Published Paper Details:

Accurate demand forecasting is critical in the food industry to minimize waste and optimize inventory management, particularly for perishable products. This study uses the "Food Demand Forecasting" dataset by Genpact to compare 16 forecasting models, including traditional machine learning (Random Forest, Gradient Boosting, LightGBM, XGBoost, CatBoost, ElasticNet, SVR), statistical methods (Prophet, TBATS, ARIMA), deep learning models (LSTM, BiLSTM, TCN, ETSformer, GNN), and a hybrid LSTM + XGBoost approach. We applied advanced feature engineering, such as lag features, Exponentially Weighted Moving Average (EWMA), and cyclical encoding of temporal features, to capture seasonality and trends. The primary focus was on achieving high accuracy, with the Hybrid LSTM + XGBoost model achieving the highest accuracy of 93.47%, followed closely by the BiLSTM model at 91.84%. The TCN model achieved a competitive accuracy of 90.66%, demonstrating its effectiveness in capturing temporal patterns. This study improves upon prior work by Panda and Mohanty (2023) by introducing a broader model set, explicit seasonality modeling, and a focus on maximizing prediction accuracy. We provide detailed reasoning for model selection, feature engineering, and evaluation metrics, emphasizing the practical implications of high accuracy in food supply chain management.