Integrating demand forecasting & inventory management in food manufacturing industry : hierarchical forecasting approach

Abstract

This study focuses on production planning in the food manufacturing industry utilizing hierarchical forecasting in order to enhance demand prediction for products that share a common core ingredient. Creating strategic supply chain techniques that maximize resource allocation and production efficiency requires accurate demand predictions for this ingredient. The study aims to increase forecast accuracy in order to improve predictions at various operational levels. The study considers two scenarios: (1) forecasting the aggregated total demand for the common ingredients of all five products and (2) forecasting the demand for the common ingredients of each product individually and then summing the results. Following an exploratory data analysis to understand the dataset's characteristics, both statistical and machine learning (ML) methods are utilized. The statistical methods, including SARIMA and Prophet, serve as benchmarks, while ML models, including LR, SVR, RF, and XGBoost, are employed to enhance forecasting accuracy. Reconciliation methods such as bottom-up, MinT, and OLS are evaluated, with the base forecast serving as a benchmark. The dataset exhibits sudden spikes and drops, where the spikes impose an additional penalty cost and adversely impact forecasting accuracy. To address this, the study adopts a dual approach: retaining the original dataset in one instance and creating an anomaly-free dataset in another. Results demonstrate that the XGBoost model, combined with MinT reconciliation, outperforms other reconciliation techniques across all products and hierarchical levels. The forecasting process was applied to two datasets: one with the original data and the other after anomaly removal. This research underscores the critical role of data preprocessing in the forecasting process, particularly the value of anomaly detection in improving forecasting accuracy. Key findings reveal that forecasting the aggregated demand for the common ingredients of all five products outperforms the individual forecasting and summation approach for the selected industry case. Moreover, this study aids a middle-scale local food manufacturer by integrating academic insights with industry practices. The results have significant implications for enhancing production efficiency, reducing inventory costs, and improving supply chain resilience in the food and beverage sector.