THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE09
67
https://www.theamericanjournals.com/index.php/tajet
PUBLISHED DATE: - 20-09-2024
https://doi.org/10.37547/tajet/Volume06Issue09-09
PAGE NO.: - 67-80
OPTIMIZING RETAIL DEMAND
FORECASTING: A PERFORMANCE
EVALUATION OF MACHINE LEARNING
MODELS INCLUDING LSTM AND GRADIENT
BOOSTING
Md Shujan Shak
Master of Science in Information Technology, Washington University of
Science and Technology, USA
Md Shahin Alam Mozumder
Master of Science in Information Technology, Washington University of
Science and Technology, USA
Md Amit Hasan
Master of Science in Information Technology, Washington University of
Science and Technology, USA
Ashim Chandra Das
Master of Science in Information Technology, Washington University of
Science and Technology, USA
Md Rashel Miah
Department of Digital Communication and Media/Multimedia, Westcliff
University, USA
Salma Akter
Department of Public Administration, Gannon University, Erie, PA, USA
Md Nur Hossain
Master’s in information technology management, Webster University
, USA
RESEARCH ARTICLE
Open Access
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE09
68
https://www.theamericanjournals.com/index.php/tajet
INTRODUCTION
Retail demand forecasting plays a critical role in
supply chain management, inventory control, and
overall business planning for retail organizations.
Accurate forecasting enables retailers to optimize
stock levels, reduce costs, and enhance customer
satisfaction by ensuring products are available
when and where they are needed (Chopra &
Meindl, 2016). However, retail demand is
influenced by a multitude of factors, including
seasonality, promotional activities, holidays, and
external economic conditions, making demand
forecasting a complex task (Fildes et al., 2019). In
recent years, advancements in machine learning
(ML) algorithms have shown significant promise in
improving the accuracy of demand forecasts by
learning from historical data and capturing
intricate patterns in consumer behavior (Zhao et
al., 2021).
Traditional forecasting methods such as Linear
Regression (LR) have been widely used in retail
but often fall short in handling non-linear
relationships and complex interactions between
demand drivers (Makridakis et al., 2018). Machine
learning models like Decision Tree Regressor
(DTR), Random Forest Regressor (RFR), Gradient
Boosting (GB), and Long Short-Term Memory
(LSTM) offer more sophisticated approaches by
leveraging the power of non-parametric modeling,
ensemble learning, and deep learning techniques
(Bajari et al., 2019). These models have been
effective in accounting for external influences such
as weather conditions, economic indicators, and
promotional campaigns, which significantly
impact consumer demand patterns (Taddy, 2019).
The objective of this study is to evaluate the
performance of multiple machine learning models
in retail demand forecasting and compare their
ability to capture temporal dependencies,
seasonality, and other factors that influence retail
demand. By analyzing models such as LR, DTR,
RFR, GB, and LSTM, this research aims to identify
the most effective algorithm for accurately
forecasting retail demand and assisting retailers in
making data-driven decisions.
The importance of demand forecasting in the retail
industry has been well-documented in both
academic and industry research. Traditionally,
statistical models such as Exponential Smoothing
and ARIMA have been employed for demand
forecasting (Hyndman & Athanasopoulos, 2018).
However, these models often struggle to capture
complex relationships in retail data, particularly
when non-linear factors such as promotions,
seasonality, and economic fluctuations come into
Abstract
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE09
69
https://www.theamericanjournals.com/index.php/tajet
play (Syntetos et al., 2009).
With the advent of machine learning, researchers
have explored the use of more advanced
algorithms for improving demand forecasting
accuracy. One such advancement is the application
of Decision Tree Regressors (DTRs), which can
model non-linear relationships by recursively
splitting the dataset into subsets based on decision
criteria (Breiman, 2017). DTRs have been found to
perform well in handling categorical data and
capturing key drivers of demand, such as product
characteristics and promotional activities (Keerthi
& Lin, 2020). However, they are prone to
overfitting, especially when used without
regularization or ensemble techniques (Hastie et
al., 2009).
Random Forest Regressors (RFR), an ensemble of
decision trees, have been proposed as a solution to
the overfitting problem (Breiman, 2001). By
averaging the results of multiple decision trees,
RFR reduces variance and improves generalization
to unseen data (Liaw & Wiener, 2002). Several
studies have demonstrated the effectiveness of
RFR in retail demand forecasting, particularly in
handling complex, high-dimensional datasets
(Cortez et al., 2021). Random Forests have shown
strong performance in capturing seasonality and
other recurring patterns in demand data
(Hyndman et al., 2021).
LITERATURE REVIEW
Accurate retail demand forecasting is crucial for
optimizing supply chain management, reducing
inventory costs, and enhancing customer
satisfaction. Traditionally, statistical models such
as Linear Regression (LR) and the Autoregressive
Integrated Moving Average (ARIMA) have been
used for predicting retail demand due to their
simplicity and interpretability (Box & Jenkins,
1970). However, these methods often struggle to
capture complex, non-linear relationships and are
less effective in addressing the dynamic and
unpredictable
nature
of
modern
retail
environments (Hyndman & Athanasopoulos,
2018). As a result, researchers and practitioners
have increasingly turned to advanced machine
learning (ML) models to improve demand
forecasting accuracy.
