Enhancing Supply Chain Decision-Making with Large Language Models: A Comparative Study of AI-Driven Optimization

A comparative analysis of LLMs versus traditional machine learning models across four key supply chain tasks

Sakib Salam Jamee; M. Hossain; Mahabub Hasan; Mohammad Kawsur Sharif; Md Sayem Khan; Md Iftakhayrul Islam; Shaidul Islam Suhan· International Journal of Economics Finance & Management Science· 2025· DOI 10.55640/ijefms/volume10issue04-02

The core problem

Supply chain decision-making increasingly relies on data-driven approaches to enhance efficiency, reduce costs, and improve responsiveness. Traditional machine learning models such as Random Forest (RF), Gradient Boosting Machines (GBM), Support Vector Machines (SVM), and Deep Neural Networks (DNN) have been widely applied to tasks like demand forecasting, supplier selection, inventory management, and logistics optimization. However, these models often struggle to integrate unstructured data (e.g., supplier feedback, market trends) with structured transactional records. Large Language Models (LLMs) offer a promising alternative due to their ability to process and analyze complex, multi-dimensional data from diverse sources. This study aims to evaluate the performance of LLMs relative to traditional models across four key supply chain tasks, providing insights into their potential to revolutionize supply chain management.

Innovation

The LLM significantly outperformed traditional models across all four tasks. In demand forecasting, the LLM achieved superior accuracy by effectively integrating transactional records with market trends. For supplier selection, the LLM demonstrated enhanced capability in analyzing supplier feedback and historical performance. Inventory management and logistics optimization also showed marked improvements, with the LLM reducing costs and improving efficiency. However, the LLM required more computational resources compared to traditional models. Quantitative results (e.g., accuracy, F_{1} -score) are summarized in the table below (values are illustrative based on the study's findings). | Task | RF | GBM | SVM | DNN | LLM | |------|----|-----|-----|-----|-----| | Demand Forecasting | 0.82 | 0.84 | 0.80 | 0.86 | **0.92** | | Supplier Selection | 0.75 | 0.78 | 0.73 | 0.80 | **0.88** | | Inventory Management | 0.79 | 0.81 | 0.77 | 0.83 | **0.89** | | Logistics Optimization | 0.77 | 0.79 | 0.75 | 0.82 | **0.90** | *Note: Accuracy values are for demonstration; actual metrics may vary.*

Supply chain decision-making increasingly relies on data-driven approaches to enhance efficiency, reduce costs, and improve responsiveness. Traditional machine learning models such as Random Forest (RF), Gradient Boosting Machines (GBM), Support Vector Machines (SVM), and Deep Neural Networks (DNN) have been widely applied to tasks like demand forecasting, supplier selection, inventory management, and logistics optimization. However, these models often struggle to integrate unstructured data (e.g., supplier feedback, market trends) with structured transactional records. Large Language Models (LLMs) offer a promising alternative due to their ability to process and analyze complex, multi-dimensional data from diverse sources. This study aims to evaluate the performance of LLMs relative to traditional models across four key supply chain tasks, providing insights into their potential to revolutionize supply chain management.

The study compares LLMs with four traditional machine learning models: Random Forest (RF), Gradient Boosting Machines (GBM), Support Vector Machines (SVM), and Deep Neural Networks (DNN). The evaluation focuses on four supply chain tasks: demand forecasting, supplier selection, inventory management, and logistics optimization. Data sources include transactional records, supplier feedback, and market trends, encompassing both structured and unstructured data. Performance metrics include accuracy, computational resource requirements, and overall decision-making effectiveness. The LLM is implemented using a state-of-the-art architecture (e.g., transformer-based) and fine-tuned on supply chain-specific data. Traditional models are trained on the same datasets for fair comparison. The experimental setup is illustrated in the Mermaid diagram below.

Why it matters

The results indicate that LLMs offer a substantial advantage in supply chain decision-making, particularly when tasks involve both structured and unstructured data. The LLM's ability to analyze complex, multi-dimensional data from sources such as transactional records, supplier feedback, and market trends enables more accurate and holistic decisions. This is especially valuable in dynamic environments where traditional models may fail to capture nuanced patterns. However, the higher computational cost of LLMs poses a challenge for real-time applications. Future research should focus on optimizing LLM efficiency, exploring hybrid approaches that combine LLMs with traditional models, and extending applications to broader supply chain contexts. The findings suggest that LLMs can revolutionize supply chain management by improving efficiency, reducing costs, and enhancing decision-making accuracy.

Who should read this

CS practitioners and researchers

Opening member content…