Investigating the role of using AI and machine learning for demand forecasting in supply chain management

A.DEMAND FORECASTING IN THE MANUFACTURING SECTOR

Demand forecasting has a big impact on all aspects of manufacturing since it affects how many goods have to be ordered, when they should be ordered, how much product can be produced, and where to allocate resources (i.e., labor and warehouse space). A comprehensive review by [17] points out that AI, ML, and especially Scientific Machine Learning (SciML) are now proving to be increasingly important in predicting demand because they can use hybrid forecasting methods to perform much quicker and with greater precision than traditional methods. This type of forecasting is enhanced through automated workflows, including ETL pipelines and AutoML, as well as through using reinforcement learning and hybrid algorithms to develop more effective scheduling and logistics systems in constantly changing environments. According to [18], traditional statistical methods often do not adequately reflect the complexities of industrial processes. ML techniques including regression, classification, and clustering can provide adaptive and continuously updated forecasting models. Furthermore, algorithms such as random forests (RFs) and artificial neural networks (ANNs) have demonstrated exceptional predictive performance. Research by [19] shows that AI models (e.g., support vector regression) have consistently outperformed traditional demand forecasting methods in complex environments (e.g., forecasting demand of spare parts). As a result of these emerging trends, AI has emerged as a critical success factor for efficient, adaptive, and Industry 4.0-compliant manufacturing systems.

B.AI-DRIVEN DEMAND FORECASTING IN THE RETAIL SECTOR

Recent advancements in retail, e-commerce, and big data have led to a paradigm shift for the industry toward the implementation of demand forecasting techniques. The use of traditional statistical tools has become less useful in generating accurate demand forecasts due to their ability to model nonlinear relationships. Therefore, businesses are now looking to take advantage of AI technologies to assist with demand forecasting. ML algorithms, such as ANNs, long short-term memory (LSTM), and recurrent neural networks (RNN) within the fast-moving consumer goods (FMCG) sector, have proven to be more predictive than traditional demand forecasting methods as they effectively reflect how consumers’ behavior will change over time. While hybrid ARIMA and ML modeling techniques can enhance accuracy, their higher costs and need for the organization’s internal expertise limit their implementation [20]. As shown in [21], the success of using hybrid ensemble methods to forecast demand can be demonstrated through the efforts made in creating a hybrid RF–extreme gradient boosting–linear regression (RF-XGBoost-LR) model that has a high level of predictive accuracy (R² = 0.9551). The ability of these models to produce highly generalized results and scale well across multiple different datasets has demonstrated the effectiveness of hybrid ensemble methods. Reference [22] also looked at various methods of forecasting by converting them to a classification method to help produce probabilistic forecasts with a lower incidence of forecast error than traditional forecasting methods, particularly during instances of irregular demands. In summary, many of the studies reviewed indicate a shift toward the use of hybrid and adaptive probabilistic AI-based techniques for demand forecasting. However, significant organizational and infrastructural barriers still exist which will limit widespread use of these techniques. Despite strong predictive performance reported by hybrid and deep learning models, most studies prioritize accuracy over interpretability and operational integration. Additionally, these models often assume stable demand distributions, limiting their effectiveness in volatile retail environments. This reveals a key limitation in current research: the lack of frameworks that balance predictive performance with adaptability and managerial usability.

C.AI-ENHANCED DEMAND FORECASTING IN THE HEALTHCARE SECTOR

Due to their multifaceted and unpredictable nature, it is inherently difficult to forecast healthcare demand because of the diversity and randomness of patient needs; therefore, healthcare demand forecasting must include clinical, behavioral, environmental, and genomic data. AI and ML are among the best solutions to address these issues. In [23], the authors compiled a list of key data sources to support this solution and proposed a six-step predictive analytics framework that included the need for big-data infrastructures such as Hadoop and NoSQL databases, along with a means of managing heterogeneous datasets to be used for suitable forecasting. Practical work on this approach includes the use of ML models, including RF models, to predict the demand for essential medicines, with very high levels of accuracy and utilization of the information to develop a chain of supply to hospitals and other health facilities and to demonstrate project regional variances in demand [24]. A separate study by [25] demonstrates the use of hybrid deep learning techniques to create a multi-source time-series fusion model that combined both environmental data and clinical data, thus greatly enhancing predictive robustness. These works illustrate how AI is changing the focus of healthcare systems from being reactive to being proactive, thus improving the allocation of resources in the healthcare system, as well as improving patient care outcomes. Nonetheless, there are still significant challenges to be addressed in the areas of data privacy, interoperability, and the transparency of the models that will continue to be of interest in future research.