Recent advancements in machine learning have
introduced more sophisticated techniques that
outperform traditional models in various
predictive tasks. For example, Random Forest
Regressor (RFR) and Gradient Boosting (GB) have
demonstrated superior performance in handling
non-linear relationships and high-dimensional
data (Breiman, 2001; Friedman, 2001). These
models leverage ensemble learning methods to
improve accuracy and robustness by combining
the predictive capabilities of multiple decision
trees. Studies have shown that these models are
effective at forecasting retail demand by capturing
complex interactions between variables, such as
the effects of promotions, holidays, and external
factors (Lima et al., 2021; Zhao et al., 2021).
Moreover, the application of Long Short-Term
Memory (LSTM) networks has further advanced
the field of retail demand forecasting. LSTM, a type
of recurrent neural network (RNN), is particularly
adept at modeling time series data due to its ability
to retain and utilize past information to predict
future outcomes (Hochreiter & Schmidhuber,
1997). Unlike traditional models that assume
independence between observations, LSTM excels
at capturing temporal dependencies, making it
highly suitable for retail scenarios where demand
patterns fluctuate over time due to seasonality and
other time-dependent factors (Brownlee, 2018).
Research has shown that LSTM consistently
outperforms both traditional statistical methods
and simpler machine learning models, especially
when forecasting tasks involve large datasets and
long-term patterns (Livieris et al., 2020).
The literature clearly indicates that while
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE09
70
https://www.theamericanjournals.com/index.php/tajet
traditional models like LR and ARIMA provide a
solid foundation for demand forecasting, machine
learning models, particularly RFR, GB, and LSTM,
offer significant improvements in accuracy. These
advanced models are better suited to capturing the
complex, non-linear relationships inherent in
retail data, particularly when enriched with
additional contextual information such as product
details, promotions, and external factors (Zhao et
al., 2021). This study builds on these
advancements by evaluating the performance of
various machine learning models in retail demand
forecasting, focusing on their ability to handle
temporal patterns, seasonality, and promotional
events.
Another powerful machine learning model,
Gradient Boosting (GB), has gained popularity for
its ability to iteratively learn from the errors of
previous models, gradually improving its
prediction accuracy (Friedman, 2001). Gradient
Boosting models are known for their high accuracy
in a variety of applications, including demand
forecasting (Natekin & Knoll, 2013). Studies have
highlighted the ability of GB models to capture
complex interactions between features, such as the
combined effects of promotions and holidays, and
adjust for them over time (Chen & Guestrin, 2016).
More recently, Long Short-Term Memory (LSTM)
networks, a type of recurrent neural network, have
been employed in retail demand forecasting due to
their capacity to model temporal dependencies
(Hochreiter & Schmidhuber, 1997). LSTM models
have proven highly effective in forecasting tasks
involving time series data, as they can retain and
utilize past information to predict future
outcomes. In the retail context, LSTMs excel at
capturing seasonality, demand spikes, and other
time-dependent patterns (Brownlee, 2018).
Numerous studies have shown that LSTM
outperforms traditional statistical methods and
simpler machine learning models in demand
forecasting tasks, particularly when dealing with
long-term patterns and large datasets (Livieris et
al., 2020).
Overall, the literature suggests that while
traditional models like LR and ARIMA provide a
solid baseline for retail demand forecasting,
machine learning models such as RFR, GB, and
LSTM offer significant improvements in capturing
complex, non-linear relationships and temporal
patterns in retail data. The combination of these
advanced models with rich datasets, including
sales transactions, product information, and
external factors, holds great promise for more
accurate demand forecasting and better inventory
management (Zhao et al., 2021).
METHODOLOGY
The methodology for this study involved several
critical steps to ensure accurate and reliable retail
demand forecasting using different machine
learning algorithms. These steps included data
collection and preprocessing, feature engineering,
model selection and training, performance
evaluation, and final comparison. Each stage of the
methodology was designed to address the unique
challenges of retail demand forecasting, such as
handling high-dimensional data, capturing
complex demand patterns, and accounting for
seasonality and promotions.
1. Data Collection
The success of any machine learning model is
heavily dependent on the quality and relevance of
the data used to train it, and this is especially true
in retail demand forecasting. In this study, data
collection played a critical role in building accurate
models capable of predicting future demand based
on historical patterns. The dataset used in this
research was sourced from a retail organization,
encompassing a wide range of factors that
influence consumer purchasing behavior. The
dataset included daily sales transactions, product
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE09
71
https://www.theamericanjournals.com/index.php/tajet
details, promotional information, and external
factors such as holidays and economic indicators.
