Review of Economics and Finance, 2023, 21, 2603-2610 2603

Improving Global Stock Market Prediction with XGBoost and LightGBM

Machine Learning Models

Omaima Guennioui*, Dalila Chiadmi, Mustapha Amghar

Mohammadia School of Engineering, Mohammed V University in Rabat, Rabat, Morocco.

Abstract: Predicting stock market prices has been a challenging task for analysts and researchers over the years due
to the numerous factors that affect stock prices, making it difficult to achieve accurate predictions. This paper focus-
es on the application of machine learning techniques, specifically Light Gradient-Boosting Machine (LightGBM)
and Extreme Gradient Boosting (XGBoost), to forecast stock market trends. Furthermore, it introduces a hybrid
model called Global XGBoost-LightGBM, which combines these individual models. The study adopts a global ap-
proach by consistently applying parameter settings across multiple series and evaluates the models using a signifi-
cant number of stocks. The effectiveness of this approach is validated through experimentation in two distinct stock
markets: US and Morocco. Experimental results clearly demonstrate that the integration of LightGBM and XGBoost
enhances the ability to identify and interpret fluctuations in stock prices, enabling these models to effectively address
the challenges associated with stock market prediction.
Keywords: Global Forecasting Models, Machine learning, Regression, Stock forecasting.

et al., 2020), but most research has focused on implementing

1. INTRODUCTION
local models with unique parameter sets for each individual
Effective investment selection is crucial for organizations stock. Recently, there have been promising results in global
like pension funds, as their managed funds have a limited forecasting models trained on multiple time series data
viability horizon. To assist in decision-making, implement- (Hewamalage, Bergmeir & Bandara, 2021). A global model
ing stock market prediction models can aid managers in shares the same set of parameters for all series, while a local
making better investment choices. By doing so, they can model uses distinct parameter sets for each individual series
improve their portfolio performance and extend the viability (Januschowski et al., 2020). Compared to local models,
horizon of their managed funds. Predicting stock prices of- global forecasting models have the ability to acquire cross-
fers the potential benefit of enabling investors and managers series information during model training, enabling them to
to make informed decisions, helping them anticipate future manage complexity and avoid over-fitting on a global scale
movements in the stock market and identify potential oppor- (Montero-Manso & Hyndman, 2021). The abundance of time
tunities for profit. series data related to the stock market makes this approach
highly interesting for predicting stock market trends. Using
In most cases, two primary methods are employed for pre- global forecasting models that are trained on multiple, relat-
dicting the direction of stock prices: fundamental analysis ed time series data can improve prediction accuracy.
and technical analysis. Fundamental analysis evaluates the
intrinsic value of a stock by conducting an in-depth analysis Light Gradient-Boosting Machine (LightGBM) and Extreme
of the company's balance sheet, strategic initiatives, and mi- Gradient Boosting (XGBoost) are both gradient trees boost-
croeconomic indicators. On the other hand, technical analy- ing machine learning algorithms that are designed to handle
sis uses technical indicators and charts to identify patterns large and complex datasets (Chen & Guestrin, 2016) (Ke et
and signals of future price fluctuations. al., 2017). XGBoost uses an ensemble learning approach that
builds a sequence of decision trees to make predictions,
Over the past few years, the implementation of machine while LightGBM uses a leaf-wise growth strategy to build
learning methods increased significantly with a potential use decision trees. Both algorithms excel in terms of speed and
in many domains for various purposes (Sarker, 2021). The
accuracy, rendering them well-suited for machine learning
financial industry is one of the fields where machine learning tasks in areas like finance.
has made significant progress and found potential applica-
tions in many areas (Hoang & Wiegratz, 2022). Machine This article investigates the effectiveness of employing
learning methods have been widely validated for stock price LightGBM and XGBoost models in predicting stock prices.
forecasting (Henrique, Sobreiro & Kimura, 2019) (Strader The research adopts a global perspective, employing both
models to forecast a substantial number of stocks in two dis-
tinct and contrasting stock markets: the US and the Moroc-
*Address correspondence to this author at the Mohammadia School of En-
gineers, Av. Ibn Sina, Rabat (Morocco); Tel: (212) 537 68 71 50;
can market. These two markets differ significantly in terms
E-mail: omaimaguennioui@[Link] of maturity and complexity. Moreover, we introduce a hy-
2604 Review of Economics and Finance, 2023, Vol. 21, No. 1 Guennioui et al.

