an end-to-end crisp-dm machine learning pipeline for forecasting demand in fmcg chain stores

Objective: accurate forecasting of customer demand is necessary to optimize the efficiency of a supply chain, maximize profits through reduced inventory costs, and increase customer satisfaction. this research presents a new machine learning methodology based on the crisp-dm for customer order forecasting that is both interpretive and interpretable and validates it with a real-world application from the ofogh kourosh company, which offers the largest number of physical retail locations in iran.methods: the dataset analyzed for this research contained 844,275 sales transactions from 40 separate physical locations. six advanced ensemble machine learning models were developed to forecast customer order demand. a beneficial factor of this research was the ability to automate hyperparameter tuning of the six predictive models using the optuna framework. the performance of the predictive models was then evaluated using mae, rmse, mse, and r² metrics.results: based on r² score, lightgbm was the most accurate predictive model with an r² score of 0.536. feature importance analysis from lightgbm demonstrated that the three factors that would most determine customer order demand were the percentage of discount, price, and store location.conclusion: this research contributes both theoretically and practically to the development of a forecast model that is regionally, culturally, and contextually relevant within the iranian retail marketplace. compared to the literature, this study uses actual transactional data with ml models to narrow the theory-practice gap. future research should emanate from this development, incorporating external influences such as climate, advertising, and macroeconomic influences for even greater forecast accuracy