This diverse data provided a rich foundation for
the models to learn from and make precise
forecasts.
1.1 Sales Data
The core of the dataset comprised historical sales
data from the retail organization. This data
included transactional records at a granular level,
with information on the number of units sold,
revenue generated, and product categories. For
each transaction, details such as the product
identifier, store location, and transaction date
were recorded. This granular-level data allowed
the models to track trends over time, detect
seasonality, and identify spikes in demand related
to specific products or categories. The sales data
spanned multiple years, which provided a robust
foundation for understanding long-term trends
and recurring patterns in demand.
1.2 Product Information
To enhance the sales data, additional product-level
information was incorporated. This included
details about the type of product, its price, and
product category. Product information is crucial
for demand forecasting, as different products
exhibit varying demand patterns depending on
their attributes. For example, high-priced items
may have less frequent but larger demand spikes,
while everyday consumer goods may show steady
demand with minimal fluctuation. By including
product features in the data, the models could
make more nuanced predictions that take into
account the inherent characteristics of each item.
Additionally, stock-keeping unit (SKU) identifiers
were used to track individual products, ensuring
that forecasts could be made at both the aggregate
and SKU-specific levels.
1.3 Promotional and Discount Data
Promotional campaigns and discounts are some of
the most significant drivers of demand fluctuations
in the retail industry. To capture the impact of
these factors, data on promotional activities such
as discounts, coupons, and special sales events
were included. This promotional data was aligned
with the transactional sales data to enable the
models to understand how demand surged or
declined during promotional periods. Variables
such as the type of promotion, its duration, and the
discount percentage were crucial for predicting
demand spikes during promotional events. By
incorporating this data, the machine learning
models were able to anticipate short-term
increases in demand, making the forecasts more
accurate during sale periods.
1.4 Calendar and Holiday Data
Retail demand is often influenced by seasonal
factors, holidays, and events, which lead to
predictable shifts in consumer behavior. To
account for these effects, data on holidays, special
events, and calendar dates were integrated into the
dataset. National holidays, religious festivals, and
annual shopping events such as Black Friday and
Cyber Monday were included to help the models
forecast demand surges during these periods.
Additionally, calendar-based features such as the
day of the week, month, and quarter were added to
capture recurring weekly and monthly trends. For
instance, weekends typically experience higher
sales in some product categories, while certain
months might show increased demand due to
seasonal factors.
1.5 Weather and External Data
To further enrich the dataset and improve the
predictive accuracy of the models, external data
such as weather conditions and economic
indicators were incorporated. Weather data,
including temperature, precipitation, and extreme
weather events, was collected for each store
location. Weather can significantly affect consumer
behavior, as severe weather conditions often lead
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE09
72
https://www.theamericanjournals.com/index.php/tajet
to changes in shopping habits. For example,
extreme heat or cold may discourage in-store
shopping, while rainy weather might increase the
demand for certain products, such as umbrellas or
cold weather gear. By including weather variables,
the models were able to capture these external
influences on retail demand. Additionally,
economic indicators such as inflation rates,
unemployment levels, and consumer confidence
indices were integrated into the dataset. Economic
conditions can play a major role in shaping
consumer spending patterns. For instance, during
periods of economic downturn, consumers may
reduce discretionary spending, while in times of
economic growth, they may increase purchases.
Incorporating these macroeconomic variables
allowed the models to better account for long-term
shifts in demand driven by changes in the broader
economy.
2. Data Preprocessing
Preprocessing the data was a crucial step to ensure
the models could effectively learn from the data.
The raw data contained missing values, outliers,
and inconsistencies that needed to be addressed
before training the machine learning models. The
following preprocessing steps were applied:
•
Handling Missing Values: Missing sales or
feature data were imputed using statistical
methods like mean or median values for numerical
data or the most frequent category for categorical
data. For time series data, missing entries were
handled by forward or backward filling
techniques.
•
Outlier Detection and Removal: Sales spikes
or drops that were not related to actual market
trends or promotions were identified as outliers.
These outliers were either removed or treated
using techniques such as capping or transforming
the data to avoid skewing the model’s predictions.
•
Feature Scaling: To ensure that the models
could properly interpret the data, feature scaling
was applied where necessary. Continuous features
such as sales volume and price were scaled using
normalization or standardization techniques to
ensure they fell within a similar range.
•
One-Hot Encoding for Categorical Variables:
Categorical variables such as promotion types,
holidays,
and
product
categories
were
transformed into numerical values using one-hot
encoding to ensure the machine learning models
could process them effectively.
•
Time Series Transformation: For models
such as Long Short-Term Memory (LSTM), the data
was transformed into sequences to capture the
temporal relationships between sales at different
time points. Lag features were created to help the
models understand how previous days' sales
influenced future demand.
3. Feature Engineering
Feature engineering was performed to create new
variables from the existing data, providing the
models with more informative inputs. Features
such as moving averages, rolling windows, and lag
variables were generated to capture temporal
dependencies in the data. Additionally, interaction
terms were created to model complex
relationships between variables, such as the
interaction between promotions and holidays.