brid model named Global XGBoost-LightGBM, aiming to models: random forest, logistic regression, and LightGBM to
assess its performance in comparison to individual models. forecast the prices of three selected stocks. The models' per-
Finally, our findings are compared with other models such as formance was compared using metrics such as accuracy,
Long Short-Term Memory (LSTM), Gated Recurrent Units recall, precision, and f1 value across different time windows.
(GRU), and Convolutional Neural Networks (CNN). Our The findings provided valuable insights into enhancing both
main contributions in this study are summarized as follows: long-term and short-term forecasting accuracy. The study
1. We have identified two models based on LightGBM highly recommended LightGBM as the preferred model for
medium and long-term share price forecasting due to its su-
and XGBoost that demonstrate strong performance
in global stock market forecasting. perior performance. However, many of these studies and
existing literature heavily rely on stock-specific data as in-
2. We introduce a hybrid model that combines indi- puts for their predictive machine learning algorithms. This
vidual models and investigate its responsiveness approach might overlook crucial market-related information
and predictive capabilities for stock prices. and other relevant factors, which could potentially restrict
3. Our research presents an adaptive global approach the accuracy and efficiency of the models.
capable of effectively predicting a diverse range of In recent years, global time series models have been devel-
stocks in various markets. oped to capitalize on similarities among related time series.
4. To assess the effectiveness of our proposed models, Bojer and Meldgaard (2021) conducted a review of six re-
we evaluate their performance using real data ex- cent Kaggle forecasting competitions, comparing them to the
tracted from two distinct stock markets: the US and M4 Time Series Forecasting Competition. The study empha-
Moroccan market. sizes that the forecasting community can gain valuable in-
sights from the Kaggle community, particularly in forecast-
5. Finally, we pay special attention to analysing the re- ing daily and weekly business time series. Ensemble models
sults and drawing meaningful insights from our ex- consistently outperformed local single models, reaffirming
periments, providing valuable implications for fu- the findings from the M4 competition. Remarkably, machine
ture research. learning methods surpassed conventional time series and
The rest of this paper is organized as follows. The second statistical methods in the recent Kaggle competitions, at-
section reviews related works on stock price prediction, tributed to their use of external information for cross-
global forecasting models and gradient tree boosting ma- learning and modeling the effects of exogenous factors. The
chine learning algorithms. In section 3, we describe the top-ranked solutions in Kaggle competitions and the top two
methodology applied, followed by our experimental results, solutions in the M4 competition relied on either Gradient
analysis, and discussion in Section 4. In Section 5, we sum- Boosting Decision Trees (GBDT) or neural networks. Simi-
marize some conclusions and suggest directions for future larly, Makridakis, Spiliotis, and Assimakopoulos (2022),
research. based on M5 Kaggle's Forecasting Competitions, identified
three key findings for forecasting: (1) LightGBM demon-
2. RELATED WORKS strates remarkable accuracy in predicting retail sales, (2)
incorporating external adjustments and explanatory variables
Numerous studies have explored the application of machine improves precision, and (3) utilizing cross-learning and
learning techniques for predicting stock market trends. Guo, cross-validation is beneficial. Additionally, they highlighted
Li and Xu (2021) developed a hybrid financial time series the dominance of machine learning methods in recent com-
model called LSTM_LightGBM, which uses data such as petitions, underlining the importance of merging statistics
opening, closing, highest, lowest prices, trading volume, and and data science. In this article, we have implemented the
adjusted closing price to make predictions. They found their suggestions and recommendations provided, specifically
model to be stable and effective in forecasting stock price those aimed at improving the accuracy of our forecasting.
fluctuations. Similarly, Yun, Yoon and Won (2021) present-
ed a hybrid GA-XGBoost prediction system that incorporates In conclusion, the review of existing literature highlights the
advanced feature engineering processes and optimal feature limited utilization of ensemble learning algorithms and glob-
set selection. Their research demonstrated that incorporating al forecasting models in the domain of stock market predic-
technical indicators through feature expansion enhances pre- tion. In our study, we aim to address this gap by employing a
diction performance, and the GA-XGBoost algorithm can global forecasting approach and utilizing ensemble tech-
produce an optimal feature set with fewer features. Vijh, niques to develop global forecasting models. Our methodol-
Chandola, Tikkiwal and Kumar (2020) employed two widely ogy involves evaluating both individual models and com-
used machine learning methods, Artificial Neural Network bined model on two distinct stock markets. To validate the
(ANN) and Random Forest (RF), to forecast the next day's effectiveness of our approach, we conducted a comparative
closing stock prices of five companies from various sectors. analysis with other models known for their accurate time
They generated new variables from the financial data (Open, series prediction capabilities.
High, Low, and Close prices) of each company to serve as
inputs to the models. The results revealed that both ANN and 3. MATERIALS AND METHODOLOGY
RF models effectively predict the next day's closing price for 3.1. Model Dataset
the stocks. Hu, Qin, and Zhang (2023) conducted a study on
stock price trend prediction by selecting twelve technical We retrieved the data from Bloomberg and Casablanca Stock
indicators through correlation analysis. They utilized three Exchange website ([Link]) for the
Improving Global Stock Market Prediction Review of Economics and Finance, 2023, Vol. 21, No. 1 2605

