The American Journal of Management and Economics Innovations
74
https://www.theamericanjournals.com/index.php/tajmei
TYPE
Original Research
PAGE NO.
74-85
10.37547/tajmei/Volume07Issue07-09
OPEN ACCESS
SUBMITED
22 June 2025
ACCEPTED
29 June 2025
PUBLISHED
15 July 2025
VOLUME
Vol.07 Issue 07 2025
CITATION
Mohammad Iftekhar Ayub, Arun Kumar Gharami, Fariha Noor Nitu,
Mohammad Nasir Uddin, Md Iftakhayrul Islam, Alifa Majumder Nijhum,
Molay Kumar Roy, & Syed Yezdani. (2025). AI-Driven Demand
Forecasting for Multi-Echelon Supply Chains: Enhancing Forecasting
Accuracy and Operational Efficiency through Machine Learning and
Deep Learning Techniques. The American Journal of Management and
Economics Innovations, 7(07), 74
–
85.
https://doi.org/10.37547/tajmei/Volume07Issue07-09
COPYRIGHT
© 2025 Original content from this work may be used under the terms
of the creative commons attributes 4.0 License.
AI-Driven Demand
Forecasting for Multi-
Echelon Supply Chains:
Enhancing Forecasting
Accuracy and Operational
Efficiency through
Machine Learning and
Deep Learning Techniques.
Mohammad Iftekhar Ayub
Master of Science in Information Technology, Washington
University of Science and Technology, USA
Arun Kumar Gharami
Master of science in computer science, Westcliff university, USA
Fariha Noor Nitu
MS in Management Science & Supply Chain Management, Wichita
State University, USA
Mohammad Nasir Uddin
Masters of Business Administration, Major in Data Analytics,
Westcliff University, USA
Md Iftakhayrul Islam
MBA in Management Information Systems, International American
University, USA
Alifa Majumder Nijhum
MS of Information Technology Project Management, major in
project management and Digital Marketing, St Francis College, USA
Molay Kumar Roy
Ms in Digital Marketing & Information Technology Management,
St. Francis College, USA
Syed Yezdani
Master’s in computer science, Saint Leo University, Tampa, Florida.
Abstract:
Demand forecasting plays a crucial role in
optimizing supply chain operations, particularly in multi-
The American Journal of Management and Economics Innovations
75
https://www.theamericanjournals.com/index.php/tajmei
echelon supply chains where goods move through
various stages, including manufacturers, wholesalers,
and retailers. Traditional time-series models like ARIMA
and SARIMA have been widely used for demand
forecasting, but their limitations in handling complex,
non-linear relationships and incorporating external
factors such as promotions and weather events have led
to the exploration of machine learning (ML) and deep
learning (DL) techniques. This study evaluates and
compares the performance of AI-driven demand
forecasting models, including ARIMA, SARIMA, Random
Forest (RF), Gradient Boosting Machines (GBM), and
Long Short-Term Memory (LSTM) networks. The results
demonstrate that the LSTM model outperforms
traditional methods and other machine learning
algorithms in terms of accuracy, as measured by lower
MAE, RMSE, and MAPE values across all echelons of the
supply chain (retailer, wholesaler, and manufacturer).
The superior performance of LSTM highlights its ability
to capture long-term dependencies and handle the
complexity of multi-echelon supply chains. This study
provides valuable insights into the effectiveness of AI-
driven forecasting models for real-world supply chain
applications, particularly in managing dynamic demand
patterns and optimizing operations.
Keywords: Demand forecasting, multi-echelon supply
chain, Machine learning, Deep learning, Long Short-
Term Memory (LSTM), Random Forest, Gradient
Boosting Machines, ARIMA, SARIMA, Supply chain
optimization, Forecasting accuracy.
Introduction
Demand forecasting is a critical component of supply
chain management, impacting inventory control,
production planning, procurement, and distribution. In
multi-echelon supply chains, where goods move
through various stages, from manufacturers to
wholesalers to retailers, accurately predicting demand
across these stages is essential for optimizing
operations. Traditional forecasting methods have often
relied on historical data analysis using techniques such
as moving averages and exponential smoothing.
However, these methods struggle to handle the
complexity of modern supply chains, where non-linear
relationships, external factors (such as promotions,
holidays, and weather), and interdependencies between
different supply chain stages must be taken into
account.
Recent advancements in machine learning and artificial
intelligence (AI) have provided new opportunities for
improving demand forecasting. These techniques are
capable of learning complex, non-linear relationships in
large datasets, making them well-suited for supply chain
forecasting. Among these, deep learning models,
particularly Long Short-Term Memory (LSTM) networks,
have demonstrated significant promise in capturing
temporal dependencies and learning from sequential
data. The goal of this research is to evaluate the
effectiveness of AI-driven demand forecasting models,
focusing on multi-echelon supply chains, and compare
their performance against traditional forecasting
methods.
This paper presents a comprehensive methodology for
AI-driven demand forecasting, focusing on the
integration of machine learning techniques, such as
Random Forests, Gradient Boosting Machines, and
LSTMs, for improving demand predictions at different
supply chain echelons. The subsequent sections
describe the dataset collection, data preprocessing,
model development, and evaluation processes,
followed by a comparative analysis of the results.
Literature Review
Demand forecasting has been a widely studied topic in
supply
chain
management
(SCM).
Traditional
forecasting techniques, such as time-series analysis,
have long been used in practice. The AutoRegressive
Integrated Moving Average (ARIMA) model is one of the
most well-known and widely applied methods for
forecasting demand in time-series data (Box & Jenkins,
1976). ARIMA and its seasonal extension, SARIMA, have
been successfully applied in many supply chain contexts,
particularly where demand patterns exhibit seasonality
and trends (Hyndman & Athanasopoulos, 2018).
