Journal of Economics, Finance and Accounting Studies
ISSN: 2709-0809
DOI: 10.32996/jefas
JEFAS
AL-KINDI CENTER FOR RESEARCH
Journal Homepage: www.al-kindipublisher.com/index.php/jefas AND DEVELOPMENT
| RESEARCH ARTICLE
Novel AI-Powered Dynamic Inventory Management Algorithm in the USA: Machine
Learning Dimension
Md Zahidul Islam1, Nisha Gurung2, Md Sumon Gazi3 and Md Rokibul Hasan4 ✉
1234MBA Business Analytics, Gannon University, Erie PA, USA
Corresponding Author: Md Rokibul Hasan, E-mail:
[email protected]| ABSTRACT
Dynamic inventory management revolves around the practice of progressively modifying inventory degrees to adapt to
fluctuations in client demand, production, and supply chain dynamics. At the center, inventory management focuses on upholding
enhanced levels of stock to balance consumer service via availability with the costs related to holding excess inventory. This
research paper aimed to explore the dynamic inventory management activities employed by organizations in the USA, shedding
light on the machine learning strategies that can be deployed and their implications. The performance of the algorithms was
empirically evaluated in a Python program experiment utilizing real-world data. To facilitate the data for input into the Neural
Network, feature engineering, and selection were imposed to affirm its suitability. This study proposes the Sequence-to-Sequence
(Seq2Quant) algorithm, a neural network-powered technique for demand prediction in inventory management. The current
experiment compared and contrasted the performance of the Neural Networks against the following baselines, most notably,
Naïve Seasonal Forecast, Moving Average Forecast, ARIMA, Naïve Seasonal Forecast with Averaging over four periods, SARIMAX.
From the experiment, it was evident that the Seq2Seq had the lowest MAE (17.44) and the lowest SMAPE (66.91), suggesting that
it was the best-performing algorithm overall. Besides, SARIMAX and ARIMAX also performed well, with MAE values of 18.33 and
18.09, respectively.
| KEYWORDS
Dynamic Inventory; Neural Network (NN); Recurrent Neural Network (RNN); Seq2Quant; Naïve Seasonal Forecast; Moving
Average Forecast; ARIMA; SARIMAX.
| ARTICLE INFORMATION
ACCEPTED: 02 April 2024 PUBLISHED: 27 April 2024 DOI: 10.32996/jefas.2024.6.2.12
1. Introduction
Chen & Chao (2020), states that inventory management is an imperative function for companies across sectors in the United States,
allowing them to balance demand and supply efficiently. Dynamic inventory management revolves around the practice of
progressively modifying inventory degrees to adapt to fluctuations in client demand, production, and supply chain dynamics.
Efficient inventory management is a strategic advantage for companies, and dynamic methods afford several competitive benefits.
By assessing demand trends and modifying inventory levels respectively, organizations can diminish stockouts, minimize carrying
costs, enhance production schedules, and boost customer satisfaction. As per, Cuartas & Aguilar (2023), dynamic inventory
management also enables businesses to respond to market changes rapidly, improving overall agility and competitiveness. This
research paper intends to explore the dynamic inventory management activities employed by organizations in the USA, shedding
light on the strategies deployed and their implications.
Chaudhary et al. (2023), indicate that at the center, inventory management focuses on upholding enhanced levels of stock to
balance consumer service via availability with the costs related to holding excess inventory. Conventional inventory management
frameworks were grounded on presumptions of supply lead times and solid demand patterns. schedules and replenishment
Copyright: © 2024 the Author(s). This article is an open access article distributed under the terms and conditions of the Creative Commons
Attribution (CC-BY) 4.0 license (https://creativecommons.org/licenses/by/4.0/). Published by Al-Kindi Centre for Research and Development,
London, United Kingdom.
Page | 156
JEFAS 6(2): 156-168
quantities were fixed. Nevertheless, in the contemporary complex business atmosphere, demand changes continuously and is
impacted by a different factor such as new product launches, seasonal trends, competitor activities, and the macroeconomic state.
Supply chains have also become internationally circulated with variable lead times and a greater risk of disruptions.
