest prediction accuracy among all compared studies. [6] T. H. Nguyen, H. A. Nguyen, Q. C. Tran, and Q. L. Le, “Dividend
policy and share price volatility: empirical evidence from Vietnam,”
However, this research still needs improvements to provide Accounting, pp. 67–78, 2020, doi: 10.5267/j.ac.2019.12.006.
more accurate forecasting results, enabling investors to have [7] K. Mishev, A. Gjorgjevikj, I. Vodenska, L. T. Chitkushev, and D.
greater confidence in the model for future stock price Trajanov, “Evaluation of Sentiment Analysis in Finance: From
predictions. Therefore, it is recommended for future research Lexicons to Transformers,” IEEE Access, vol. 8, pp. 131662–131682,
to make several enhancements to improve the accuracy of the 2020, doi: 10.1109/ACCESS.2020.3009626.
LSTM model in stock price forecasting. Handling outliers is [8] A. Thakkar and K. Chaudhari, “A Comprehensive Survey on Portfolio
Optimization, Stock Price and Trend Prediction Using Particle Swarm
crucial, with more effective detection and removal methods, Optimization,” Archives of Computational Methods in Engineering,
as well as appropriate data transformations to mitigate the vol. 28, no. 4, pp. 2133–2164, Jun. 2021, doi: 10.1007/s11831-020-
negative impact of outliers on the model. Additionally, to 09448-8.
address overfitting issues, the use of regularization [9] W. S. Udo, N. A. Ochuba, O. Akinrinola, and Y. J. Ololade,
techniques, dropout, and cross-validation is necessary to “Theoretical approaches to data analytics and decision-making in
finance: Insights from Africa and the United States,” GSC Advanced
ensure the model is not overly complex and generalizable to Research and Reviews, vol. 18, no. 3, pp. 343–349, Mar. 2024, doi:
different datasets. Furthermore, improving the model 10.30574/gscarr.2024.18.3.0114.
architecture through more extensive hyperparameter tuning, [10] A. Thakkar and K. Chaudhari, “Fusion in stock market prediction: A
the use of Bidirectional LSTM [44], [80], [81], and Stacked decade survey on the necessity, recent developments, and potential
LSTM [82] can help capture more complex data features and future directions,” Information Fusion, vol. 65, pp. 95–107, Jan. 2021,
doi: 10.1016/j.inffus.2020.08.019.
enhance prediction accuracy. Moreover, data augmentation
[11] A. Thakkar and K. Chaudhari, “A comprehensive survey on deep
and the integration of relevant external features, such as neural networks for stock market: The need, challenges, and future
market sentiment and macroeconomic indicators, will directions,” Expert Syst Appl, vol. 177, p. 114800, Sep. 2021, doi:
provide additional context that can improve prediction 10.1016/j.eswa.2021.114800.
outcomes. Implementing monitoring systems and feedback [12] M. Nabipour, P. Nayyeri, H. Jabani, A. Mosavi, E. Salwana, and S. S.,
loops is also essential to ensure the model remains optimal “Deep Learning for Stock Market Prediction,” Entropy, vol. 22, no. 8,
p. 840, Jul. 2020, doi: 10.3390/e22080840.
and can adapt to data changes over time.
[13] A. Richards, C. C. French, A. J. Calder, B. Webb, R. Fox, and A. W.
IV. CONCLUSSION Young, “Anxiety-related bias in the classification of emotionally
ambiguous facial expressions.,” Emotion, vol. 2, no. 3, pp. 273–287,
This study successfully demonstrates the superior 2002, doi: 10.1037/1528-3542.2.3.273.
performance of the Sequential LSTM model in predicting [14] I. Blanchette and A. Richards, “The influence of affect on higher level
closing stock prices. The model, consisting of a single hidden cognition: A review of research on interpretation, judgement, decision
making and reasoning,” Cogn Emot, vol. 24, no. 4, pp. 561–595, Jun.