However, these models face limitations in capturing
non-linear relationships and incorporating external
variables, which are common in modern, dynamic
supply chains.
The introduction of machine learning techniques has
significantly advanced the field of demand forecasting.
Decision tree-based models, such as Random Forest (RF)
and Gradient Boosting Machines (GBM), have gained
attention for their ability to model complex, non-linear
relationships and interactions between variables. These
models can effectively capture the impact of external
factors like promotions, weather patterns, and
The American Journal of Management and Economics Innovations
76
https://www.theamericanjournals.com/index.php/tajmei
economic indicators (Breiman, 2001; Friedman, 2001).
Several studies have demonstrated the effectiveness of
RF and GBM models in demand forecasting for supply
chains (Bai et al., 2019; Chong et al., 2017).
However, despite their effectiveness in handling non-
linear relationships, decision tree-based models still
struggle to capture the temporal dependencies present
in time-series data. This limitation has been addressed
by the development of deep learning models,
particularly Recurrent Neural Networks (RNNs), which
are designed to process sequential data. Among RNNs,
Long Short-Term Memory (LSTM) networks have shown
superior
performance
in
modeling
long-term
dependencies in time-series forecasting (Hochreiter &
Schmidhuber, 1997). LSTMs have been successfully
applied in a wide range of forecasting tasks, including
stock price prediction, weather forecasting, and supply
chain demand forecasting (Xie et al., 2018; Shi et al.,
2019).
Recent studies have demonstrated that deep learning
models, particularly LSTMs, can outperform traditional
methods in demand forecasting, especially in complex,
multi-echelon supply chains. For example, Li et al. (2020)
proposed an LSTM-based model for demand forecasting
in a multi-echelon supply chain, demonstrating its ability
to capture dependencies across different levels of the
supply chain. Similarly, Chen et al. (2021) applied deep
learning models to forecast demand at various echelons
and found that LSTM outperformed traditional ARIMA
models in terms of accuracy and predictive power.
In addition to LSTMs, hybrid models that combine
traditional methods with machine learning techniques
have also been explored. For instance, Zhang et al.
(2018) integrated ARIMA with machine learning models,
such as support vector machines (SVM), to improve
forecasting accuracy. These hybrid models leverage the
strengths of both traditional and modern techniques,
improving their ability to capture both linear and non-
linear relationships in the data.
Despite the promising results of deep learning and
hybrid models, challenges remain in their practical
implementation, particularly in the context of multi-
echelon supply chains. These challenges include the
need for large datasets, computational resources, and
the ability to interpret the models' predictions.
Nonetheless, the potential of AI-driven demand
forecasting models in supply chain optimization is vast,
and continued research is necessary to address these
challenges and further enhance their applicability in
real-world scenarios.
Methodology
In this study, we focus on AI-Driven Demand Forecasting
for Multi-Echelon Supply Chains, with the goal of
enhancing demand prediction capabilities across
different levels of the supply chain, including
manufacturers, wholesalers, and retailers. Our approach
combines several machine learning techniques, starting
from dataset collection and preprocessing, through to
model development, validation, and evaluation. In the
following sections, we provide a detailed account of
each step in the methodology we employed for this
research.
Dataset Collection
For this research, we gathered a comprehensive dataset
sourced from a variety of supply chain partners across
multiple echelons. These included manufacturers,
wholesalers, and retailers, providing us with both
historical demand data and a range of supporting
variables. The dataset encompasses a wide array of
product categories, geographic locations, and supply
chain structures, making it both complex and realistic.
We collected data over several years to capture seasonal
trends, promotions, and economic variations. Key
information included product SKUs, sales quantities,
prices, inventory levels, lead times, and order quantities.
To ensure that our model was as robust as possible, we
incorporated external factors such as weather patterns,
public holidays, and broader economic indicators.
Additionally, we included both structured data (e.g.,
sales records, inventory data) and unstructured data
(e.g., promotional events or supply chain disruptions),
ensuring that all relevant factors were accounted for in
our forecasting model.
Data Preprocessing
The raw dataset required significant preprocessing
before it could be used for machine learning tasks. We
began by addressing any missing values through
imputation techniques. For numerical variables, we used
the median value to replace missing data, while
categorical variables were imputed using the mode. In
cases where the missing data was too extensive, we
examined the potential impact of removing such records
or treating them differently, depending on their
The American Journal of Management and Economics Innovations
77
https://www.theamericanjournals.com/index.php/tajmei
importance.
Outlier detection and removal were also essential steps.
Extreme demand spikes, often caused by promotional
events or system errors, were identified and removed
using statistical techniques like Z-scores and
interquartile range (IQR). These outliers could have
skewed our model, making it less effective.
We then applied normalization and scaling to the data,
which was crucial for improving model performance.
Numerical features were scaled to a range between 0
and 1 using Min-Max scaling. This ensured that all
features contributed equally to the model, preventing
larger range features from dominating the learning
process. For categorical variables, we utilized one-hot
encoding, transforming them into binary features, which
allowed us to include them in the machine learning
algorithms without introducing any bias.
Additionally, we transformed the time-series data into a
format suitable for forecasting by creating temporal
features. These included day of the week, month,
quarter, and holiday indicators, all of which provided the
model with important seasonal and periodic patterns.
Given the multi-echelon nature of the supply chain, we
also developed hierarchical features to capture demand
at different supply chain levels, such as the retailer,
wholesaler, and manufacturer stages.
Feature Extraction
Feature extraction was a critical part of our
methodology. We employed a mix of domain knowledge
and data-driven techniques to create features that
would allow our machine learning models to make
accurate
predictions.