Demizu et al. (2023), argue that to resolve the uncertain and dynamic business landscape, dynamic inventory management has
arisen as an advanced method. Dynamic inventory management depends on the progressive assessment of supply chain events,
real-time demand signals, sales predictions, and other associated data sources to adapt inventory replenishment frameworks
dynamically on a progressive basis. The prime concept is that supply chain variables and demand trends are in constant flux, so
inventory frameworks need to be adaptive and fluid based on the most up-to-date real-time data. Contrary to the mainstream
fixed policies, dynamic inventory management targets to right-size stock levels in real-time by taking a more data-oriented and
responsive approach.
2. Literature Review
While the present prevalence revolving around Artificial Intelligence is relatively new, the notion of utilizing machine learning for
predictive purposes has been in existence for many decades. In a comprehensive examination of current literature, De Melo (2019),
highlighted that the first usage of Neural Networks in prediction as far back as 1974. Nonetheless, research in this domain was
quite limited until the development of the backpropagation models in the 1980s, which enabled the training of deep networks.
This era of early investigations is attributed to enthusiasm for the new technique, which was deemed a capable technique for
modeling non-linear time series. Based on an extensive review of the existing literature, De Melo (2019), concluded that Neural
Networks exemplified usefulness and suitability in delivering satisfactory performance and predicting tasks.
Lolli et al. (2022) conducted a comprehensive empirical study on predicting inventories using Neural Networks. Particularly, they
performed a comparative examination between an automated Box-Jenkins forecasting and a Multilayer Perceptron (MLP) network
system, evaluating a dataset comprising 75-time series. Their research exposed that the simple neural network algorithm achieved
similar level outcomes compared to the Box-Jenkins predictive system. On the other hand, Meisheri et al. (2023) found out through
their research on 384 demographic and economic time series that Neural Networks produced relatively accurate predictions
compared to traditional methods. The predictions produced by the networks were comparatively superior to those of a linear
regression algorithm and a weighted average of six simpler techniques.
Even though Machine Learning techniques, exemplified good performance, Namir & Labriji (2022), presented a hybrid approach
that consolidated Machine Learning with a statistical technique, which yielded excellent results. Chaudhary et al. (2020) observed
that this specific algorithm employed a Neural Network to learn and predict seasonality trends. Impressively, the Neural Network
did not only learn these trends for every individual series independently but rather from a consolidation of all the series. These
results led Makridakis and coworkers to hypothesize that utilizing information from multiple series to forecast individual series
yields favorable outcomes.
In their research, Demizu et al. (2023) proposed a global Long Short-Term Memory (LSTM) framework. As an extension to their
method, they adopted different clustering models to pinpoint identical time series and suggested training the frameworks
exclusively on these clusters. Their investigation ascertained that this clustering method led to enhanced predictions. The
recommended algorithms demonstrated competitive performance when assessed on the datasets from two previous prediction
competitions.
3. Methodology
The performance of the algorithms was empirically evaluated in a Python program experiment utilizing real-world data. The
algorithm will forecast the demand for several hundred inventories in the product segment of an American wholesale organization
over a three-month timeline. Subsequently, a simulation was performed to compute the effect on the inventory advancement of
the product segment under situations that closely resemble reality (proAIrokibul, 2024). The system will be compared and
compared to a statistical baseline framework to ascertain if the possible enhancement in performance is substantial enough to
justify the increased computational complexity.
3.1 Dataset
For the current research experiment, the organization supplied extensive data associated with their product segment. The data
was broadly classified into three main categories, most notably, logistical process information, historical data, and product-specific
data coupled with logistical properties. The historical data comprised the period from April 1, 2018, to April 31, 2020, which
indicated the maximum timeline available from the company system during that time. This data was gathered from the
organization's five primary warehouses situated in America, which attend to both clients and intermediate warehouses in the
supply chain (proAIrokibul, 2024). The available data comprised three different datasets. The initial dataset comprises data on daily
Page | 157
Novel AI-Powered Dynamic Inventory Management Algorithm in the USA: Machine Learning Dimension
sales and the quantity of discarded products. The second dataset concentrates on client orders for items and the corresponding
degree of fulfillment. Ultimately, the third historical dataset offers details regarding past promotional activities carried out for the
product.