Calendar features like day of the week, month, and
season were also incorporated to account for
seasonal patterns in consumer demand.
4. Model Selection
The study involved the evaluation of several
machine learning algorithms, each chosen for its
specific strengths in handling different aspects of
demand forecasting. The selected models included:
•
Linear Regression (LR): Used as a baseline
model to provide a simple and interpretable
forecast based on a linear relationship between
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE09
73
https://www.theamericanjournals.com/index.php/tajet
features and demand.
•
Decision Tree Regressor (DTR): Selected for
its ability to handle non-linear relationships in the
data by splitting features into decision nodes based
on their influence on demand.
•
Random Forest Regressor (RFR): Chosen for
its ensemble learning technique that combines
multiple decision trees to reduce overfitting and
improve predictive performance.
•
Gradient Boosting (GB): A boosting
algorithm selected for its iterative approach, which
allows it to fine-tune predictions by learning from
previous errors and capturing complex feature
interactions.
•
Long Short-Term Memory (LSTM): A deep
learning model chosen for its ability to capture
long-term dependencies in time series data,
making it highly suitable for retail demand
forecasting where temporal patterns are crucial.
5. Model Training and Hyperparameter Tuning
Each model was trained using the preprocessed
and engineered data. A split was performed to
divide the dataset into training and test sets,
ensuring that the models were trained on
historical data and validated on unseen data.
Cross-validation techniques were employed to
minimize overfitting and improve generalization
performance. To optimize model performance,
hyperparameter tuning was conducted using grid
search and random search techniques. For each
model, the most critical hyperparameters, such as
the number of trees in Random Forest, the learning
rate in Gradient Boosting, and the number of units
in LSTM, were tuned to identify the best
configuration for the dataset. This step ensured
that each model operated at peak efficiency,
providing the best possible forecast for retail
demand.
6. Model Evaluation
After training, the models were evaluated using
standard regression metrics to assess their
performance in predicting future demand. The
chosen evaluation metrics included:
•
Mean Absolute Error (MAE): To measure the
average magnitude of errors in the predictions,
regardless of direction.
•
Root Mean Square Error (RMSE): To
penalize larger errors more significantly,
providing a measure of the model’s accuracy.
•
R-squared (R²): To determine the
proportion of the variance in the dependent
variable that is predictable from the independent
variables, indicating how well the model fits the
data.
These metrics allowed for a detailed comparison of
each model’s accuracy, error rates, and ability to
handle complex retail demand patterns.
7. Performance Comparison and Final Selection
Once all models were trained and evaluated, their
performance metrics were compared to identify
the best-performing model. The comparison
focused on the ability of each model to handle
temporal patterns, seasonality, demand spikes,
and long-term trends. The LSTM model emerged as
the top performer, demonstrating superior
accuracy in capturing temporal dependencies.
Gradient Boosting and Random Forest also
performed well, providing robust forecasts for
non-linear and seasonal demand patterns. Linear
Regression and Decision Tree Regressor, on the
other hand, showed limitations in their ability to
handle complex relationships and variability in the
retail data.
The final selection was based on the balance
between model accuracy, interpretability, and
computational efficiency, with LSTM being
recommended for deployment due to its superior
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE09
74
https://www.theamericanjournals.com/index.php/tajet
performance. However, Gradient Boosting and
Random Forest were considered strong
alternatives,
especially
in
cases
where
computational resources or model interpretability
were prioritized.
RESULT
1. Linear Regression (LR)
Linear Regression (LR) was selected as the
baseline model for retail demand forecasting in
this study. As a simple and interpretable algorithm,
LR assumes a linear relationship between the input
features and the target variable, making it a widely
used approach for regression tasks. While LR
performed reasonably well in capturing general
demand trends, it exhibited significant limitations
in forecasting more complex demand patterns,
particularly during periods of sharp fluctuations
such as peak sale seasons and promotional events.
This
model's
tendency
to
oversimplify
relationships between the variables led to higher
errors when it encountered nonlinear behavior,
such as sudden demand spikes or seasonal
variations. Additionally, the model struggled with
high-dimensional data, where the assumption of
linearity did not hold. The performance metrics
reflect these challenges, with a Mean Absolute
Error (MAE) of 15.34, a Root Mean Square Error
(RMSE) of 20.57, and an R-squared (R²) score of
0.71. These results indicate that while LR can
provide a quick and basic forecast, it lacks the
sophistication needed for accurate demand
prediction in retail environments characterized by
high variability.