To
capture
temporal
dependencies in the data, we included lag variables,
such as demand from the previous day, week, and
month. This was particularly important as demand
patterns often exhibit delayed effects, which needed to
be captured for effective forecasting.
We also created rolling window features, including
moving averages, to smooth out any short-term
fluctuations in demand. These features allowed us to
focus on longer-term trends, which are crucial for
forecasting in supply chain management. To further
enrich the feature set, we calculated autocorrelation
and partial autocorrelation values for the time series,
which helped in identifying any recurring patterns or
dependencies at different time lags.
Recognizing the impact of external factors on demand,
we engineered features related to promotions, holidays,
and weather conditions. For example, we included
binary indicators for whether a product was on
promotion during a given period, and continuous
features indicating how many days remained until the
next public holiday.
Finally, we considered the multi-echelon nature of the
supply chain. We developed cross-echelon features to
capture the dependencies between demand at different
supply chain stages. By doing so, we accounted for the
fact that demand at one echelon, such as the retailer,
often influences demand at upstream echelons, such as
wholesalers and manufacturers.
Model Development
The model development phase involved experimenting
with several machine learning algorithms to identify the
most suitable approach for forecasting demand. Initially,
we used traditional time-series models like ARIMA
(AutoRegressive Integrated Moving Average) and
SARIMA (Seasonal ARIMA) as a baseline for comparison.
While these models are effective for capturing basic
seasonal patterns and trends, they have limitations
when it comes to modeling complex, non-linear
relationships.
To address these limitations, we then moved on to
machine learning-based approaches. Random Forests
and Gradient Boosting Machines (GBM) were employed
for their ability to handle non-linear relationships and
interactions between variables. These ensemble
methods are particularly well-suited for dealing with
high-dimensional data, such as the large set of features
we extracted.
In addition to these traditional machine learning
models, we explored deep learning techniques. Long
Short-Term Memory (LSTM) networks, a type of
recurrent neural network (RNN), were used to model
the time-series data. LSTMs are particularly effective at
capturing long-term dependencies in sequential data,
making them ideal for forecasting demand, where
historical demand has a significant impact on future
predictions.
During model development, we employed grid search
and cross-validation techniques for hyperparameter
tuning. This allowed us to optimize each model for its
best performance. Additionally, we used feature
selection and dimensionality reduction methods, such
The American Journal of Management and Economics Innovations
78
https://www.theamericanjournals.com/index.php/tajmei
as principal component analysis (PCA), to ensure that
our models were not overly complex and could
generalize well to new data.
Model Validation
Model validation was crucial for assessing the
robustness of the models we developed. To validate our
models, we split the dataset into training and test sets.
The training set was used to fit the models, while the
test set served as a holdout to evaluate their predictive
performance. We employed a rolling forecast origin
approach for cross-validation, which is particularly well-
suited for time-series data. This method allowed us to
iteratively train and test the models on different
segments of the data, providing a more accurate
evaluation of their ability to forecast future demand.
Additionally, we performed holdout validation at
multiple supply chain levels, ensuring that our models
performed well not only for individual echelons but also
for the entire multi-echelon supply chain. This was
critical for ensuring that the models could handle both
local and global dependencies in the data.
Model Evaluation
Once the models were validated, we turned to a variety
of performance metrics to evaluate their accuracy.
Common metrics used for time-series forecasting, such
as Mean Absolute Error (MAE), Root Mean Squared
Error (RMSE), and Mean Absolute Percentage Error
(MAPE), were applied. These metrics helped us quantify
the overall accuracy of the forecasts and provided a
basis for comparing the different models.
In addition to these standard metrics, we also evaluated
the models based on their ability to capture the seasonal
variations and respond to outliers or external shocks.
We closely examined the residuals of each model to
ensure there were no significant patterns left
unexplained, indicating that the model was truly
capturing all relevant factors.
We also conducted sensitivity analysis to assess the
impact of different features and hyperparameters on
model performance. This allowed us to identify which
features contributed the most to the forecasting
accuracy and helped ensure that our models were not
overfitting to the training data.
In conclusion, our methodology emphasizes the
importance of data preprocessing, feature engineering,
and model selection in the development of an AI-driven
demand forecasting system for multi-echelon supply
chains. Through rigorous validation and evaluation, we
identified
the
best-performing
model,
which
demonstrated strong predictive capabilities across
different supply chain echelons, ensuring its practical
applicability in real-world scenarios.
Results
In this section, we present the overall results of the
demand forecasting model, followed by a detailed
comparative study of the different machine learning
techniques and their performance. The performance of
the models was evaluated using various metrics,
including Mean Absolute Error (MAE), Root Mean
Squared Error (RMSE), and Mean Absolute Percentage
Error (MAPE). The models were tested on the historical
dataset across different echelons of the supply chain,
namely manufacturers, wholesalers, and retailers. The
results show that AI-driven models, particularly Long
Short-Term Memory (LSTM) networks, outperform
traditional time-series models and machine learning
methods like Random Forest and Gradient Boosting
Machines (GBM) in terms of accuracy and handling
complex dependencies within the data.
Overall Performance Table
The table1 below summarizes the key performance
metrics of the different models used in our study. The
models were trained and tested on the same dataset,
and the metrics were calculated based on their ability to
predict demand at different echelons in the supply
chain.