LSTM layer with 200 units, shows strong capability in 2010, doi: 10.1080/02699930903132496.
capturing temporal patterns and long-term relationships in the [15] G. Kumar, S. Jain, and U. P. Singh, “Stock Market Forecasting Using
data. Evaluation results indicate the lowest MAE value of Computational Intelligence: A Survey,” Archives of Computational
0.0142 on testing data, signifying high accuracy in stock price Methods in Engineering, vol. 28, no. 3, pp. 1069–1101, May 2021, doi:
predictions. However, the future forecasting results exhibit a 10.1007/s11831-020-09413-5.
higher MAE value, indicating a need for model improvement [16] X. Song et al., “Time-series well performance prediction based on
Long Short-Term Memory (LSTM) neural network model,” J Pet Sci
for long-term predictions. This model outperforms several Eng, vol. 186, p. 106682, Mar. 2020, doi:
previous studies referenced. Nonetheless, there is still room 10.1016/j.petrol.2019.106682.
for improvement to achieve even better results. [17] Y. Ji, A. W.-C. Liew, and L. Yang, “A Novel Improved Particle Swarm
Optimization With Long-Short Term Memory Hybrid Model for Stock
REFERENCES Indices Forecast,” IEEE Access, vol. 9, pp. 23660–23671, 2021, doi:
[1] D. Mariyani, H. Hariyanti, and D. R. Novida, “The Effect Of Economic 10.1109/ACCESS.2021.3056713.
Value Added (Eva), Market Value Added (MVA), Refined Economic [18] G. Van Houdt, C. Mosquera, and G. Nápoles, “A review on the long
Value Added (REVA) On Stock Prices And Stock Returns In short-term memory model,” Artif Intell Rev, vol. 53, no. 8, pp. 5929–
Manufacturing Companies Listed In Indonesia Stock Exchange,” 5955, Dec. 2020, doi: 10.1007/s10462-020-09838-1.
Winter Journal: Imwi Student Research Journal, vol. 3, no. 1, pp. 10– [19] R. Huang et al., “Well performance prediction based on Long Short-
22, Oct. 2023, doi: 10.52851/wt.v3i1.46. Term Memory (LSTM) neural network,” J Pet Sci Eng, vol. 208, p.
[2] A. Kartono, V. W. Fatmawati, S. T. Wahyudi, and Irmansyah, 109686, Jan. 2022, doi: 10.1016/j.petrol.2021.109686.
“Numerical Solution of Nonlinear Schrodinger Approaches Using the [20] G. Ding and L. Qin, “Study on the prediction of stock price based on
Fourth-Order Runge-Kutta Method for Predicting Stock Pricing,” J the associated network model of LSTM,” International Journal of
Phys Conf Ser, vol. 1491, no. 1, p. 012021, Mar. 2020, doi: Machine Learning and Cybernetics, vol. 11, no. 6, pp. 1307–1317, Jun.
10.1088/1742-6596/1491/1/012021. 2020, doi: 10.1007/s13042-019-01041-1.
[3] R. Wu, Z. Qin, and B.-Y. Liu, “A systemic analysis of dynamic [21] A. Kothari, A. Kulkarni, T. Kohade, and C. Pawar, “Stock Market
frequency spillovers among carbon emissions trading (CET), fossil Prediction Using LSTM,” in Smart Trends in Computing and
energy and sectoral stock markets: Evidence from China,” Energy, vol. Communications, pp. 143–164, 2024, doi: 10.1007/978-981-97-1326-
254, p. 124176, Sep. 2022, doi: 10.1016/j.energy.2022.124176. 4_13.
[4] B. Dhingra, S. Batra, V. Aggarwal, M. Yadav, and P. Kumar, “Stock [22] R. Kumar, P. Kumar, and Y. Kumar, “Analysis of Financial Time
market volatility: a systematic review,” Journal of Modelling in Series Forecasting using Deep Learning Model,” in 2021 11th
Management, vol. 19, no. 3, pp. 925–952, Mar. 2024, doi: International Conference on Cloud Computing, Data Science &
10.1108/JM2-04-2023-0080. Engineering (Confluence), pp. 877–881, 2021, doi:
[5] A. Firmansyah, W. Utami, H. Umar, and S. D. Mulyani, “Do 10.1109/Confluence51648.2021.9377158.