D.CROSS-SECTORAL DISCUSSION AND EMERGING TRENDS

Comparing the three sectors, namely manufacturing, retail, and healthcare, reveals both shared advances and distinct challenges. Across all domains, AI-driven demand forecasting has evolved from isolated analytical tools to integrated decision support systems. Manufacturing emphasizes production optimization and process control, retail focuses on demand variability and hybrid model adoption, while healthcare prioritizes data integration and predictive accuracy for service delivery. Despite contextual differences, three overarching trends define the current state of the art. First, hybridization has become the cornerstone of advanced forecasting. The convergence of statistical, ML, and domain-specific models as demonstrated by [21,25] enhances adaptability and precision across diverse data environments. Second, **data fusion and feature engineering** are critical enablers, allowing the synthesis of structured, semi-structured, and unstructured datasets for richer predictive insights. Finally, the **shift toward explainable and ethical AI** reflects growing awareness of trust and accountability, especially in sensitive sectors like healthcare. While AI and ML have demonstrated immense potential, the journey toward full-scale industrial adoption remains ongoing. Future research must focus on developing interpretable, energy-efficient, and context-aware forecasting systems that not only predict but also prescribe optimal actions. By aligning technological innovation with ethical governance and domain-specific needs, AI-driven demand forecasting can continue to redefine efficiency, resilience, and sustainability across all sectors.

E.DISCUSSION AND ANALYSIS

There has been an increasing trend in using AI for demand forecasting. However, much of the current research on this topic has been primarily focused on either supervised models that are aimed toward improving prediction accuracy or unsupervised models that focus on identifying patterns within datasets rather than combining the strengths of both. This lack of comprehensive studies is problematic, as there is currently little data available to show how these two methods could successfully work together within a single framework to achieve both greater predictive accuracy and better identify the hidden demand structures within the dynamic nature of the retail industry. Furthermore, many of the current demand forecasting models do not provide generalizable results due to their inability to handle this type of heterogeneous, high-dimensional data that incorporates external factors such as oil price fluctuations, seasonal adjustability, and holiday schedules. Many current studies fail to validate their models properly by not sufficiently applying validation techniques, optimizing hyperparameters, and/or using weak generalizability techniques, which further limits their practical applications in real-world supply chain systems. Thus, a need for a complete, robust, and scientifically validated framework exists that integrates multiple AI methodologies to overcome the limitations previously discussed.

Accordingly, the objective of this study is to develop and evaluate a conceptual framework that integrates supervised (RF) and unsupervised (clustering) models to improve demand forecasting performance under dynamic market conditions. This study aims to assess how these AI models adapt to changing consumer behaviors and external factors, thereby addressing the critical gap in existing research regarding the efficiency of AI-driven demand forecasting. Creating an RF supervised model and unsupervised clusters provides insights into which model performs better in handling the complexity and variability of data in SCM, ultimately aiding organizations in making more informed and timely decisions.

A.PROBLEM IDENTIFICATION

Despite extensive research and advancements, a common theme emerges from the literature review indicating that the effectiveness of demand forecasting varies considerably across different sectors influenced by the rapid evolution of market conditions. This variability can be due to differing market dynamics, consumer behaviors, and the specific nature of products within each sector. While some sectors experience slow and predictable changes, others face rapid shifts that demand quicker responsiveness from forecasting models. A major challenge frequently discussed is that current AI models are not quick enough to accurately forecast demand in these fast-changing markets. Both supervised and unsupervised models often fail to keep up with sudden changes in consumer preferences and market trends, making their forecasts outdated soon after they are generated. Moreover, there isn’t a single AI model capable of effectively forecasting demand across a wide range of products and industries. Each model has its strengths and weaknesses and is usually optimized for specific types of data or market conditions, limiting their usefulness across the broad spectrum of global industries. The identified research gap is the “lack of evaluation of the capability of AI models to efficiently forecast demand concerning changing market conditions, consumer behavior, and external factors.” This gap underscores the need for a study that not only tests the adaptability of these models to dynamic environments but also evaluates their effectiveness across a broad spectrum of scenarios. The primary objective of this paper, therefore, is to develop and assess models that enhance the predictive accuracy and operational efficiency of demand forecasting in SCM under varying market conditions. This will involve the design and development of an RF supervised model and the application of unsupervised clustering techniques. These models will be rigorously tested to determine how well they can adapt to changing consumer behaviors and external factors.

