Synthetic time series for multi series demand forecasting

Global feature importance methods are one of the core tools for interpreting the role of explanatory variables in machine learning models,however, using them more complex forecasting tasks involving multiple time series can be challenging.This study focuses on tree-based ensemble models, applied to multi-series product demand forecasting.To evaluate the different global feature importance methods, we generate simulated datasets with controlled dependencies onlagged values and external demand drivers.We compare model-specific and model-agnostic global importance methods, including SHAP values, permutation importance, and tree-specific gain- and split-based importance.Our analysis focuses on uncovering pitfalls in applying these methods, including problems introduced by auto-correlation and feature scaling.