Derivative Instruments Increase Firm Risk for Indonesia Non- [23] A. Kurniawati, E. M. Yuniarno, and Y. K. Suprapto, “Deep Learning
Financial Companies?,” International Journal of Business, Economics for Multi-Structured Javanese Gamelan Note Generator,” Knowledge

Furizal, Stock Price Forecasting with Multivariate Time Series Long Short-Term Memory: A Deep Learning Approach
Journal of Robotics and Control (JRC) ISSN: 2715-5072 1334

Engineering and Data Science, vol. 6, no. 1, p. 41, Jul. 2023, doi: Normalization Techniques for Energy demands Time Series
10.17977/um018v6i12023p41-56. Forecasting,” Knowledge Engineering and Data Science, vol. 7, no. 1,
[24] A. K. S. Lenson and G. Airlangga, “Comparative Analysis of MLP, pp. 1–12, 2024.
CNN, and RNN Models in Automatic Speech Recognition: Dissecting [41] F. F. Rahani and P. A. Rosyady, “Quadrotor Altitude Control using
Performance Metric,” Buletin Ilmiah Sarjana Teknik Elektro, vol. 5, Recurrent Neural Network PID,” Buletin Ilmiah Sarjana Teknik
no. 4, pp. 576–583, 2023, doi: 10.12928/biste.v5i4.9668. Elektro, vol. 5, no. 2, pp. 279–290, 2023, doi:
[25] J. Du, Q. Liu, K. Chen, and J. Wang, “Forecasting stock prices in two 10.12928/biste.v5i2.8455.
ways based on LSTM neural network,” in 2019 IEEE 3rd Information [42] M. F. Maulana, S. Sa’adah, and P. Eko Yunanto, “Crude Oil Price
Technology, Networking, Electronic and Automation Control Forecasting Using Long Short-Term Memory,” Jurnal Ilmiah Teknik
Conference (ITNEC), pp. 1083–1086, 2019, doi: Elektro Komputer dan Informatika (JITEKI), vol. 7, no. 2, p. 286, 2021,
10.1109/ITNEC.2019.8729026. doi: 10.26555/jiteki.v7i2.21086.
[26] D. H. D. Nguyen, L. P. Tran, and V. Nguyen, “Predicting Stock Prices [43] Z. Chang, Y. Zhang, and W. Chen, “Effective Adam-Optimized LSTM
Using Dynamic LSTM Models,” in Communications in Computer and Neural Network for Electricity Price Forecasting,” in 2018 IEEE 9th
Information Science (CCIS), pp. 199–212, 2019, doi: 10.1007/978-3- International Conference on Software Engineering and Service
030-32475-9_15. Science (ICSESS), pp. 245–248, 2018, doi:
[27] S. Kumar S, C. D, and S. Rajan, “Stock price prediction using deep 10.1109/ICSESS.2018.8663710.
learning LSTM (long short-term memory),” in 2022 2nd International [44] K. Yudhaprawira Halim, D. Turianto Nugrahadi, M. Reza Faisal, R.
Conference on Advance Computing and Innovative Technologies in Herteno, and I. Budiman, “Gender Classification Based on
Engineering (ICACITE), pp. 1787–1791, 2022, doi: Electrocardiogram Signals Using Long Short Term Memory and
10.1109/ICACITE53722.2022.9823639. Bidirectional Long Short Term Memory,” Jurnal Ilmiah Teknik Elektro
[28] P. Hoang Vuong, T. Tan Dat, T. Khoi Mai, P. Hoang Uyen, and P. The Komputer dan Informatika (JITEKI), vol. 9, no. 3, pp. 606–618, 2023,
Bao, “Stock-Price Forecasting Based on XGBoost and LSTM,” doi: 10.26555/jiteki.v9i3.26354.
Computer Systems Science and Engineering, vol. 40, no. 1, pp. 237– [45] W. T. Handoko and A. N. Handayani, “Forecasting Solar Irradiation
246, 2022, doi: 10.32604/csse.2022.017685. on Solar Tubes Using the LSTM Method and Exponential Smoothing,”
[29] M. K. Ho, H. Darman, and S. Musa, “Stock Price Prediction Using Jurnal Ilmiah Teknik Elektro Komputer dan Informatika (JITEKI), vol.