period from January 2012 to January 2023. The main varia- BCP
bles of interest were the daily closing prices of individual
stocks and their corresponding sectors. BMCI

In our research, we concentrated on assessing the effective- CDM

ness of our models in two distinct stock markets. The first
CIH
being the US stock market, representing a mature market
characterized by high levels of volatility and returns. The
AFRIC INDUSTRIES SA
second being the Moroccan stock market, representing an
emerging stock market with diverse challenges in the predic- ALUMINIUM DU MAROC
tion of stock prices.
CIMENTS DU MAROC
We chose to study the US stock market through its index,
The Dow Jones Industrial Average (DJIA), which is a stock COLORADO
Building and Construction
market index that tracks the performance of thirty large
Materials
companies listed on various stock exchanges in the United JET CONTRACTORS
States. To do so, we specifically selected stocks from two
major sectors within this index: Financials and Information LAFARGEHOLCIM MAR
Technology. According to data from January 2023, these two
sectors collectively account for more than 35% of the total SONASID
composition of the DJIA. TGCC S. A
Concerning the Moroccan stock market, there are three sec-
Telecommunication ITISSALAT AL-MAGHRIB
tors that collectively account for over 50% of the market's
overall movement. Based on data from January 2023, these Note that our study excluded companies listed after 2012, such as TGCC
sectors consist of the banking index, playing a crucial role in S.A (Building and Construction Materials, Moroccan Stock Market) as there
the Moroccan economy, the building and construction mate- was insufficient data available.
rials index, reflecting the growth in the country's construc-
tion industry, and the telecommunication index. Our study 3.2. Background: Machine learning algorithms
specifically focuses on these three major sectors.
Table 1 and 2 provide the details of the stocks selected for 3.2.1. XGBoost
each stock market. XGBoost by Chen and Guestrin (2016) is an extensively
used open-source software library designed to implement
Table 1. Stocks selected from the US market. gradient boosting machine algorithms. As one of the most
popular boosting tree algorithms in the data science commu-
Sector Members
nity, it has gained significant popularity due to its high prob-
GOLDMAN SACHS GROUP lem-solving performance and minimal feature engineering
requirements (Tamayo et al., 2016). XGBoost is an enhanced
VISA framework of the GBRT model, which is a boosting model
Financials TRAVELERS COS
comprising a series of basic regression trees built using a
sequential ensemble technique. The model can adaptively
AMERICAN EXPRESS add more trees to increase its capacity. Chen and Guestrin
introduced a regularization term to the conventional gradient
JPMORGAN CHASE
boosted regression Trees loss function in XGBoost to penal-
MICROSOFT CORP ize model complexity and prevent overfitting.
SALESFORCE XGBoost's large adoption in the industry stems from its abil-
ity to handle various structured and unstructured datasets.
APPLE Furthermore, XGBoost has consistently given high perfor-
Information Technology mance in diverse data science tasks, including classification,
INTERNATIONAL BUSINESS MA-
regression, and ranking.
CHINES