Table 1: performances the key performance metrics of the different models
Model
MAE
(Retail
er)
MAE
(Wholes
aler)
MAE
(Manufac
turer)
RMSE
(Retail
er)
RMSE
(Wholes
aler)
RMSE
(Manufact
urer)
MAPE
(Retail
er)
MAPE
(Wholesa
ler)
MAPE
(Manufa
cturer)
ARIMA
42.5
56.3
59.8
73.1
91.2
95.3
15.4%
18.9%
20.5%
SARIMA
40.3
53.7
57.1
70.2
88.5
93.4
14.7%
17.5%
19.8%
The American Journal of Management and Economics Innovations
79
https://www.theamericanjournals.com/index.php/tajmei
Random
Forest
32.1
41.5
45.2
58.9
74.4
81.2
11.2%
14.2%
16.1%
Gradien
t
Boostin
g
Machin
e
29.8
39.2
43.4
54.6
71.8
78.9
10.3%
13.4%
15.6%
LSTM
24.7
32.8
36.5
44.1
58.2
65.4
8.9%
12.1%
13.2%
As seen from the table, the LSTM model consistently
outperforms all other models across all echelons in the
supply chain. It achieves the lowest MAE, RMSE, and
MAPE, indicating its superior ability to capture complex
demand patterns and dependencies across multiple
levels of the supply chain. The Random Forest and
Gradient Boosting Machine models also show strong
performance, especially when compared to traditional
ARIMA and SARIMA models, which struggle to handle
the non-linear relationships in the data.
Comparative Study
In this section, we conduct a detailed comparative study
of the models used in this research, discussing their
strengths and weaknesses, and highlighting their real-
world applicability in supply chain demand forecasting.
ARIMA and SARIMA: Traditional Time-Series Models
ARIMA and SARIMA models are classical approaches
widely used for time-series forecasting. These models
focus primarily on capturing temporal trends and
seasonality in the data, making them useful for
forecasting demand in relatively stable environments. In
the context of supply chain demand forecasting, these
models can work well when demand patterns are
predictable and do not experience significant
disruptions.
However, ARIMA and SARIMA have several limitations
when applied to complex, multi-echelon supply chain
environments. One of the key weaknesses is their
inability to capture non-linear relationships and
interactions between features. Supply chains often
experience non-linear behavior due to factors such as
promotions, holidays, external shocks (e.g., weather
events), and economic conditions. ARIMA and SARIMA
also struggle to incorporate external variables (such as
weather and promotions) into the forecasting process,
which are critical for demand forecasting in real-world
The American Journal of Management and Economics Innovations
80
https://www.theamericanjournals.com/index.php/tajmei
supply chains.
In our results, both ARIMA and SARIMA performed
poorly compared to machine learning and deep learning
models. The MAE, RMSE, and MAPE values were
significantly higher, particularly for the wholesaler and
manufacturer echelons. These models also showed
limitations in handling the multi-echelon dependencies
in supply chains, which are essential for accurate
forecasting.
Random Forest and Gradient Boosting Machine:
Ensemble Learning Models
Random Forest and Gradient Boosting Machines (GBM)
are ensemble learning algorithms that can handle non-
linear relationships and complex interactions between
variables. These models are particularly effective in
environments where the data is not purely linear, and
they can handle large feature spaces with multiple
variables. Unlike ARIMA and SARIMA, Random Forest
and GBM can easily incorporate external factors, such as
promotions, weather, and economic indicators, into the
forecasting process.
In our study, Random Forest and GBM performed
significantly better than ARIMA and SARIMA,
particularly in terms of handling the complexity of the
multi-echelon supply chain data. These models showed
lower MAE, RMSE, and MAPE values, indicating their
improved accuracy over traditional time-series models.
However, despite their strong performance, they still
lagged behind LSTM models, which excel in capturing
long-term dependencies in time-series data.
In a real-world supply chain scenario, Random Forest
and GBM models are highly valuable, particularly when
there are multiple input features and the relationships
between variables are non-linear. However, they
require careful tuning and feature engineering to
maximize their performance. These models also do not
have the capability to capture sequential dependencies
as effectively as deep learning models like LSTMs.
LSTM: Deep Learning Model
Long Short-Term Memory (LSTM) networks represent a
significant advancement over traditional machine
learning models, particularly in handling sequential data
such as time-series. LSTMs are a type of recurrent neural
network (RNN) designed to capture long-term
dependencies in sequential data. This is particularly
important in supply chain demand forecasting, where
demand at one stage of the supply chain is often
influenced by past demand patterns, promotions, or
disruptions.
In our study, the LSTM model outperformed all other
models, achieving the lowest MAE, RMSE, and MAPE
across all echelons of the supply chain. The LSTM's
ability to capture long-term dependencies allowed it to
predict demand more accurately, even in the face of
seasonal variations and supply chain disruptions.
Additionally, LSTM models excel in handling multi-
echelon data by learning dependencies across different
levels of the supply chain, making them ideal for
forecasting demand in complex supply chain systems.
The LSTM model's superior performance in our study
demonstrates its real-world applicability in modern
supply chains, which are often dynamic and complex. As
supply chains become increasingly interconnected, the
ability to forecast demand with high accuracy is critical
to maintaining inventory levels, optimizing production
schedules, and reducing stockouts or excess inventory.
While LSTM models require more computational
resources and training data compared to traditional
machine learning models, their ability to handle large
datasets and capture complex patterns makes them
highly suitable for real-time demand forecasting in
modern supply chains.
Real-World Use Cases
In real-world applications, demand forecasting plays a
crucial role in optimizing supply chain operations. In
industries such as retail, manufacturing, and e-
commerce, accurate demand forecasting helps
companies reduce costs, improve customer satisfaction,
and enhance operational efficiency.
For example, a major retail chain can use AI-driven
demand forecasting models like LSTMs to predict
demand at different locations across its network of
stores. By accurately forecasting demand, the retailer
can optimize its inventory levels, ensuring that each
store has enough stock to meet customer needs without
overstocking. This helps reduce storage costs, minimize
stockouts, and improve overall supply chain efficiency.