ARIMA, Neural Network and LSTM Models,” J Phys Conf Ser, vol. 9, no. 3, pp. 649–660, 2023, doi: 10.26555/jiteki.v9i3.26395.
1988, no. 1, p. 012041, Jul. 2021, doi: 10.1088/1742- [46] N. M. Azahra and E. B. Setiawan, “Sentence-Level Granularity
6596/1988/1/012041. Oriented Sentiment Analysis of Social Media Using Long Short-Term
[30] J. Qiu, B. Wang, and C. Zhou, “Forecasting stock prices with long- Memory (LSTM) and IndoBERTweet Method,” Jurnal Ilmiah Teknik
short term memory neural network based on attention mechanism,” Elektro Komputer dan Informatika (JITEKI), vol. 9, no. 1, pp. 85–95,
PLoS One, vol. 15, no. 1, p. e0227222, Jan. 2020, doi: 2023, doi: 10.26555/jiteki.v9i1.25765.
10.1371/journal.pone.0227222. [47] T. A. Armanda, I. P. Wardhani, T. M. Akhriza, and T. M. A. Admira,
[31] M. Rana, Md. M. Uddin, and Md. M. Hoque, “Effects of Activation “Recurrent Session Approach to Generative Association Rule based
Functions and Optimizers on Stock Price Prediction using LSTM Recommendation,” Knowledge Engineering and Data Science, vol. 6,
Recurrent Networks,” in Proceedings of the 2019 3rd International no. 2, p. 199, Nov. 2023, doi: 10.17977/um018v6i22023p199-214.
Conference on Computer Science and Artificial Intelligence, pp. 354– [48] M. Hayaty and A. H. Pratama, “Performance of Lexical Resource and
358, 2019, doi: 10.1145/3374587.3374622. Manual Labeling on Long Short-Term Memory Model for Text
[32] M. Nikou, G. Mansourfar, and J. Bagherzadeh, “Stock price prediction Classification,” Jurnal Ilmiah Teknik Elektro Komputer dan
using Deep learning algorithm and its comparison with machine Informatika (JITEKI), vol. 9, no. 1, pp. 74–84, 2023, doi:
learning algorithms,” Intelligent Systems in Accounting, Finance and 10.26555/jiteki.v9i1.25375.
Management, vol. 26, no. 4, pp. 164–174, Oct. 2019, doi: [49] G. Sonkavde, D. S. Dharrao, A. M. Bongale, S. T. Deokate, D.
10.1002/isaf.1459. Doreswamy, and S. K. Bhat, “Forecasting Stock Market Prices Using
[33] C. Y. Lai, R.-C. Chen, and R. E. Caraka, “Prediction Stock Price Based Machine Learning and Deep Learning Models: A Systematic Review,
on Different Index Factors Using LSTM,” in 2019 International Performance Analysis and Discussion of Implications,” International
Conference on Machine Learning and Cybernetics (ICMLC), pp. 1–6, Journal of Financial Studies, vol. 11, no. 3, p. 94, Jul. 2023, doi:
2019, doi: 10.1109/ICMLC48188.2019.8949162. 10.3390/ijfs11030094.
[34] T. Boechel, L. M. Policarpo, G. de O. Ramos, R. da Rosa Righi, and [50] T. B. Shahi, A. Shrestha, A. Neupane, and W. Guo, “Stock Price
D. Singh, “Prediction of Harvest Time of Apple Trees: An RNN-Based Forecasting with Deep Learning: A Comparative Study,” Mathematics,
Approach,” Algorithms, vol. 15, no. 3, p. 95, Mar. 2022, doi: vol. 8, no. 9, p. 1441, Aug. 2020, doi: 10.3390/math8091441.
10.3390/a15030095. [51] S. Mirzaei, J.-L. Kang, and K.-Y. Chu, “A comparative study on long
[35] A. Testas, “Decision Tree Regression with Pandas, Scikit-Learn, and short-term memory and gated recurrent unit neural networks in fault
PySpark,” in Distributed Machine Learning with PySpark, pp. 75–113, diagnosis for chemical processes using visualization,” J Taiwan Inst
2023, doi: 10.1007/978-1-4842-9751-3_4. Chem Eng, vol. 130, p. 104028, Jan. 2022, doi:
10.1016/j.jtice.2021.08.016.