CISCO SYSTEMS 3.2.2. LightGBM

INTEL CORP LightGBM, a machine learning algorithm introduced by Mi-

crosoft in 2017, belongs to the gradient boosting decision
Table 2. Stocks selected from the Moroccan market. tree family. It is specifically designed for efficiently training
large-scale supervised learning tasks such as regression,
Sector Members classification, and ranking. LightGBM achieves high accura-
cy in these tasks by iteratively building decision trees and
ATTIJARIWAFA BANK optimizing the leaf-wise split strategy, selecting the optimal
Banking feature value for each tree leaf to maximize the overall accu-
BANK OF AFRICA
racy gain (Ke et al., 2017).
2606 Review of Economics and Finance, 2023, Vol. 21, No. 1 Guennioui et al.

Table 3. Hyperparameter tuning details for XGBoost and LightGBM.

Algorithm Hyperparameters Meanings Search Ranges

learning_rate Shrinkage coefficient of each tree (0.05, 0.1, 0.2)

max_depth Maximum depth of a tree (3, 5, 7)

n_estimators Number of iterations (150, 200, 250)

subsample sub-sample size (0.8, 0.9, 1.0)

XGBoost
colsample_bytree Random sampling ratio of features (0.8, 0.9, 1.0)

gamma Minimum loss function degradation required for node splitting (0, 0.1, 0.2)

reg_alpha Regularization on weights (0, 0.1, 0.2)

reg_lambda Regularization on weights (0, 0.1, 0.2)

n_estimators Number of iterations (100, 200, 300)

learning rate Shrinkage coefficient of each tree (0.01, 0.05, 0.1)

num_leaves Maximum number of leaves in a tree (10, 50, 100)

max_depth Maximum tree depth (-1, 5, 10)

LightGBM min_child_samples Minimum number of samples required to create a leaf node (10, 20, 30)

subsample sub-sample size (0.8, 0.9, 1.0)

colsample_bytree Random sampling ratio of features (0.8, 0.9, 1.0)

reg_alpha Regularization on weights (0, 0.1, 0.5)

reg_lambda Regularization on weights (0, 0.1, 0.5)

