Research Summary

Abstract:

Inflation is never merely a statistic. It is the visible surface of a complex, historically contingent phenomenon shaped by supply chains, monetary institutions, exchange-rate regimes, and the psychology of expectations. This paper asks whether machine learning (ML) algorithms— Random Forest, XGBoost, LASSO-regularised VAR, and Long Short-Term Memory networks—can replace or substantively outperform traditional Vector Autoregressive (VAR) frameworks for inflation analysis in Morocco, a small open economy whose price dynamics are most conditioned by import dependence, a managed exchange-rate peg, and thin financial markets. Drawing on quarterly Moroccan macroeconomic data (1990–2024) and a rolling out-of-sample evaluation framework, we document a regime-contingent pattern: ML models achieve statistically significant forecast improvements during structurally turbulent periods— most acutely during the 2022–2023 commodity price shock—while VAR retains decisive advantages in structural identification, causal inference, and the counterfactual policy exercises that central banks require. We conclude that ML complements but cannot replace VAR in the Moroccan monetary policy toolkit, and we sketch a hybrid forecasting architecture appropriate for Bank Al-Maghrib’s ongoing inflation-targeting transition.