[36] Y. Yu, K. Adu, N. Tashi, P. Anokye, X. Wang, and M. A. Ayidzoe,
“RMAF: Relu-Memristor-Like Activation Function for Deep [52] K. Khalil, B. Dey, A. Kumar, and M. Bayoumi, “A Reversible-Logic
Learning,” IEEE Access, vol. 8, pp. 72727–72741, 2020, doi: Based Architecture for Long Short-Term Memory (LSTM) Network,”
10.1109/ACCESS.2020.2987829. in 2021 IEEE International Symposium on Circuits and Systems
(ISCAS), pp. 1–5, 2021, doi: 10.1109/ISCAS51556.2021.9401395.
[37] S. R. Dubey, S. K. Singh, and B. B. Chaudhuri, “Activation functions
in deep learning: A comprehensive survey and benchmark,” [53] M. Moradi A., S. A. Sadrossadat, and V. Derhami, “Long Short-Term
Neurocomputing, vol. 503, pp. 92–108, Sep. 2022, doi: Memory Neural Networks for Modeling Nonlinear Electronic
10.1016/j.neucom.2022.06.111. Components,” IEEE Trans Compon Packaging Manuf Technol, vol.
11, no. 5, pp. 840–847, May 2021, doi:
[38] S.-L. Shen, N. Zhang, A. Zhou, and Z.-Y. Yin, “Enhancement of neural
10.1109/TCPMT.2021.3071351.
networks with an alternative activation function tanhLU,” Expert Syst
Appl, vol. 199, p. 117181, Aug. 2022, doi: [54] W. K. Sari, D. P. Rini, and R. F. Malik, “Text Classification Using
10.1016/j.eswa.2022.117181. Long Short-Term Memory With GloVe Features,” Jurnal Ilmiah
Teknik Elektro Komputer dan Informatika (JITEKI), vol. 5, no. 2, p.
[39] D. Pebrianti, H. Kurniawan, L. Bayuaji, and R. Rusdah, “XgBoost
85, 2020, doi: 10.26555/jiteki.v5i2.15021.
Hyper-Parameter Tuning Using Particle Swarm Optimization for Stock
Price Forecasting,” Jurnal Ilmiah Teknik Elektro Komputer dan [55] A. Pranolo, Y. Mao, A. P. Wibawa, A. B. P. Utama, and F. A.
Informatika (JITEKI), vol. 9, no. 4, pp. 1179–1195, 2024, doi: Dwiyanto, “Optimized Three Deep Learning Models Based-PSO
10.26555/jiteki.v9i4.27712. Hyperparameters for Beijing PM2.5 Prediction,” Knowledge
Engineering and Data Science, vol. 5, no. 1, p. 53, Jun. 2022, doi:
[40] A. Pranolo, X. Zhou, Y. Mao, and B. Widi, “Exploring LSTM-based
10.17977/um018v5i12022p53-66.
Attention Mechanisms with PSO and Grid Search under Different

Furizal, Stock Price Forecasting with Multivariate Time Series Long Short-Term Memory: A Deep Learning Approach
Journal of Robotics and Control (JRC) ISSN: 2715-5072 1335

[56] G. Van Houdt, C. Mosquera, and G. Nápoles, “A review on the long highlight wide-ranging digitization opportunities and major
short-term memory model,” Artif Intell Rev, vol. 53, no. 8, pp. 5929– challenges,” SSRN Electronic Journal, 2020, doi:
5955, Dec. 2020, doi: 10.1007/s10462-020-09838-1. 10.2139/ssrn.3529924.