A primary advantage of LightGBM is its speed and scalabil- the hyperparameter set that delivers the best model perfor-
ity, as it can handle large datasets with millions of examples mance (Liashchynskyi & Liashchynskyi, 2019). By evaluat-
and features. It accomplishes this by employing several tech- ing all potential hyperparameter combinations, we could
niques. Firstly, it uses a histogram-based approach for fea- select the optimal set that produced the best model perfor-
ture discretization, quickly binning continuous features into mance. Grid search is a widely used technique for optimizing
discrete values. Secondly, it adopts a leaf-wise data layout machine learning models, and in our study, we combined it
instead of a level-wise layout, only computing the gradients with cross-validation to improve the generalization perfor-
and hessians for examples within the same leaf, thus reduc- mance of both XGBoost and LightGBM models.
ing computational cost. Lastly, LightGBM supports parallel
The specific details of the hyperparameter tuning can be
and distributed training, allowing it to scale to even larger found in Table 3, which provides information about the hy-
datasets. perparameters and their respective meanings. Furthermore,
we established a search range for each hyperparameter to
3.3. Hyperparameters Tuning explore various values during the tuning procedure.
Numerous researchers have discovered that hyperparameters Taking into consideration the complexity of the American
can significantly influence the final performance of ML stock market and its heightened volatility, and since we be-
models (Rodriguez-Galiano et al., 2015) (Du, Xu and Zhu, lieve that factors influencing stock movements in the United
2021). Therefore, determining the optimal hyperparameters States are also relevant to the Moroccan market. We opted to
is crucial for maximizing the performance of ML models. conduct our parameter tuning using data from this market
Both XGBoost and LightGBM models have various internal and subsequently apply the tuned parameters for predicting
hyperparameters that can affect their prediction accuracy, but the Moroccan stock market. Furthermore, the parameter se-
their default values may not be ideal for stock price predic- lection was made in a way that the tuning process can com-
tion. To identify the best combination of hyperparameters for prehensively explore various aspects of the two model's be-
our specific dataset, we employed a hyperparameter tuning haviours. The chosen values within each parameter range
method called Grid Search. This approach involves system- cover a diverse set of options, enabling the tuning process to
atically assessing the model's performance for each hyperpa- consider a range of settings, including both conservative and
rameter combination within a predefined grid, aiming to find alternative approaches.
Improving Global Stock Market Prediction Review of Economics and Finance, 2023, Vol. 21, No. 1 2607

Fig. (1). Applied process chart.

works (CNN) were also considered, as they can detect tem-

3.4. Global XGBoost-LightGBM poral and spatial patterns in high-dimensional data. Addi-
Global XGBoost-LightGBM is a hybrid model that adopts a tionally, Gated Recurrent Unit (GRU), another type of recur-
global approach for predicting stock prices, where the final rent neural network, was included in our analysis, given its
results are assembled from the predictions of each individual growing popularity in time series prediction tasks.
model: XGBoost and LightGBM. Ensemble learning is a For the programming language, we utilized Python, carrying
machine learning technique that combines the predictions of out the development on Google Colab. For the deep learning
multiple diverse and independent models to improve overall models, TensorFlow and Keras packages were employed. It's
performance and reduce generalization error. important to note that we did not employ any GPU or cloud
In our experiment, we employed a voting ensemble method computing resources during our experimentation.
that uses multiple models to enhance system performance for
both classification and regression problems. Voting ensem- 4.2. Performance Metrics
bles for regression problems are referred to as voting regres- We evaluate prediction performance of our regression prob-
sors (VRs), lem with three different evaluation metrics, including Root
where estimators from all models are averaged to obtain a Mean Square Error (RMSE), Mean Absolute Error (MAE)
final estimate. There are two voting methods: average voting and R-squared (R2). Calculated as follow:
(AV) and weighted voting (WV). In our particular situation,
as XGBoost and LightGBM produced comparable outcomes, (1)
we chose to adopt the average voting approach (Sagi &
Rokach, 2018).
(2)
It is important to mention that prior to combining XGBoost
and LightGBM, we performed hyperparameter tuning for
both models. R2 (3)

4. RESULTS AND DISCUSSION

In this context, the symbol y denotes the actual closing price
4.1. Datasets & Baselines of the test set, while the symbol ỹ represents the predicted
The aim of this study is to forecast the next day's closing value of that price. The symbol ȳ represents the mean value
price for a selected group of stocks. The dataset used com- of y and n denotes the number of data points in the test set.
prises daily closing prices for these stocks throughout the
study period, with adjustments made for bank holidays and 4.3. Experimental Results
alignment across stocks. To train the prediction model, the To prevent issues such as underfitting or overfitting that can
raw dataset was partitioned into training and testing sets at a arise from manually specifying parameters, it is necessary to
ratio of 0.8 to 0.2, respectively. tune models’ parameters. In our study we use GridSearchCV
The performance of our proposed model was compared with for parameter tuning. It employs an exhaustive search ap-
the performance of deep learning models that are known for proach, testing every possible combination of parameters
their effectiveness in time series forecasting (Guennioui, from the given candidates, and selects the best-performing
Chiadmi & Amghar, 2022). Specifically, Long Short-Term parameter set as the final result. Thus, it ensures the accuracy
Memory (LSTM) networks, a type of Recurrent Neural Net- and reliability of results. The evaluation is based on the
works, were chosen due to their capacity to retain infor- search results of different parameter combinations, ultimate-
mation over extended periods. Convolutional Neural Net- ly leading to an optimal result within the specified parameter
2608 Review of Economics and Finance, 2023, Vol. 21, No. 1 Guennioui et al.

