MIT SCM Symposium 2024

As Artificial Intelligence (AI) and Machine Learning (ML) continue to advance, their application in supply chain demand forecasting has surged, significantly impacting the field. These ML techniques improve forecasting accuracy, ultimately supporting managerial inventory and planning decisions. In this project, we evaluate the underlying hypothesis that suggests the application of targeted exogenous variables as model features, would further improve the predictive accuracy of the popular ML algorithm “Prophet”. Provided with statistically normalized Fast Moving Consumer Goods (FMCG) sales data from the Indian Subcontinent, we investigated 5,184 data points from 2011 to 2023 across 36 products. We screened and shortlisted exogenous variables within economic, health, climate, and political areas and selected 10 for evaluation in the predictive model. Each product was cross-validated with variables built as univariate regressors (features), and it was found that the introduction of exogenous variables reduced the Mean Average Percent Error (MAPE) by as much as -40.93%. Furthermore, incorporating targeted exogenous variables together with Hyper-Parameter tuning (calibration), had astounding results, with MAPE decreases by as much as -143.48%. This study proves targeted exogenous variables beneficial for practical application in improving forecasting accuracy within the FMCG industry.