2. Decision Tree Regressor (DTR)
The Decision Tree Regressor (DTR) showed an
improvement over Linear Regression by capturing
nonlinear relationships between features and
demand. DTR is a non-parametric model that splits
the dataset into branches based on feature values,
making it more adaptable to complex data
patterns. In the context of retail demand
forecasting, the DTR model was able to identify
decision points where certain features, such as
holidays or sales promotions, significantly
influenced demand. This flexibility allowed DTR to
better handle the fluctuations and seasonality in
retail data. However, a notable drawback of the
model was its propensity to overfit the training
data, especially when dealing with high-
dimensional datasets. This overfitting resulted in
diminished generalization capability when
predicting on new, unseen data. The model
achieved a Mean Absolute Error (MAE) of 12.78, a
Root Mean Square Error (RMSE) of 18.11, and an
R-squared (R²) score of 0.75. While the model
demonstrated improved accuracy over Linear
Regression, its susceptibility to overfitting
suggests that further optimization, such as pruning
or regularization, would be necessary to enhance
its robustness for demand forecasting.
3. Random Forest Regressor (RFR)
The Random Forest Regressor (RFR) provided a
significant leap in performance compared to both
Linear Regression and Decision Tree Regressor. As
an ensemble learning technique, RFR builds
multiple decision trees and averages their
predictions to reduce overfitting and improve
predictive accuracy. This characteristic proved to
be highly beneficial for retail demand forecasting,
where randomness in feature selection and the
aggregation of diverse trees helped the model
capture complex patterns, such as seasonality and
sudden demand shifts, without succumbing to
overfitting. The Random Forest model was
particularly adept at handling the dynamic nature
of retail data, where multiple factors like holidays,
promotions, and market trends influence demand
simultaneously. With its ability to handle large
amounts of data and provide robust results, RFR
outperformed its simpler counterparts with a
Mean Absolute Error (MAE) of 11.22, a Root Mean
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE09
75
https://www.theamericanjournals.com/index.php/tajet
Square Error (RMSE) of 16.45, and an R-squared
(R²) score of 0.83. The improved accuracy and
reduced error margins make RFR a strong
candidate for retail demand forecasting, especially
when the data exhibits high variability and
complexity.
4. Gradient Boosting (GB)
Gradient Boosting (GB) emerged as one of the top-
performing models in this study, offering high
accuracy in retail demand forecasting. Unlike
Random Forest, which builds trees independently,
Gradient Boosting builds trees sequentially, with
each tree attempting to correct the errors made by
the previous one. This iterative approach enabled
the GB model to fine-tune its predictions, making it
particularly effective at capturing intricate
patterns in the data, including both short-term
fluctuations and long-term seasonal trends. In
retail demand forecasting, GB's ability to model
complex interactions between features, such as the
impact of pricing, promotions, and external factors
like holidays, proved to be advantageous. The
model consistently delivered strong predictive
performance, with a Mean Absolute Error (MAE) of
10.68, a Root Mean Square Error (RMSE) of 15.89,
and an R-squared (R²) score of 0.87. These metrics
demonstrate GB's capacity to handle non-linear
and complex relationships, making it an ideal
choice for predicting retail demand where multiple
factors interact in unpredictable ways. However,
one limitation is the computational intensity of the
model, which can be resource-heavy and time-
consuming, especially when dealing with large
datasets.
5. Long Short-Term Memory (LSTM)
The Long Short-Term Memory (LSTM) model
outperformed all traditional machine learning
models in this study, showcasing its exceptional
ability to forecast retail demand. LSTM is a type of
recurrent neural network (RNN) specifically
designed for time series data, making it
particularly well-suited for retail forecasting,
where demand patterns often exhibit temporal
dependencies. Unlike traditional models, LSTM can
retain information over long periods, allowing it to
effectively model both short-term demand spikes
(such as during a sale) and long-term seasonal
trends (like holiday shopping periods). In this
study, the LSTM model was able to capture
complex temporal patterns, identifying crucial
factors like recurring weekly and monthly sales
patterns, as well as the effects of special
promotions and holidays. The model's strong
learning capabilities are reflected in its
performance metrics: a Mean Absolute Error
(MAE) of 9.53, a Root Mean Square Error (RMSE)
of 14.67, and an R-squared (R²) score of 0.90.
These results demonstrate that LSTM is highly
effective at capturing the temporal dynamics
inherent in retail data, making it the most accurate
model for forecasting future demand. Despite its
superior performance, LSTM does require more
computational resources and longer training times
compared to traditional models, which could be a
consideration for deployment in real-time retail
forecasting environments.
6. Performance Comparison
Among all the models, LSTM demonstrated the
best performance, particularly in handling
temporal patterns and demand spikes. Gradient
Boosting and Random Forest also performed well,
providing high accuracy without significant
overfitting. On the other hand, Linear Regression
struggled with non-linear trends and seasonality,
making it the least effective model for demand
forecasting. In the table 1 we illiterate the result
comparison.
