In recent years, food supply chain has gained considerable attention as compared to other supply chain systems. This can be attributed to the fact that fresh food products are perishable items which inherently has very short shelf life. Further, manual estimation of demand of these products often leads to demand underestimation and overestimation, which adversely affects revenues of the retailer. Therefore, effective demand forecasting can help to reduce food wastage as well as financial losses. The primary objective of this research is to investigate advanced machine learning models such as  Random Forest Regressor, XGBoost, and Polynomial Regression, for improving demand forecasting accuracy. In this work, we have specifically considered highly perishable items such as ladyfinger and tomato. Performance of the proposed models are evaluated on six years of market data from Maharashtra, India, using the Mean Absolute Percentage Error (MAPE) metric. Findings indicate that the Random Forest Regressor achieves the highest accuracy, reducing forecasting errors and enabling better decision-making in inventory and resource management. The proposed approach provides valuable insights for stakeholders, including farmers, distributors, and retailers, to minimize waste, optimize inventory, and ensure sustainable supply chain practices