[57] J. Duan, P.-F. Zhang, R. Qiu, and Z. Huang, “Long short-term [71] N. P. W. Rahayu, S. Bangsawan, Mahrinasari, and A. Ahadiat, “Service
enhanced memory for sequential recommendation,” World Wide Web, Quality, Employee Ethics, Bank Customer Trust: The Role of Size
vol. 26, no. 2, pp. 561–583, Mar. 2023, doi: 10.1007/s11280-022- Moderation and Bank Reputation in Indonesia,” International Journal
01056-9. of Pharmaceutical Research, vol. 13, no. 2, Jul. 2021, doi:
[58] G. Airlangga, “Performance Evaluation of Deep Learning Techniques 10.31838/ijpr/2021.13.02.522.
in Gesture Recognition Systems,” Buletin Ilmiah Sarjana Teknik [72] M. Kautsar, E. E. Merawaty, S. Bahri, and A. H. Sutawidjaya,
Elektro, vol. 6, no. 1, pp. 83–90, 2024, doi: 10.12928/biste.v6i1.10120. “Corporate Strategy Analysis in Increasing the Value of the Firm
[59] I. H. Arsytania, E. B. Setiawan, and I. Kurniawan, “Movie through Mergers and Acquisitions in the Banking Sector (Case Study
Recommender System with Cascade Hybrid Filtering Using on the BCA Group),” International Journal of Economics, Business
Convolutional Neural Network,” Jurnal Ilmiah Teknik Elektro and Management Research, vol. 7, no. 1, pp. 01–09, 2023, doi:
Komputer dan Informatika (JITEKI), vol. 9, no. 4, pp. 1262–1274, 10.51505/IJEBMR.2023.7101.
2024, doi: 10.26555/jiteki.v9i4.28146. [73] A. E. Parung, L. S. Oppusunggu, and T. Sunaryo, “The Role of Risk
[60] K. A. Althelaya, E.-S. M. El-Alfy, and S. Mohammed, “Evaluation of Management in Competitiveness at PT Bank Central Asia TBK with
bidirectional LSTM for short-and long-term stock market prediction,” an ERM (Enterprise Risk Management) Approach using Likelihood
in 2018 9th International Conference on Information and Table,” Journal of Economics, Finance and Management Studies, vol.
Communication Systems (ICICS), pp. 151–156, 2018, doi: 6, no. 12, Dec. 2023, doi: 10.47191/jefms/v6-i12-01.
10.1109/IACS.2018.8355458. [74] S. M. N. Siahaan, I. Sadalia, and A. S. Silalahi, “Effect of Financial
[61] W. Wei, H. Wu, and H. Ma, “An AutoEncoder and LSTM-Based Ratios on Stock Returns with Earning Per Share as Moderating
Traffic Flow Prediction Method,” Sensors, vol. 19, no. 13, p. 2946, Jul. Variable in Banking Companies on the Indonesia Stock Exchange
2019, doi: 10.3390/s19132946. (2012-2017 Period),” International Journal of Research and Review,
vol. 8, no. 8, pp. 398–406, Aug. 2021, doi: 10.52403/ijrr.20210855.
[62] M. Ma, C. Liu, R. Wei, B. Liang, and J. Dai, “Predicting machine’s
performance record using the stacked long short‐term memory (LSTM) [75] M. L. Widyanto, “Comparative Analysis of PT Bank Central Asia Tbk
neural networks,” J Appl Clin Med Phys, vol. 23, no. 3, Mar. 2022, doi: Performance before and after the COVID 19 Pandemic,” East African
10.1002/acm2.13558. Scholars Journal of Economics, Business and Management, vol. 5, no.
1, pp. 16–21, Jan. 2022, doi: 10.36349/easjebm.2022.v05i01.003.
[63] B. D. Satoto, R. T. Wahyuningrum, and B. K. Khotimah,
“Classification of Corn Seed Quality Using Convolutional Neural [76] B. Hanif, A. Larasati, R. Nurdiansyah, and T. Le, “The Effect of the
Network with Region Proposal and Data Augmentation,” Jurnal Ilmiah Number of Hidden Layers on The Performance of Deep Q-Network for
Teknik Elektro Komputer dan Informatika (JITEKI), vol. 9, no. 2, pp. Traveling Salesman Problem,” Knowledge Engineering and Data
348–362, 2023, doi: 10.26555/jiteki.v9i2.26222. Science, vol. 6, no. 2, p. 188, Oct. 2023, doi:
10.17977/um018v6i22023p188-198.