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE09
76
https://www.theamericanjournals.com/index.php/tajet
Table 1: Evaluation of model performance
Model
MAE
RMSE
R²
Linear Regression
15.34
20.57
0.71
Decision Tree Regressor
12.78
18.11
0.75
Random Forest Regressor
11.22
16.45
0.83
Gradient Boosting
10.68
15.89
0.87
LSTM
9.53
14.67
0.90
The performance comparison of the machine
learning models in the chart 1 for retail demand
forecasting highlights the strengths and
weaknesses of each algorithm in terms of accuracy,
error rates, and ability to handle complex patterns
in the data. Among the evaluated models, the Long
Short-Term Memory (LSTM) network stood out as
the top performer. With a Mean Absolute Error
(MAE) of 9.53, Root Mean Square Error (RMSE) of
14.67, and R-squared (R²) value of 0.90, LSTM
demonstrated the best ability to capture the
temporal dependencies in the data, such as
seasonal fluctuations, demand spikes during
promotions, and long-term sales patterns. Its
recurrent structure allowed it to remember and
utilize information from past time steps, making it
the most suitable model for handling time series
data like retail demand.
Chart 1: Comparison of different machine learning Model performance
Following LSTM, Gradient Boosting (GB) emerged
as another strong contender, with a MAE of 10.68,
RMSE of 15.89, and R² of 0.87. The GB model
performed particularly well due to its iterative
nature, where each subsequent model corrected
the errors of the previous one. This made GB highly
effective in capturing complex, non-linear
relationships between input features and demand.
15.3
4
12.7
8
11.22
10.6
8
9.53
20.5
7
18.1
1
16.4
5
15.8
9
14.6
7
0.71
0.75
0.83
0.87
0.9
L I N E A R
R E G R E S S I O N
D E C I S I O N T R E E
R E G R E S S O R
R A N D O M F O R E S T
R E G R E S S O R
G R A D I E N T
B O O S T I N G
L S T M
M A C H I N E L E A R N I N G M O D E L P E R F O R M A N C E
MAE
RMSE
R²
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE09
77
https://www.theamericanjournals.com/index.php/tajet
Although it did not match LSTM's performance in
handling temporal patterns, GB provided excellent
accuracy, making it one of the top-performing
models in this study.
The Random Forest Regressor (RFR) also
performed impressively, achieving a MAE of 11.22,
RMSE of 16.45, and R² of 0.83. Random Forest's
ability to reduce overfitting by averaging the
predictions of multiple decision trees resulted in a
robust model that was particularly effective in
handling the inherent variability in retail demand.
The model was able to accommodate seasonality
and other intricate demand patterns, though it was
slightly less accurate than Gradient Boosting and
LSTM.
In contrast, Decision Tree Regressor (DTR), while
showing an improvement over the baseline model,
struggled with overfitting. It achieved a MAE of
12.78, RMSE of 18.11, and an R² of 0.75. While the
decision tree model captured non-linear
relationships better than Linear Regression, its
performance was hindered by the model's
sensitivity to small changes in the data, leading to
overfitting when applied to complex retail demand
patterns.
Lastly, Linear Regression (LR), which served as the
baseline model, performed the weakest with a
MAE of 15.34, RMSE of 20.57, and R² of 0.71. The
linear nature of this model limited its ability to
capture non-linear trends and interactions
between variables, making it less suitable for the
intricate dynamics of retail demand forecasting. It
struggled particularly with seasonality and
demand spikes, which require more sophisticated
models to forecast accurately.
In summary, LSTM emerged as the most effective
model due to its ability to model temporal
dependencies, followed by Gradient Boosting and
Random Forest, which both performed well in
handling non-linear relationships and seasonal
demand. Linear Regression, on the other hand, was
the least effective, highlighting the importance of
using more advanced models for accurate demand
forecasting in retail environments.
CONCLUSION
In this study, we explored the application of
various machine learning models for retail demand
forecasting, comparing their performance based
on accuracy, error rates, and their ability to handle
complex patterns such as seasonality and demand
spikes. The models evaluated included Linear
Regression (LR), Decision Tree Regressor (DTR),
Random Forest Regressor (RFR), Gradient
Boosting (GB), and Long Short-Term Memory
(LSTM). Among these, the LSTM model emerged as
the top performer due to its ability to capture long-
term dependencies and temporal patterns in the
data, making it particularly suitable for time series
forecasting in retail. LSTM's recurrent structure
allowed it to handle fluctuations caused by
holidays, promotions, and other seasonal factors
with superior accuracy. Gradient Boosting and
Random Forest also delivered strong results,
effectively managing non-linear relationships and
providing robust forecasts, albeit with slightly less
precision than LSTM. Both models demonstrated
their suitability for retail demand forecasting by
reducing overfitting and capturing intricate
patterns in the data.