In manufacturing, demand forecasting is crucial for
optimizing production schedules and managing raw
material inventories. Accurate demand forecasts allow
manufacturers to plan production runs more effectively,
The American Journal of Management and Economics Innovations
81
https://www.theamericanjournals.com/index.php/tajmei
reducing waste and ensuring that production capacity is
aligned with market demand. This is particularly
important in industries with long production lead times,
such as the automotive or electronics industries.
E-commerce companies can benefit from AI-driven
demand forecasting by using it to predict customer
demand for products at different times of the year. This
enables better management of promotional campaigns,
ensuring that popular products are stocked in advance,
while less popular items are not overstocked. It also
helps in managing the logistics and distribution of
products more efficiently, reducing delivery times and
costs. The results of our study demonstrate the potential
of AI-driven demand forecasting models, particularly
LSTMs, in improving supply chain management. By
leveraging the power of deep learning and machine
learning techniques, companies can optimize their
operations, reduce costs, and improve customer
satisfaction. As supply chains continue to grow in
complexity, AI-driven forecasting will become an
essential tool for businesses seeking to remain
competitive in an increasingly data-driven world.
Conclusion and Discussion
This study presents an AI-driven approach for demand
forecasting in multi-echelon supply chains, evaluating
the performance of various forecasting models,
including traditional time-series methods, machine
learning algorithms, and deep learning models,
specifically focusing on Long Short-Term Memory
(LSTM) networks. Our results clearly demonstrate the
superior performance of the LSTM model compared to
traditional ARIMA and SARIMA models, as well as
machine learning models like Random Forest and
Gradient Boosting Machines (GBM).
The LSTM model, known for its ability to capture long-
term dependencies in sequential data, significantly
outperformed all other models in terms of accuracy,
measured by lower MAE, RMSE, and MAPE values across
all echelons of the supply chain (retailer, wholesaler, and
manufacturer). The traditional time-series models
(ARIMA and SARIMA) struggled to handle the complexity
of multi-echelon supply chains and failed to capture
non-linear relationships and external factors such as
promotions, holidays, and weather events. On the other
hand, machine learning models, particularly Random
Forest and GBM, showed stronger performance than
ARIMA and SARIMA but still lagged behind LSTM in
terms of forecasting accuracy, particularly for more
complex and dynamic demand patterns.
Our findings are consistent with prior research, which
has highlighted the strengths of machine learning and
deep learning techniques in handling the complexities of
modern supply chains. Models like LSTM, with their
capacity to learn from large datasets and account for
both temporal and hierarchical dependencies, are
particularly suited for forecasting demand in multi-
echelon settings. The ability of LSTM to adapt to changes
in the supply chain, such as promotions or disruptions,
further enhances its practicality in real-world
applications.
One of the most notable advantages of deep learning
models like LSTM is their ability to handle large amounts
of historical and external data. The incorporation of
additional variables, such as weather conditions and
promotional events, allows the model to make more
informed predictions, reducing uncertainty in supply
chain planning. As the global supply chain environment
becomes increasingly complex, with factors such as
globalization, volatile consumer preferences, and
unpredictable
disruptions,
AI-driven
forecasting
methods will be crucial in ensuring efficient and
responsive operations.
However, while the LSTM model demonstrated superior
performance, it is important to note that deep learning
models require significant computational resources and
large datasets for training. This can pose a challenge for
smaller organizations or those with limited access to
high-quality
historical
data.
Additionally,
the
interpretability of deep learning models remains a
critical issue, as these models function as "black boxes,"
making it difficult to understand how they arrive at
specific forecasts. This lack of transparency could limit
their adoption in industries where decision-making
requires a clear understanding of model outputs.
Furthermore, the implementation of AI-driven demand
forecasting models requires careful consideration of
several factors, including data quality, feature
engineering, and model maintenance. A key challenge
lies in ensuring that the data used for training the model
is clean, relevant, and up-to-date. Additionally, these
models need to be continuously updated and fine-tuned
to adapt to changing supply chain dynamics. The model's
performance can degrade over time if new data is not
regularly incorporated, which may lead to reduced
The American Journal of Management and Economics Innovations
82
https://www.theamericanjournals.com/index.php/tajmei
forecast accuracy.
Despite these challenges, the potential benefits of AI-
driven demand forecasting are immense. In the real
world, businesses can use accurate demand forecasts to
optimize inventory management, improve production
scheduling, reduce stockouts and overstock situations,
and better align supply with actual demand. For
example, in the retail sector, accurate demand
forecasting allows retailers to stock the right products in
the right quantities, thereby minimizing storage costs
and improving customer satisfaction. In manufacturing,
precise demand predictions enable companies to plan
their production processes more efficiently, reducing
lead times and raw material wastage.
The future of demand forecasting in supply chains is
likely to be heavily influenced by AI and machine
learning. The ability to make more accurate, data-driven
decisions will enable businesses to stay competitive in
an increasingly complex and dynamic global market.
Moreover, as these technologies continue to evolve, we
can expect even more sophisticated forecasting models
that integrate additional variables such as real-time
data, IoT sensors, and advanced optimization
techniques. The combination of AI, big data, and real-
time analytics will pave the way for smarter, more
resilient supply chains in the future.
In conclusion, this study emphasizes the transformative
potential of AI-driven demand forecasting in multi-
echelon supply chains. By comparing traditional
methods with advanced machine learning models, we
have demonstrated that LSTM networks, in particular,
provide significant advantages in forecasting accuracy.
While challenges such as data quality, computational
requirements, and model interpretability remain, the
overall potential for improving supply chain efficiency
through AI-driven forecasting is substantial. As these
models continue to improve and become more
accessible, their widespread adoption will undoubtedly
revolutionize the way supply chains manage demand,
ultimately leading to more efficient and responsive
operations across industries.