B.OBJECTIVE REQUIREMENT

This study has employed the DSR method to develop and evaluate AI demand forecasting models in the supply chain. We first gathered information from various sources, including sales history, promotions, and other variables that impacted demand (such as economic conditions and holidays), to develop a better grasp of the changes in demand. After determining how to gather demand-related information, we characterized forecasting challenges, especially surrounding the inability of current models to adapt quickly to changes in the marketplace. During model development, we created various models (e.g., RF, neural networks based on Keras, and unsupervised clustering) with the objective of providing greater accuracy and versatility in predicting future demand. A functional prototype was then developed, providing a demonstration of the practical usability and feasibility of integrating the models into supply chain systems. Through the use of scenario-based testing to simulate real-world settings, the models were tested to allow researchers to evaluate their robustness to variations in the input data. Each mode was evaluated through a combination of quantitative measures (e.g., accuracy, precision, recall, F1-score, mean squared error (MSE), and coefficient of determination (R²)) and qualitative assessments from industry experts on how well each mode performed. The system is intended to meet key attributes, including scalability, usability, reliability, and seamless ERP integration. At the conclusion of the project, we anticipate that our findings will provide practical, efficient, and versatile demand forecasting solutions for SCM while at the same time producing useful insights to advance both academic research and practical applications in the supply chain arena.

C.ANALYSIS

The analysis phase of this paper focuses on improving demand forecasting in SCM by evaluating advanced AI and ML techniques. Existing research highlights that traditional forecasting methods are limited in handling complex, dynamic market environments. Studies demonstrate that ML models including ANNs, RNNs, and support vector machines (SVMs) offer improved accuracy compared to conventional techniques. However, these models also present notable challenges. SVMs handle linear and nonlinear data well but require extensive parameter tuning, which increases computational cost and complexity. ANNs are capable of learning deep nonlinear patterns but can become trapped in local minima during training and require substantial computational power. Both models also struggle with optimal feature selection, risking overfitting or loss of critical information.

RF provides a more robust alternative by constructing multiple decision trees and aggregating their predictions. This reduces variance, prevents overfitting, and enhances generalization, making RF highly suitable for large and complex forecasting tasks. Research also shows the complementary value of KNNs, which enable fast experimentation through a structured pipeline of model definition, compilation, training, evaluation, and prediction. To enrich predictive modeling, the analysis integrates unsupervised clustering techniques, recognizing that supervised methods alone may overlook hidden market structures. K-means clustering assists in identifying natural groupings among stores or products based on sales patterns, demographics, or purchasing behaviors. These insights support targeted strategies such as inventory optimization and market segmentation.

The analytical process begins with compiling diverse datasets essential for understanding demand drivers, including store attributes, item characteristics, over 125 million transaction records, oil price fluctuations, and holiday calendars. These datasets undergo cleaning and integration to ensure quality and consistency. Exploratory data analysis then identifies trends, anomalies, and external influences such as oil price shifts or seasonal patterns that impact purchasing behavior. This comprehensive analysis strengthens the foundation for building accurate, adaptive forecasting models capable of supporting data-driven decision-making in SCM. Tables I, II, and III give a detailed view of the datasets and their attributes. During initial clustering experiments, noninformative identifiers such as “id” and “store_nbr” were inadvertently included. These variables were removed in the refined model, as they do not carry meaningful analytical information and can distort clustering outcomes.

Table I. Store dataset summary

Table II. Item dataset summary

Table III. Train dataset summary

D.CONCEPTUAL MODEL

Figure 1 shows a conceptual model that involves a dataset merged from different databases and systems such as ERP, TPS, and SCM. The data obtained from these systems alongside the merged dataset obtained from combining all those are used to train and test the model using the supervised methods and the hidden layer where the implementation of the artifact occurs. The conceptual model displayed represents the framework for business analytics, focusing on demand forecasting through a structured process that integrates ML techniques and data from various business systems. This process is presented in layers, illustrating the flow from data input to business decision-making.

Fig. 1. Demand forecasting system architecture flowchart.

At the foundational level (input layer), various data sources feed into the system. These include merged datasets, an ERP system, organizational historical data, an SCM system, and a transaction processing system. These diverse data sources provide a comprehensive input base ensuring that the forecasting models have access to all relevant data sources necessary for accurate analysis.

The hidden layer represents the core analytical engine of the architecture where data are transformed and insights are generated. It acts as a bridge between raw data inputs and the outputs that inform decision-making. The primary activities in this layer include data cleaning and preprocessing which is crucial as it prepares the raw data for modeling by cleaning and structuring it, ensuring that the data modeling techniques can be effectively applied. The next step is data modeling where three various modeling techniques are employed to analyze the data such as the RF model. This supervised technique uses multiple decision trees to output a forecast that improves accuracy by reducing the risk of overfitting typical of single decision tree models. The second model is KNN which is a deep learning approach where data are processed through layers of neurons adapting and improving accuracy through iterative training on the input data. The third and final unsupervised model is clustering where it groups data into clusters based on similarity, which can be particularly useful for identifying patterns or behaviors in data that are not explicitly labeled. The results from the data modeling are compiled into forecasts, such as demand forecasts obtained from different models (RF, neural network, and clustering). These forecasts are then presented in a manner that allows straightforward interpretation and application in business contexts.