range that best fits the model. The results of our parameter error is the average of the errors calculated for each individ-
tuning for XGBoost and LightGBM are summarized in Table ual stock prediction.
4: Table 5 presents the evaluation metrics for the prediction
results of different models, including RMSE, MAE, and R2
Table 4. Hyperparameter Tuning Results.
score. Among the tested models, parameter optimization was
Algorithm Hyperparameters Optimal Values
employed to achieve the highest level of accuracy.
The table highlights the performances of different models
learning_rate 0.2
applied to our data in the two selected stock markets: US and
max_depth 3 Morocco.
For stock predictions in the US market, it is observed that the
n_estimators 200
G-LightGBM, G-XGBoost, and Global XGBoost-LightGBM
subsample 0.9 methods achieve similar performances, with a slight im-
XGBoost provement in accuracy when using the model built by com-
colsample_bytree 0.9 bining LightGBM and XGBoost. The RMSE and MAE cal-
culated for the three models, G-LightGBM, G-XGBoost, and
gamma 0
Global XGBoost-LightGBM, are very low (around 0.018 for
reg_alpha 0 RMSE and 0.0133 for MAE). Furthermore, all three models
exhibit a high R2 score, indicating the ability to account for
reg_lambda 0.1 about 99% of the variance in the target variable. These re-
n_estimators 300
sults indicate that these methods are accurate in predicting
stocks in the US market. As for the methods based on neural
learning rate 0.1 networks (LSTM, GRU, and CNN), lower performances are
observed. The RMSE and MAE values for these methods are
num_leaves 10 higher (0.0188 / 0.0144 for LSTM, 0.0255/ 0.0201 for GRU,
max_depth -1
and 0.0194/0.0146 for CNN). And the R2 score is lower for
LSTM, GRU, and CNN (0.967, 0.9913, and 0.963, respec-
LightGBM min_child_samples 10 tively), indicating that these models explain less variance in
the data compared to the proposed methods. Additionally, as
subsample 0.8 these models are designed to build individual models for
colsample_bytree 0.8
each dataset, they require significant computational re-
sources.
reg_alpha 0.0
Regarding stock predictions in the Moroccan market, the G-
reg_lambda 0.5 LightGBM, G-XGBoost, and Global XGBoost-LightGBM
methods achieve similar results to those in the US market,
Baseline models were constructed for each stock in the se- with very low RMSE values (0.0192, 0.0195, and 0.0191,
lected group using a local approach, where past observations respectively) and very low MAE values (0.0119, 0.0124, and
were used to develop the model. The reported prediction 0.0119, respectively). The R2 score is also high for these

Table 5. Performance Results.

Market Method Approach RMSE MAE R2 score

G-LightGBM Global 0.0188 0.0140 0.989

G-XGBoost Global 0.0187 0.0137 0.989

Global XGBoost-LightGBM Global 0.0180 0.0133 0.990

US
LSTM Local 0.0188 0.0144 0.967

GRU Local 0.0255 0.0201 0.913

CNN Local 0.0194 0.0146 0.963

G-LightGBM Global 0.0192 0.0119 0.991

G-XGBoost Global 0.0195 0.0124 0.991

Global XGBoost-LightGBM Global 0.0191 0.0119 0.991

Morocco
LSTM Local 0.0194 0.0138 0.978

GRU Local 0.0197 0.0143 0.977

CNN Local 0.0192 0.0135 0.979

Improving Global Stock Market Prediction Review of Economics and Finance, 2023, Vol. 21, No. 1 2609