Pre-processing
The data preprocessing entailed two main phases. Initially, the data required to be formatted in a manner that the software could
read and understand. The initial data was unstructured and came in Excel and CSV files. The data was normalized by transforming
it into preserved panda data frames in Python with constant feature labeling. The second segment of preprocessing was pilot data
filtering. An initial assessment of the sales data via visualization identified outlier values that potentially skewed the analysis
(proAIrokibul, 2024). This demonstrated the demand for some initial filtering of outliers. Subsequently, a rolling window filter was
employed utilizing z-scores to point out outlier records in the sales data.
3.2 Feature Engineering and Selection
To facilitate the data for input into the Neural Network, feature engineering, and selection were imposed to affirm its suitability.
In particular, to compute equidistant annotations within the time series, the data was resampled to comprise a single annotation
for each business day, ranging from Monday to Friday. In the rare scenario where demand happened on Saturdays, possibly
because of shifted delivery schedules due to holidays, this demand was viewed as if it had materialized on the previous Friday
(proAIrokibul, 2024). This technique affirmed that the time series affirmed a constant rhythm, with observations partitioned at
consistent intervals, a pivotal prerequisite for efficient analysis by the Neural Network.
Fig 1: Showcases normalized seasonality attribute for every individual time series
Page | 158
JEFAS 6(2): 156-168
For the training algorithm, the system utilized two types of features - exogenous features and engineered features. Apart from the
actual product sales values, different categorical variables were also inserted into the system as supplementary predictors or
covariates (proAIrokibul, 2024). The categorical attributes that were passed entailed the product element that the product
belonged to, the place where the item was sold, and a distinct index comprised of the product ID and location name. This index
facilitated the independent time series for every item-location combination to be observed.
Feature Depiction Utilized in
Categorical Data Index, location, and product segment Decoder & Encoder
Time Series Observation Definite Sales per SKU per geographical Encoder
location
Date Features Day, Month, Year, or Weekday Decoder & Encoder
Weather Data Average temperature, precipitation, and Decoder & Encoder
Sunshine hours
Related Item Sales Weekly and Daily average of product segment Encoder
and location
Holiday Sixteen binary attributes Decoder & Encoder
Models and Metrics
Neural Networks
The neural network algorithm is likened to the functionality of the human brain, which thrives at pinpointing patterns. A basic form
of the neural network is the single-layer NN, also termed a "perceptron." A Neural Network encompasses an input layer, an
activation function, and an output layer. In the displayed diagram, x1...xn denotes the input attribute characteristic. Each attribute
characteristic is multiplied by the equivalent weight w1...wn, which ascertains the intensity of the nodes (proAIrokibul, 2024). The
bias value b is utilized to modify the activation formula, facilitating it to be shifted to lower or higher levels.
𝑦 = 𝑥0𝑤0 + 𝑥1𝑤1 + 𝑥2𝑤2 + ⋯ + 𝑥𝑛𝑤𝑛 + 𝑏
Where,
• Y is the targeted variable.
• X.=X1……Xn denotes the inserted feature characteristic.
• Wi=W1………. is the weight
3.3 Recurrent Neural Network
Recurrent Neural Networks (RNNs) are a form of neural network that integrates a feedback loop system, enabling them to uphold
information from previous inputs. These networks pass a vector via a distinct hidden layer, deemed as the hidden state (h_t), at
every input phase. This hidden condition facilitates RNNs to offer a setting for interpreting inputs, hindering them from being
modeled in seclusion. As a consequence, RNNs can process progressive information and manage sequences of varying lengths. A
typical Recurrent Neural Network's (RNN) basic structure is demonstrated on the right-hand side of Figure 3. To showcase how
data flows across sequential inputs, the network is showcased unrolled overtime on the left-hand side of the figure.
Page | 159
Novel AI-Powered Dynamic Inventory Management Algorithm in the USA: Machine Learning Dimension
In this algorithm, time 𝑡, the input is represented as 𝑥𝑡, while the output is 𝑜𝑡. The hidden state ℎ𝑡 is impacted by 𝑥𝑡 and the
previous hidden state ℎ𝑡-1, regulated by the weight matrices W and U. The output is retrieved from the present hidden state and
the weight matrix V. This association can be mathematically depicted as:
ℎ𝑡 = 𝑓9(𝑈𝑥𝑡 + 𝑊ℎ𝑡&1),
𝑜𝑡 = 𝑓𝛼(𝑉ℎ𝑡),
where 𝑓𝛼 & 𝑓9, represent the non-linear formula as the ReLu-function or the tanh-function.