[64] A. A. Waskita, S. Yushady, C. H. Bissa, I. A. Satya, and R. S. Alwi,
“Development of Novel Machine Learning to Optimize the Solubility [77] H. Aini and H. Haviluddin, “Crude Palm Oil Prediction Based on
of Azathioprine as Anticancer Drug in Supercritical Carbon Dioxide,” Backpropagation Neural Network Approach,” Knowledge Engineering
Jurnal Ilmiah Teknik Elektro Komputer dan Informatika (JITEKI), vol. and Data Science, vol. 2, no. 1, p. 1, Jun. 2019, doi:
9, no. 1, pp. 49–57, 2023, doi: 10.26555/jiteki.v9i1.25608. 10.17977/um018v2i12019p1-9.
[65] D. N. Muhammady, H. A. E. Nugraha, V. R. S. Nastiti, and C. S. K. [78] A. Azhari, A. Susanto, A. Pranolo, and Y. Mao, “Neural Network
Aditya, “Students Final Academic Score Prediction Using Boosting Classification of Brainwave Alpha Signals in Cognitive Activities,”
Regression Algorithms,” Jurnal Ilmiah Teknik Elektro Komputer dan Knowledge Engineering and Data Science, vol. 2, no. 2, p. 47, Dec.
Informatika (JITEKI), vol. 10, no. 1, p. 154, 2024, doi: 2019, doi: 10.17977/um018v2i22019p47-57.
10.26555/jiteki.v10i1.28352. [79] P. Purnawansyah, H. Haviluddin, H. Darwis, H. Azis, and Y. Salim,
[66] A. G. Putrada, N. Alamsyah, I. D. Oktaviani, and M. N. Fauzan, “A “Backpropagation Neural Network with Combination of Activation
Hybrid Genetic Algorithm-Random Forest Regression Method for Functions for Inbound Traffic Prediction,” Knowledge Engineering
Optimum Driver Selection in Online Food Delivery,” Jurnal Ilmiah and Data Science, vol. 4, no. 1, p. 14, Aug. 2021, doi:
Teknik Elektro Komputer dan Informatika (JITEKI), vol. 9, no. 4, pp. 10.17977/um018v4i12021p14-28.
1060–1079, 2023, doi: 10.26555/jiteki.v9i4.27014. [80] K. K. Agustiningsih, E. Utami, and O. M. A. Alsyaibani, “Sentiment
[67] E. Sitompul, R. M. Putra, H. Tarigan, A. Silitonga, and I. Bukhori, Analysis and Topic Modelling of The COVID-19 Vaccine in Indonesia
“Implementation of Digital Feedback Control with Change Rate on Twitter Social Media Using Word Embedding,” Jurnal Ilmiah
Limiter in Regulating Water Flow Rate Using Arduino,” Buletin Ilmiah Teknik Elektro Komputer dan Informatika (JITEKI), vol. 8, no. 1, p.
Sarjana Teknik Elektro, vol. 6, no. 1, pp. 72–82, 2024, doi: 64, 2022, doi: 10.26555/jiteki.v8i1.23009.
10.12928/biste.v6i1.10234. [81] A. P. Wibawa et al., “Deep Learning Approaches with Optimum Alpha
[68] H. D. Trung, “Estimation of Crowd Density Using Image Processing for Energy Usage Forecasting,” Knowledge Engineering and Data
Techniques with Background Pixel Model and Visual Geometry Science, vol. 6, no. 2, p. 170, Oct. 2023, doi:
Group,” Buletin Ilmiah Sarjana Teknik Elektro, vol. 6, no. 2, pp. 142– 10.17977/um018v6i22023p170-187.
154, 2024, doi: 10.12928/biste.v6i2.10785. [82] Y. Sujatna et al., “Stacked LSTM-GRU Long-Term Forecasting Model
[69] M. Jameaba, “Digitalization, Emerging Technologies, and Financial for Indonesian Islamic Banks,” Knowledge Engineering and Data
Stability: Challenges and Opportunities for the Indonesian Banking Science, vol. 6, no. 2, p. 215, 2023, doi:
Sector and Beyond,” SSRN Electronic Journal, 2024, doi: 10.17977/um018v6i22023p215-250.
10.2139/ssrn.4808469.
[70] M. Jameaba, “Digitization revolution, FinTech disruption, and
financial stability: Using the case of Indonesian banking ecosystem to

Furizal, Stock Price Forecasting with Multivariate Time Series Long Short-Term Memory: A Deep Learning Approach