Fig. (2). Predicted closing price of Goldman Sachs and Attijariwafa Bank.

three methods (0.991), indicating an accurate predictive ca- predictions and the proportion of variance explained. Addi-
pacity in the Moroccan market. As for the methods based on tionally, we observed that, under the evaluated conditions,
neural networks, lower performances are observed compared there is no benefit in utilizing the hybrid model in the Mo-
to tree-based methods. Furthermore, The RMSE and MAE roccan market compared to the US market. Moreover, the
values for LSTM, GRU, and CNN are all close, indicating a proposed models were designed to predict a significant por-
similar prediction error for these methods. tion of the markets and have succeeded in doing so. Finally,
These results suggest that, for both markets, the proposed optimizing the parameters in the US market and transferring
these same parameters to the Moroccan market has yielded
models: G-LightGBM, G-XGBoost, and especially their en-
semble, Global XGBoost-LightGBM, outperform LSTM, accurate results. This suggests the possibility of generalizing
GRU, and CNN models in terms of both the average error of to stock markets belonging to different economies and influ-
enced by the same factors driving stock movements.
2610 Review of Economics and Finance, 2023, Vol. 21, No. 1 Guennioui et al.

Fig. (2) showcases a composite chart of the best-performing Hewamalage, H., Bergmeir, C., & Bandara, K. (2021). Global Models for
algorithms from both the baseline and the proposed models. Time Series Forecasting: A Simulation Study. Pattern Recognition
(Vol 124).
It displays the prediction results for two major banks in the Hoang, D., & Wiegratz, K. (2022). Machine Learning Methods in Finance:
US and Moroccan markets (Goldman Sachs and Attijariwafa Recent Applications and Prospects. SSRN Electronic Journal.
Bank). The predictions are made on the test set. A specific Hu, Q., Qin, S., & Zhang, S. (2023). Comparison of Stock Price Prediction
section of the data is also highlighted and presented in the Based on Different Machine Learning Approaches. Proceedings of
figure for a more detailed view. the 2022 International Conference on Bigdata Blockchain and
Economy Management (ICBBEM 2022).
[Link]
5. CONCLUSION Januschowski, T., Gasthaus, J., Wang, Y., Salinas, D., Flunkert, V., Bohlke-
Schneider, M., & Callot, L. (2020). Criteria for classifying forecast-
Predicting financial asset prices, such as stocks, is a critical ing methods. In International Journal of Forecasting (Vol. 36, Issue
task in economic management. As we enter the era of big 1).
data, with an overwhelming amount of information available, Ke, G., Meng, Q., Finley, T., Wang, T., Chen, W., Ma, W., Ye, Q., & Liu,
there's an increasing trend towards employing computer al- T.-Y. (2017). LightGBM: A Highly Efficient Gradient Boosting
gorithms to address challenges associated with market pre- Decision Tree. Curran Associates, Inc (Vol. 30).
Makridakis, S., Spiliotis, E., & Assimakopoulos, V. (2022). M5 accuracy
diction. This study investigates the effectiveness of utilizing competition: Results, findings, and conclusions. International Jour-
LightGBM and XGBoost, as well as a hybrid model that nal of Forecasting, 38(4), 1346–1364.
combines both, through a global forecasting approach for Montero-Manso, P., & Hyndman, R. J. (2021). Principles and Algorithms
stock price prediction. To demonstrate the efficacy of our for Forecasting Groups of Time Series: Locality and Globality. In-
approach, a comparison with deep learning models such as ternational Journal of Forecasting (Vol. 37, Issue 4).
