Improving the Accuracy and Interpretability of Sales Forecasts in ERP Systems using Predictive Analytics with Machine Learning and Large Language Models

The manufacturing industry's efficiency hinges on the seamless integration of production, inventory, and distribution processes. Enterprise Resource Planning (ERP) systems are the cornerstone of this integration, managing core functionalities from inventory to customer data. The next evolutionary step for these systems is the integration of Predictive Analytics, powered by Machine Learning (ML) and, more recently, Large Language Models (LLMs). This paper investigates the confluence of these technologies to enhance sales forecasting within ERP systems. We synthesize existing research on Machine Learning Models, which includes SARIMA, XGBoost, LSTM, and Random Forest for tasks like revenue prediction, channel-specific sales, and seasonal trend forecasting. The literature consistently indicates that ensemble methods like XGBoost and deep learning models like LSTM often achieve high accuracy by capturing complex, non-linear patterns in historical data. However, a major challenge persists due to the “black box” nature of these sophisticated models, which often limits transparency and can impede user trust and widespread adoption. 

 To address this, we propose a novel framework that integrates LLMs not for prediction, but for interpretability. This framework uses statistical techniques like Pearson correlation for robust feature selection and then leverages LLMs to translate model outputs into natural language summaries and actionable business recommendations. This synergistic approach bridges the critical gap between raw analytical output and strategic, interpretable decision-making, charting a course for the next generation of intelligent, user-centric ERP systems.