Finally, a managerial figure evaluates the outputs. This evaluation is presumably based on the comparative insights drawn from the forecasts provided by different models. The manager’s role is to interpret these forecasts, considering the business context and strategic objectives, and make informed decisions that will impact the organization’s operational or strategic directions. The manager reviews the outputs provided by the system to make informed business decisions. The decision-making is based on the analytical insights provided by the models. Data flow from the database layer up to the hidden layer where it is processed and modeled. The results then flow to the output layer where they are consolidated and presented in a form useful for managerial decision-making. The manager evaluates these outputs to make strategic business decisions, effectively closing the loop of the decision support system. This model illustrates the use of data-driven decision-making using a combination of traditional and advanced data analysis techniques in a business environment. The goal is to integrate data across various functions to provide a comprehensive view that supports strategic and operational decisions. Overall, the architecture effectively integrates data from a broad spectrum of internal systems, applies advanced analytical techniques, and outputs actionable insights. This enables a data-driven approach to decision-making, which is critical in modern business environments where efficiency and accuracy in forecasting can provide a significant competitive advantage.

E.DEMAND FORECASTING MODELS PROTOTYPE

According to [31], one example of a model design framework is the CRISP-DM model. Additionally, an addition to this model is the data mining methodology referred to as the CRISP-DM model, which consists of six phases. These phases in data mining include Business Understanding, Data Understanding, Data Preparation, Modeling, Evaluation, and Deployment, as shown in Fig. 3.

Fig. 3. Modeling output of random forest and KNN models.

The focus of this paper is on demand forecasting through three different stages: data understanding, data preparation, and modeling implementing in Google Colab with the use of Python. The way that the exploratory data analysis and visualization methods (bar plots) were used during the data understanding stage of the process was to help to identify the distributions of the data (e.g., store type) as well as to help to identify any imbalances and assist in selecting the models that provided the most amount of noise reduction. If it were not for the data preparation process, much of the model accuracy and reliability would not have been able to improve at all. For this reason, I wanted to handle any missing values in the dataset (such as oil prices) through the use of mean imputation as this would help to keep the data consistent across the board as well as to help select only the top 10 best-selling products to simplify the dataset further and increase the efficiency of the forecasting model (the validity/generalizability of the resulting prediction would be somewhat limited). Furthermore, the categorical data have been transformed using one-hot and label encoding, and the numerical data have also been normalized so that the model does not tend toward one data type or another during the learning of the predictions.

Feature engineering techniques were used to create new variables, such as time-series variables and interaction terms, that capture seasonal/behavioral patterns. The analysis involved merging multiple datasets, including sales, store, product, oil price, and holidays, so as to add additional context to the models. Dimensionality reduction was performed with principal component analysis (PCA) to improve processing speed and retain as much of the most important data as possible. The final step in the overall model development effort was to split the final dataset into training and testing sets and conduct model development with RFs and neural networks. The overall outcome of the analysis is that, without thorough preprocessing of the data and undertaking feature engineering, it is not possible to develop accurate, reliable, and scalable demand forecasting models.

(All the developing steps have been added as Figs. A1 to A6 in Appendix).

The two models are implemented on the dataset for all the records to create Fig. 3.

The final unsupervised algorithm (clustering) was then performed where the data were partitioned into three clusters grouping similar data points together based on the selected features. The resulting cluster assignments are added to the Data Frame, allowing for further analysis of the clustered data. To summarize and analyze the clusters, the code identifies numeric columns and calculates the mean values of these columns for each cluster. This provides a detailed summary of the characteristics of each cluster, highlighting the differences and similarities between them. The cluster analysis helps in understanding the underlying patterns within the data, such as distinct store types or sales behaviors. The data were segmented into three clusters based on their sales demand and with consideration to other factors such as holidays or store location, for example. This unsupervised model ensured that the data do not need to be trained in order to provide the forecast and can take into consideration several other factors which also make it more efficient and advanced. The number of clusters (K = 3) was selected based on the elbow method and supported by silhouette analysis. Figure 4 shows the clustering model.

Fig. 4. Clustering model.

The model produced different clusters based on different attributes and showed how those attributes affected the sales demand. In the next chapter, the interpretation and evaluation of the results will be further discussed. To improve experimental robustness, k-fold cross-validation was applied, and hyperparameter tuning was conducted using grid search techniques. This ensures model stability and reduces overfitting.