Liashchynskyi, Petro & Liashchynskyi, Pavlo. (2019). Grid Search, Random
CNN, LSTM and GRU is conducted. The evaluation was Search, Genetic Algorithm: A Big Comparison for NAS. ArXiv
performed in two different stock markets. abs/1912.06059.
Rodriguez-Galiano, V., Sanchez-Castillo, M., Chica-Olmo, M., & Chica-
There are several directions that can be explored following Rivas, M. (2015). Machine learning predictive models for mineral
this project. For instance, the approach could be extended to prospectivity: An evaluation of neural networks, random forest, re-
cover other diverse markets. Additionally, as an enhance- gression trees and support vector machines. Ore Geology Reviews,
ment to the proposed model, the ensembling method could 71. [Link]
be further developed to better address the challenges of stock Sagi, O., & Rokach, L. (2018). Ensemble learning: A survey. In Wiley
Interdisciplinary Reviews: Data Mining and Knowledge Discovery
price prediction. (Vol. 8, Issue 4).
Sarker, Iqbal. (2021). Machine Learning: Algorithms, Real-World Applica-
REFERENCES tions and Research Directions. SN Computer Science. 2.
10.1007/s42979-021-00592-x.
Adam-Bourdarios, C., Cowan, G., Germain-Renaud, C., Guyon, I., Kégl, B.,
Strader, T. J., Rozycki, J. J., ROOT, T. H., & Huang, Y.-H. J. (2020). Ma-
&
chine Learning Stock Market Prediction Studies: Review and Re-
Rousseau, D. (2015). The higgs machine learning challenge. Journal of
search Directions. Journal of International Technology and Infor-
Physics: Conference Series, 664(7). [Link]
mation Management, 28(4). [Link]
6596/664/7/072015
6679.1435
Bojer, C. S., & Meldgaard, J. P. (2021). Kaggle forecasting competitions:
Tamayo, D., Silburt, A., Valencia, D., Menou, K., Ali-Dib, M., Petrovich,
An
C., Huang, C. X., Rein, H., van Laerhoven, C., Paradise, A., Ober-
Overlooked learning opportunity. International Journal of Forecasting,
tas, A., & Murray, N. (2016). A Machine Learns to Predict the Sta-
37(2). [Link]
bility of Tightly Packed Planetary Systems. The Astrophysical
Chen, T., & Guestrin, C. (2016). XGBoost: A Scalable Tree Boosting Sys-
Journal, 832(2), L22. [Link]
tem.
Vijh, M., Chandola, D., Tikkiwal, V. A., & Kumar, A. (2020). Stock Clos-
Proceedings of the ACM SIGKDD International Conference on Knowledge
ing Price Prediction using Machine Learning Techniques. Procedia
Discovery and Data Mining, 13-17-August-2016, 785–794.
Computer Science, 167, 599–606.
[Link]
[Link]
Guennioui Omaima, Chiadmi Dalila, Amghar Mustapha. (2022). Systematic
Xianping Du and Hongyi Xu and Feng Zhu. (2021). Understanding the
mapping study on Applications of deep learning in stock market
Effect of Hyperparameter Optimization on Machine Learning
prediction. The 10th International Conference on Innovation and
Models for Structure Design Problems. Computer-Aided Design,
New Trends in Information Technology. In press.
135, 103013. [Link]
Guo, Y., Li, Y., & Xu, Y. (2021). Study on the application of LSTM-
Yun, K. K., Yoon, S. W., & Won, D. (2021). Prediction of stock price direc-
LightGBM Model in stock rise and fall prediction. MATEC Web of
tion using a hybrid GA-XGBoost algorithm with a three-stage fea-
Conferences, 336, 05011.
ture engineering process. Expert Systems with Applications, 186.
[Link]
[Link]
Henrique, B. M., Sobreiro, V. A., & Kimura, H. (2019). Literature review:
Machine learning techniques applied to financial market prediction.
Expert Systems with Applications, 124, 226–251.

Received: Oct 05, 2023 Revised: Oct 08, 2023 Accepted: Oct 12, 2023

Copyright © 2023– All Rights Reserved

This is an open-access article.