3.4 Experimentation
The performance of the algorithms was empirically assessed in a Python program experiment utilizing real-world data. The
algorithm was used to forecast the demand for several hundred inventories in the product segment of an American wholesale
organization over a three-month timeline. Subsequently, a simulation will be performed to compute the effect on the inventory
advancement of the product segment under situations that closely resemble reality.
Importing Libraries
Page | 160
JEFAS 6(2): 156-168
Output:
Subsequently, a code snippet was imposed to visualize the distribution of sales, the outcome can be displayed as follows:
Output:
Page | 161
Novel AI-Powered Dynamic Inventory Management Algorithm in the USA: Machine Learning Dimension
Equally important to visualize the distribution of categorical features, a code snippet was applied respectively as follows:
Page | 162
JEFAS 6(2): 156-168
Output:
Page | 163
Novel AI-Powered Dynamic Inventory Management Algorithm in the USA: Machine Learning Dimension
Furthermore, the analyst attempted to impose a code snippet to visualize the world cloud of product description as follows:
Output:
Page | 164
JEFAS 6(2): 156-168
Proposed Model
This study proposes the Sequence-to-Sequence (Seq2Quant) algorithm, a neural network-powered technique for demand
prediction in inventory management. The Seq2Seq algorithm is a relatively novel demand predictive solution crafted to address
the specific pressing issue in inventory management. The recommended model is a global modeling technique that converts
multivariate input sequences of historical data and covariates into a sequence of approximated demand quantiles. Particularly, the
algorithm is a sequence-to-sequence mapping that forecasts the quantiles of cumulative demand.
3.5 Models Performance Evaluations
The current experiment compared and contrasted the performance of the Neural Networks against the followingbaselines:
❖ Naïve Seasonal Forecast
❖ Moving Average Forecast
❖ ARIMA
❖ Naïve Seasonal Forecast with Averaging over four periods
❖ SARIMAX
4. Results
Page | 165
Novel AI-Powered Dynamic Inventory Management Algorithm in the USA: Machine Learning Dimension
From the table above, it was evident that the Seq2Seq had the lowest MAE (17.44) and the lowest SMAPE (66.91), suggesting that
it was the best-performing algorithm overall. Besides, SARIMAX and ARIMAX also perform well, with MAE values of 18.33 and
18.09, respectively.
Figure 3 above exhibits the average SMAPE of all algorithms for forecasting n steps. As expected, prediction accuracy reduces with
the length of the predicting horizon. The Seq2seq algorithm outperformed the others for the majority of predicting horizons.
Nonetheless, for short-term predictions of less than a week ahead, the Seasonal Naïve technique has a slightly better average
SMAPE.
4.1 Business Impact
4.1.1 Benefits to the Organization
1. Enhanced Inventory Optimization: By deploying the proposed Seq2quant algorithms to assess market trends, historical
sales data, and different factors, businesses in the USA can attain precise demand predictions. As a result, this algorithm
facilitates companies to optimize inventory management, reduce stockouts, and elevate overall customer satisfaction.
2. Improved Supply Chain Visibility: The seq2quant algorithm affords businesses in America real-time visibility on supply chain
data, facilitating companies to optimize inventory levels and reorder points in multiple locations. By progressively evaluating
supply chain data, the se2quant system can dynamically modify inventory levels and reorder points depending on the
fluctuating demand and supply conditions, facilitating real-time optimization.
3. Reduced Costs and Waste: By streamlining inventory levels and enhancing demand prediction, SMEs in America can minimize
the costs related to holding excess inventory. This entails obsolescence costs, storage costs, and the cost of disposing and
managing waste.
Page | 166
JEFAS 6(2): 156-168
4.1.2 How to Implement the Seq2Seq Algorithm
1. Step 1: Define the problem The problem statement for the dynamic inventory management framework should be
determined using the seq2seq algorithm revolving around predicting the inventory levels for the company demand patterns.
2. Step 2: Prepare the data- Businesses need to prepare the data for training the seq2seq algorithm. Businesses can utilize
historical data to train the algorithm.