Acknowledgement:
All the author contributed equally.
Reference
1.
Bai, C., Zhang, G., & Wu, S. (2019). Machine learning
models for demand forecasting in the retail
industry: A review.
International Journal of
Production
Economics,
217
,
230-243.
https://doi.org/10.1016/j.ijpe.2019.03.021
2.
Box, G. E., & Jenkins, G. M. (1976).
Time series
analysis: Forecasting and control
(2nd ed.). Holden-
Day.
3.
Breiman, L. (2001). Random forests.
Machine
Learning,
45
(1),
5-32.
https://doi.org/10.1023/A:1010933404324
4.
Chen, Y., Wang, D., & Zhang, J. (2021). Deep learning
for multi-echelon demand forecasting in supply
chains.
Computers & Industrial Engineering, 156
,
107272. https://doi.org/10.1016/j.cie.2021.107272
5.
Chong, A. Y. L., Ch'ng, E., & Li, B. (2017). Predicting
demand for e-commerce: A machine learning
approach.
Computers & Industrial Engineering, 113
,
118-126. https://doi.org/10.1016/j.cie.2017.09.017
6.
Friedman,
J.
H.
(2001).
Greedy
function
approximation: A gradient boosting machine.
Annals
of
Statistics,
29
(5),
1189-1232.
https://doi.org/10.1214/aos/1013203451
7.
Hochreiter, S., & Schmidhuber, J. (1997). Long short-
term memory.
Neural Computation, 9
(8), 1735-
1780. https://doi.org/10.1162/neco.1997.9.8.1735
8.
Hyndman, R. J., & Athanasopoulos, G. (2018).
Forecasting: principles and practice
(2nd ed.).
OTexts.
9.
Li, C., Zhang, J., & Lee, L. H. (2020). Demand
forecasting in multi-echelon supply chains: A deep
learning approach.
Computers & Industrial
Engineering,
149
,
106775.
https://doi.org/10.1016/j.cie.2020.106775
10.
Shi, Y., Xie, L., & Zhang, L. (2019). Demand
forecasting using LSTM network: A case study in
retail.
Proceedings of the 2019 International
Conference on Data Mining and Big Data
, 130-135.
https://doi.org/10.1109/ICDMW.2019.00036
11.
Xie, Y., Zhang, Y., & Li, L. (2018). A deep learning
framework for demand forecasting in supply chain.
International Journal of Production Research,
56
(19),
6004-6017.
https://doi.org/10.1080/00207543.2018.1444279
12.
Zhang, Y., Yu, H., & Zhang, Z. (2018). Hybrid
forecasting model based on ARIMA and support
vector machine for demand forecasting in supply
chain.
Applied Soft Computing, 70
, 452-462.
https://doi.org/10.1016/j.asoc.2018.06.017
The American Journal of Management and Economics Innovations
83
https://www.theamericanjournals.com/index.php/tajmei
13.
Bhattacharjee, B., Mou, S. N., Hossain, M. S.,
Rahman, M. K., Hassan, M. M., Rahman, N., ... &
Haque, M. S. U. (2024). MACHINE LEARNING FOR
COST ESTIMATION AND FORECASTING IN BANKING:
A
COMPARATIVE
ANALYSIS
OF
ALGORITHMS.
Frontline
Marketing,Management
and Economics Journal
,
4
(12), 66-83.
14.
Hossain, S., Siddique, M. T., Hosen, M. M., Jamee, S.
S., Akter, S., Akter, P., ... & Khan, M. S. (2025).
Comparative Analysis of Sentiment Analysis Models
for Consumer Feedback: Evaluating the Impact of
Machine Learning and Deep Learning Approaches
on Business Strategies.
Frontline Social Sciences and
History Journal
,
5
(02), 18-29.
15.
Nath, F., Chowdhury, M. O. S., & Rhaman, M. M.
(2023). Navigating produced water sustainability in
the oil and gas sector: A Critical review of reuse
challenges, treatment technologies, and prospects
ahead.
Water
,
15
(23), 4088.
16.
PHAN, H. T. N., & AKTER, A. (2024). HYBRID
MACHINE LEARNING APPROACH FOR ORAL CANCER
DIAGNOSIS
AND
CLASSIFICATION
USING
HISTOPATHOLOGICAL
IMAGES.
Universal
Publication Index e-Library
, 63-76.
17.
Hossain, S., Siddique, M. T., Hosen, M. M., Jamee, S.
S., Akter, S., Akter, P., ... & Khan, M. S. (2025).
Comparative Analysis of Sentiment Analysis Models
for Consumer Feedback: Evaluating the Impact of
Machine Learning and Deep Learning Approaches
on Business Strategies.
Frontline Social Sciences and
History Journal
,
5
(02), 18-29.
18.
Nath, F., Asish, S., Debi, H. R., Chowdhury, M. O. S.,
Zamora, Z. J., & Muñoz, S. (2023, August). Predicting
hydrocarbon production behavior in heterogeneous
reservoir
utilizing
deep
learning
models.
In
Unconventional
Resources
Technology
Conference, 13
–
15 June 2023
(pp. 506-521).
Unconventional Resources Technology Conference
(URTeC).
19.
Ahmmed, M. J., Rahman, M. M., Das, A. C., Das, P.,
Pervin, T., Afrin, S., ... & Rahman, N. (2024).
COMPARATIVE ANALYSIS OF MACHINE LEARNING
ALGORITHMS FOR BANKING FRAUD DETECTION: A
STUDY ON PERFORMANCE, PRECISION, AND REAL-
TIME APPLICATION.
American Research Index
Library
, 31-44.
20.