In contrast, Linear Regression and Decision Tree
Regressor struggled with the complexities of retail
data. Linear Regression, while easy to interpret,
lacked the sophistication needed to account for
non-linear relationships and seasonal trends,
making it the least effective model. The Decision
Tree Regressor, although an improvement over
Linear Regression, faced challenges with
overfitting, which affected its performance on
unseen data. Overall, this study highlights the
importance of selecting advanced models like
LSTM for retail demand forecasting, particularly in
environments
characterized
by
temporal
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE09
78
https://www.theamericanjournals.com/index.php/tajet
dependencies and demand volatility. The findings
suggest that businesses aiming to improve their
demand forecasting capabilities should consider
deploying LSTM or Gradient Boosting models for
more accurate and reliable predictions. Future
research could further optimize these models,
explore additional features, and evaluate their
performance across different retail segments to
refine forecasting strategies.
Acknowledgement: All the author contributed
equally
REFERENCE
1.
Shahid, R., Mozumder, M. A. S., Sweet, M. M. R.,
Hasan, M., Alam, M., Rahman, M. A., ... & Islam,
M. R. (2024). Predicting Customer Loyalty in
the Airline Industry: A Machine Learning
Approach Integrating Sentiment Analysis and
User Experience. International Journal on
Computational Engineering, 1(2), 50-54.
2.
Mozumder, M. A. S., Sweet, M. M. R., Nabi, N.,
Tusher, M. I., Modak, C., Hasan, M., ... & Prabha,
M. (2024). Revolutionizing Organizational
Decision-Making for Banking Sector: A
Machine Learning Approach with CNNs in
Business Intelligence and Management.
Journal of Business and Management Studies,
6(3), 111-118.
3.
Ferdus, M. Z., Anjum, N., Nguyen, T. N., Jisan, A.
H., & Raju, M. A. H. (2024). The Influence of
Social Media on Stock Market: A Transformer-
Based Stock Price Forecasting with External
Factors. Journal of Computer Science and
Technology Studies, 6(1), 189-194
4.
Bajari, P., Chernozhukov, V., Hortacsu, A., &
Suzuki, J. (2019). The Impact of Big Data on
Firm Performance. Journal of Applied
Econometrics, 34(4), 725-746.
5.
Breiman, L. (2001). Random forests. Machine
Learning, 45(1), 5-32.
6.
Breiman, L. (2017). Classification and
regression trees. Routledge.
7.
Brownlee, J. (2018). Deep Learning for Time
Series Forecasting: Predict the Future with
MLPs, CNNs, and LSTMs in Python. Machine
Learning Mastery.
8.
Chen, T., & Guestrin, C. (2016). XGBoost: A
scalable tree boosting system. In Proceedings
of the 22nd ACM SIGKDD International
Conference on Knowledge Discovery and Data
Mining (pp. 785-794).
9.
Box, G. E. P., & Jenkins, G. M. (1970). Time series
analysis: Forecasting and control. San
Francisco: Holden-Day.
10.
Chopra, S., & Meindl, P. (2016). Supply Chain
Management:
Strategy,
Planning,
and
Operation. Pearson.
11.
Cortez, P., Cerdeira, A., Almeida, F., Matos, T., &
Reis, J. (2021). Modeling wine preferences by
data mining from physicochemical properties.
Decision Support Systems, 47(4), 547-553.
12.
Fildes, R., Ma, S., & Kolassa, S. (2019). Retail
forecasting:
Research
and
practice.
International Journal of Forecasting, 35(2),
645-655.
13.
Friedman, J. H. (2001). Greedy function
approximation: A gradient boosting machine.
Annals of Statistics, 29(5), 1189-1232.
14.
Hastie, T., Tibshirani, R., & Friedman, J. (2009).
The Elements of Statistical Learning: Data
Mining, Inference, and Prediction. Springer.
15.
Hochreiter, S., & Schmidhuber, J. (1997). Long
short-term memory. Neural Computation,
9(8), 1735-1780.
16.
Hyndman, R. J., & Athanasopoulos, G. (2018).
Forecasting: Principles and Practice. OTexts.
17.
Hyndman, R. J., Bergmeir, C., Caceres, G., &
O'Hara-Wild, M. (2021). Forecasting with
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE09
79
https://www.theamericanjournals.com/index.php/tajet
Machine Learning. Journal of Business
Research.
18.
Keerthi, S. S., & Lin, C. J. (2020). Decision trees
for retail demand forecasting: A case study.
European Journal of Operational Research,
185(2), 789-802.
19.
Liaw, A., & Wiener, M. (2002). Classification
and regression by randomForest. R news, 2(3),
18-22.
20.
Livieris, I. E., Drakopoulou, K., & Kiriakidou, N.
(2020). Demand forecasting using machine
learning models: A case study. Expert Systems
with Applications, 145, 113089.
21.
Makridakis, S., Spiliotis, E., & Assimakopoulos,
V. (2018). Statistical and machine learning
forecasting methods: Concerns and ways
forward. PLOS ONE, 13(3), e0194889.
22.
Natekin, A., & Knoll, A. (2013). Gradient
boosting machines, a tutorial. Frontiers in
Neurorobotics, 7, 21.
23.