3. Step 3: Define the algorithm architecture- The seq2seq algorithm comprises two main elements: decoder and encoder. The
encoder obtains the input sequence and converts it into a fixed-length vector. Conversely, the decoder obtains the fixed-
length vector and produces the output sequence.
4. Step 4: Train the algorithm- Businesses can train the algorithm utilizing a backpropagation through time (BPTT) system.
Subsequently, business analysts can utilize the mean squared error (MSE) loss formula to compute the loss between the
forecasted inventory levels and the real-time inventory levels.
5. Step 5: Evaluate the model-Business can assess the algorithm utilizing root mean squared error (RMSE) or mean absolute
error (MAE) metrics. Besides, companies can also adopt visualization methods to compare the predicted inventory levels with
the actual inventory levels.
6. Step 6: Deploy the algorithm: Businesses can go ahead to implement the algorithm in the dynamic inventory management
framework to forecast the inventory levels for new demand patterns. SME can deploy the algorithm, to generate alerts when
the inventory levels are low or high.
Funding: This research received no external funding.
Conflicts of Interest: The authors declare no conflict of interest.
Publisher’s Note: All claims expressed in this article are solely those of the authors and do not necessarily represent those of
their affiliated organizations, or those of the publisher, the editors and the reviewers.
5. Conclusion
This research paper aimed to explore the dynamic inventory management activities employed by organizations in the USA,
shedding light on the machine learning strategies that can be deployed and their implications. The performance of the algorithms
was empirically evaluated in a Python program experiment utilizing real-world data. To facilitate the data for input into the Neural
Network, feature engineering, and selection were imposed to affirm its suitability. This study proposes the Sequence-to-Sequence
(Seq2Quant) algorithm, a neural network-powered technique for demand prediction in inventory management. The current
experiment compared and contrasted the performance of the Neural Networks against the followingbaselines, most notably, Naïve
Seasonal Forecast, Moving Average Forecast, ARIMA, Naïve Seasonal Forecast with Averaging over four periods, SARIMAX. From
the experiment, it was evident that the Seq2Seq had the lowest MAE (17.44) and the lowest SMAPE (66.91), suggesting that it was
the best-performing algorithm overall. Besides, SARIMAX and ARIMAX also performed well, with MAE values of 18.33 and 18.09,
respectively.
Page | 167
Novel AI-Powered Dynamic Inventory Management Algorithm in the USA: Machine Learning Dimension
References
[1] Chen, B., & Chao, X. (2020), Dynamic inventory control with stockout substitution and demand learning. Management Science, 66(11), 5108-
5127.
[2] Cuartas, C., & Aguilar, J. (2023). Hybrid algorithm based on reinforcement learning for smart inventory management. Journal of intelligent
manufacturing, 34(1), 123-149.
[3] Chaudhary, V., Bharadwaja, K., Meena, R. S., Bikash, P., Acharjee, D. N. C. C., & Gopinathan, R. (2023). Exploring the Use of Machine Learning
in Inventory Management for Increased Profitability.
[4] Demizu, T., Fukazawa, Y., & Morita, H. (2023). Inventory management of new products in retailers using model-based deep reinforcement
learning. Expert Systems with Applications, 229, 120256.
[5] De Melo, Á. S. (2019). A machine learning approach to the optimization of inventory management policies.
[6] Lolli, F., Balugani, E., Ishizaka, A., Gamberini, R., Rimini, B., & Regattieri, A. (2022). Machine learning for multi-criteria inventory classification
applied to intermittent demand. Production Planning & Control, 30(1), 76-89.
[7] Meisheri, H., Baniwal, V., Sultana, N. N., Khadilkar, H., & Ravindran, B. (2023). Using reinforcement learning for a large variable-dimensional
inventory management problem. In Adaptive learning agents workshop at AAMAS (pp. 1-9).
[8] Namir, K., & Labriji, H. (2022). Decision support tool for dynamic inventory management using machine learning, time series and
combinatorial optimization. Procedia Computer Science, 198, 423-428.
[9] ProAIrokibul. (2024). Dynamic-Inventory-Management/Model/Inventory Management.ipynb at main · proAIrokibul/Dynamic-Inventory-
Management. GitHub. https://github.com/proAIrokibul/Dynamic-Inventory-
Management/blob/main/Model/Inventory%20Management.ipynb
Page | 168