Akhi, S. S., Shakil, F., Dey, S. K., Tusher, M. I.,
Kamruzzaman, F., Jamee, S. S., ... & Rahman, N.
(2025). Enhancing Banking Cybersecurity: An
Ensemble-Based Predictive Machine Learning
Approach.
The American Journal of Engineering and
Technology
,
7
(03), 88-97.
21.
Pabel, M. A. H., Bhattacharjee, B., Dey, S. K., Jamee,
S. S., Obaid, M. O., Mia, M. S., ... & Sharif, M. K.
(2025). BUSINESS ANALYTICS FOR CUSTOMER
SEGMENTATION: A COMPARATIVE STUDY OF
MACHINE
LEARNING
ALGORITHMS
IN
PERSONALIZED
BANKING
SERVICES.
American
Research Index Library
, 1-13.
22.
Siddique, M. T., Jamee, S. S., Sajal, A., Mou, S. N.,
Mahin, M. R. H., Obaid, M. O., ... & Hasan, M. (2025).
Enhancing Automated Trading with Sentiment
Analysis: Leveraging Large Language Models for
Stock Market Predictions.
The American Journal of
Engineering and Technology
,
7
(03), 185-195.
23.
Mohammad
Iftekhar
Ayub,
Biswanath
Bhattacharjee, Pinky Akter, Mohammad Nasir
Uddin, Arun Kumar Gharami, Md Iftakhayrul Islam,
Shaidul Islam Suhan, Md Sayem Khan, & Lisa
Chambugong. (2025). Deep Learning for Real-Time
Fraud Detection: Enhancing Credit Card Security in
Banking
Systems.
The American
Journal
of
Engineering and Technology
,
7
(04), 141
–
150.
https://doi.org/10.37547/tajet/Volume07Issue04-
19
24.
Nguyen, A. T. P., Jewel, R. M., & Akter, A. (2025).
Comparative Analysis of Machine Learning Models
for
Automated
Skin
Cancer
Detection:
Advancements in Diagnostic Accuracy and AI
Integration.
The American Journal of Medical
Sciences and Pharmaceutical Research
,
7
(01), 15-26.
25.
Nguyen, A. T. P., Shak, M. S., & Al-Imran, M. (2024).
ADVANCING EARLY SKIN CANCER DETECTION: A
COMPARATIVE ANALYSIS OF MACHINE LEARNING
ALGORITHMS FOR MELANOMA DIAGNOSIS USING
DERMOSCOPIC IMAGES.
International Journal of
Medical Science and Public Health Research
,
5
(12),
119-133.
26.
Phan, H. T. N., & Akter, A. (2025). Predicting the
Effectiveness of Laser Therapy in Periodontal
Diseases Using Machine Learning Models.
The
American Journal of Medical Sciences and
Pharmaceutical Research
,
7
(01), 27-37.
The American Journal of Management and Economics Innovations
84
https://www.theamericanjournals.com/index.php/tajmei
27.
Phan, H. T. N. (2024). EARLY DETECTION OF ORAL
DISEASES
USING
MACHINE
LEARNING:
A
COMPARATIVE STUDY OF PREDICTIVE MODELS AND
DIAGNOSTIC ACCURACY.
International Journal of
Medical Science and Public Health Research
,
5
(12),
107-118.
28.
Al Mamun, A., Nath, A., Dey, S. K., Nath, P. C.,
Rahman, M. M., Shorna, J. F., & Anjum, N. (2025).
Real-Time
Malware
Detection
in
Cloud
Infrastructures
Using
Convolutional
Neural
Networks: A Deep Learning Framework for
Enhanced Cybersecurity.
The American Journal of
Engineering and Technology
,
7
(03), 252-261.
29.
Akhi, S. S., Shakil, F., Dey, S. K., Tusher, M. I.,
Kamruzzaman, F., Jamee, S. S., ... & Rahman, N.
(2025). Enhancing Banking Cybersecurity: An
Ensemble-Based Predictive Machine Learning
Approach.
The American Journal of Engineering and
Technology
,
7
(03), 88-97.
30.
Mazharul Islam Tusher, “Deep Learning Meets Early
Diagnosis: A Hybrid CNN-DNN Framework for Lung
Cancer Prediction and Clinical Translation”,
ijmsphr
,
vol. 6, no. 05, pp. 63
–
72, May 2025.
31.
Integrating Consumer Sentiment and Deep Learning
for GDP Forecasting: A Novel Approach in Financial
Industry”.,
Int Bus & Eco Adv Jou
, vol. 6, no. 05,
pp.
90
–
101,
May
2025,
10.55640/business/volume06issue05-05.
32.
Tamanna Pervin, Sharmin Akter, Sadia Afrin, Md
Refat Hossain, MD Sajedul Karim Chy, Sadia Akter,
Md Minzamul Hasan, Md Mafuzur Rahman, &
Chowdhury Amin Abdullah. (2025). A Hybrid CNN-
LSTM Approach for Detecting Anomalous Bank
Transactions: Enhancing Financial Fraud Detection
Accuracy.
The American Journal of Management
and
Economics
Innovations
,
7
(04),
116
–
123.
https://doi.org/10.37547/tajmei/Volume07Issue04
-15
33.
Mohammad
Iftekhar
Ayub,
Biswanath
Bhattacharjee, Pinky Akter, Mohammad Nasir
Uddin, Arun Kumar Gharami, Md Iftakhayrul Islam,
Shaidul Islam Suhan, Md Sayem Khan, & Lisa
Chambugong. (2025). Deep Learning for Real-Time
Fraud Detection: Enhancing Credit Card Security in
Banking
Systems.
The American
Journal
of
Engineering and Technology
,
7
(04), 141
–
150.
https://doi.org/10.37547/tajet/Volume07Issue04-
19
34.