Farabi, S. F., Prabha, M., Alam, M., Hossan, M. Z.,
Arif, M., Islam, M. R., ... & Biswas, M. Z. A. (2024).
Enhancing Credit Card Fraud Detection: A
Comprehensive Study of Machine Learning
Algorithms and Performance Evaluation.
Journal of Business and Management Studies,
6(3), 252-259.
24.
Mozumder, M. A. S., Sweet, M. M. R., Nabi, N.,
Tusher, M. I., Modak, C., Hasan, M., ... & Prabha,
M. (2024). Revolutionizing Organizational
Decision-Making for Banking Sector: A
Machine Learning Approach with CNNs in
Business Intelligence and Management.
Journal of Business and Management Studies,
6(3), 111-118.
25.
Bhuiyan, M. S., Chowdhury, I. K., Haider, M.,
Jisan, A. H., Jewel, R. M., Shahid, R., ... & Siddiqua,
C. U. (2024). Advancements in early detection
of lung cancer in public health: a
comprehensive study utilizing machine
learning algorithms and predictive models.
Journal of Computer Science and Technology
Studies, 6(1), 113-121.
26.
Nabi, N., Tusher, M. I., Modak, C., Hasan, M., ... &
Prabha,
M.
(2024).
Revolutionizing
Organizational Decision-Making for Banking
Sector: A Machine Learning Approach with
CNNs
in
Business
Intelligence
and
Management. Journal of Business and
Management Studies, 6(3), 111-118.
27.
Rahman, M. A., Modak, C., Mozumder, M. A. S.,
Miah, M. N. I., Hasan, M., Sweet, M. M. R., ... &
Alam, M. (2024). Advancements in Retail Price
Optimization: Leveraging Machine Learning
Models for Profitability and Competitiveness.
Journal of Business and Management Studies,
6(3), 103-110.
28.
Shahid, R., Mozumder, M. A. S., Sweet, M. M. R.,
Hasan, M., Alam, M., Rahman, M. A., ... & Islam,
M. R. (2024). Predicting Customer Loyalty in
the Airline Industry: A Machine Learning
Approach Integrating Sentiment Analysis and
User Experience. International Journal on
Computational Engineering, 1(2), 50-54.
29.
Modak, C., Ghosh, S. K., Sarkar, M. A. I., Sharif, M.
K., Arif, M., Bhuiyan, M., ... & Devi, S. (2024).
Machine Learning Model in Digital Marketing
Strategies for Customer Behavior: Harnessing
CNNs for Enhanced Customer Satisfaction and
Strategic
Decision-Making.
Journal
of
Economics, Finance and Accounting Studies,
6(3), 178-186.
30.
Mozumder, M. A. S., Nguyen, T. N., Devi, S., Arif,
M., Ahmed, M. P., Ahmed, E., ... & Uddin, A.
(2024). Enhancing Customer Satisfaction
Analysis Using Advanced Machine Learning
Techniques in Fintech Industry. Journal of
Computer Science and Technology Studies,
6(3), 35-41.
THE USA JOURNALS
THE AMERICAN JOURNAL OF ENGINEERING AND TECHNOLOGY (ISSN
–
2689-0984)
VOLUME 06 ISSUE09
80
https://www.theamericanjournals.com/index.php/tajet
31.
Arif, M., Hasan, M., Al Shiam, S. A., Ahmed, M. P.,
Tusher, M. I., Hossan, M. Z., ... & Imam, T.
(2024). Predicting Customer Sentiment in
Social Media Interactions: Analyzing Amazon
Help Twitter Conversations Using Machine
Learning. International Journal of Advanced
Science Computing and Engineering, 6(2), 52-
56.
32.
Md Al-Imran, Salma Akter, Md Abu Sufian
Mozumder, Rowsan Jahan Bhuiyan, Md Al Rafi,
Md Shahriar Mahmud Bhuiyan, Gourab
Nicholas Rodrigues, Md Nazmul Hossain Mir,
Md Amit Hasan, Ashim Chandra Das, & Md.
Emran
Hossen.
(2024).
EVALUATING
MACHINE LEARNING ALGORITHMS FOR
BREAST CANCER DETECTION: A STUDY ON
ACCURACY AND PREDICTIVE PERFORMANCE.
The American Journal of Engineering and
Technology,
6(09),
22
–
33.
https://doi.org/10.37547/tajet/Volume06Iss
ue09-04
33.
Md Abu Sufian Mozumder, Fuad Mahmud, Md
Shujan Shak, Nasrin Sultana, Gourab Nicholas
Rodrigues, Md Al Rafi, Md Zahidur Rahman
Farazi, Md Razaul Karim, Md. Sayham Khan, &
Md Shahriar Mahmud Bhuiyan. (2024).
Optimizing Customer Segmentation in the
Banking Sector: A Comparative Analysis of
Machine Learning Algorithms. Journal of
Computer Science and Technology Studies,
6(4),
01
–
07.
https://doi.org/10.32996/jcsts.2024.6.4.1