Mazharul Islam Tusher, Han Thi Ngoc Phan, Arjina
Akter, Md Rayhan Hassan Mahin, & Estak Ahmed.
(2025). A Machine Learning Ensemble Approach for
Early Detection of Oral Cancer: Integrating Clinical
Data and Imaging Analysis in the Public
Health.
International Journal of Medical Science and
Public
Health
Research
,
6
(04),
07
–
15.
https://doi.org/10.37547/ijmsphr/Volume06Issue0
4-02
35.
Safayet Hossain, Ashadujjaman Sajal, Sakib Salam
Jamee, Sanjida Akter Tisha, Md Tarake Siddique, Md
Omar Obaid, MD Sajedul Karim Chy, & Md Sayem Ul
Haque. (2025). Comparative Analysis of Machine
Learning Models for Credit Risk Prediction in
Banking
Systems.
The American
Journal
of
Engineering
and
Technology
,
7
(04),
22
–
33.
https://doi.org/10.37547/tajet/Volume07Issue04-
04
36.
Ayub, M. I., Bhattacharjee, B., Akter, P., Uddin, M.
N., Gharami, A. K., Islam, M. I., ... & Chambugong, L.
(2025). Deep Learning for Real-Time Fraud
Detection: Enhancing Credit Card Security in
Banking
Systems.
The American
Journal
of
Engineering and Technology
,
7
(04), 141-150.
37.
Siddique, M. T., Uddin, M. J., Chambugong, L.,
Nijhum, A. M., Uddin, M. N., Shahid, R., ... & Ahmed,
M. (2025). AI-Powered Sentiment Analytics in
Banking:
A
BERT
and
LSTM
Perspective.
International Interdisciplinary Business
Economics Advancement Journal
,
6
(05), 135-147.
38.
Thakur, K., Sayed, M. A., Tisha, S. A., Alam, M. K.,
Hasan, M. T., Shorna, J. F., ... & Ayon, E. H. (2025).
Multimodal Deepfake Detection Using Transformer-
Based Large Language Models: A Path Toward
Secure Media and Clinical Integrity.
The American
Journal of Engineering and Technology
,
7
(05), 169-
177.
39.
Al Mamun, A., Nath, A., Dey, S. K., Nath, P. C.,
Rahman, M. M., Shorna, J. F., & Anjum, N. (2025).
Real-Time
Malware
Detection
in
Cloud
Infrastructures
Using
Convolutional
Neural
Networks: A Deep Learning Framework for
Enhanced Cybersecurity.
The American Journal of
Engineering and Technology
,
7
(03), 252-261.
The American Journal of Management and Economics Innovations
85
https://www.theamericanjournals.com/index.php/tajmei
40.
Tusher, M. I., Hasan, M. M., Akter, S., Haider, M.,
Chy, M. S. K., Akhi, S. S., ... & Shaima, M. (2025).
Deep Learning Meets Early Diagnosis: A Hybrid CNN-
DNN Framework for Lung Cancer Prediction and
Clinical Translation.
International Journal of Medical
Science and Public Health Research
,
6
(05), 63-72.
41.
Sajal, A., Chy, M. S. K., Jamee, S. S., Uddin, M. N.,
Khan, M. S., Gharami, A. K., ... & Ahmed, M. (2025).
Forecasting Bank Profitability Using Deep Learning
and Macroeconomic Indicators: A Comparative
Model
Study.
International
Interdisciplinary
Business Economics Advancement Journal
,
6
(06),
08-20.
42.
Paresh Chandra Nath, Md Sajedul Karim Chy, Md
Refat Hossain, Md Rashel Miah, Sakib Salam Jamee,
Mohammad Kawsur Sharif, Md Shakhaowat
Hossain, & Mousumi Ahmed. (2025). Comparative
Performance of Large Language Models for
Sentiment Analysis of Consumer Feedback in the
Banking Sector: Accuracy, Efficiency, and Practical
Deployment.
Frontline Marketing, Management
and
Economics
Journal
,
5
(06),
07
–
19.
https://doi.org/10.37547/marketing-fmmej-05-06-
02
43.
Hossain, S., Siddique, M. T., Hosen, M. M., Jamee, S.
S., Akter, S., Akter, P., ... & Khan, M. S. (2025).
Comparative Analysis of Sentiment Analysis Models
for Consumer Feedback: Evaluating the Impact of
Machine Learning and Deep Learning Approaches
on Business Strategies.
Frontline Social Sciences and
History Journal
,
5
(02), 18-29.
44.
Jamee, S. S., Sajal, A., Obaid, M. O., Uddin, M. N.,
Haque, M. S. U., Gharami, A. K., ... & FARHAN, M.
(2025). Integrating Consumer Sentiment and Deep
Learning for GDP Forecasting: A Novel Approach in
Financial Industry.
International Interdisciplinary
Business Economics Advancement Journal
,
6
(05),
90-101.
45.
Hossain, S., Sajal, A., Jamee, S. S., Tisha, S. A.,
Siddique, M. T., Obaid, M. O., ... & Haque, M. S. U.
(2025). Comparative Analysis of Machine Learning
Models for Credit Risk Prediction in Banking
Systems.
The American Journal of Engineering and
Technology
,
7
(04), 22-33.
46.
Pabel, M. A. H., Bhattacharjee, B., Dey, S. K., Jamee,
S. S., Obaid, M. O., Mia, M. S., ... & Sharif, M. K.
BUSINESS
ANALYTICS
FOR
CUSTOMER
SEGMENTATION: A COMPARATIVE STUDY OF
MACHINE
LEARNING
ALGORITHMS
IN
PERSONALIZED BANKING SERVICES.
