Forecasting Food Inflation in Africa: A Comparison of Hybrid Models

This study investigates hybrid models for forecasting food inflation rates across six African countries: Nigeria, South Africa, Cape Verde, Morocco, Botswana, and Libya. The stacking ensemble method with Random Forest (RF) as the meta-learner was utilized as a combination technique. For robustness, a sensitivity analysis was conducted by interchanging Random Forest with Support Vector Regression (SVR), and multistep forecasts were evaluated at 3, 6, and 12-month forecast horizons. Out-of-sample forecast results reveal that different hybrid models perform optimally across countries: ARIMA-PROPHET in Nigeria and Morocco, ETS-SVR in South Africa and Botswana, SVR-MLP in Cape Verde, and NNAR-PROPHET in Libya. The results of the meta-learner sensitivity analysis indicate model changes in the countries, except Botswana and Libya, where original hybrids remained optimal. Multistep forecasts further affirm the superiority of NNAR-PROPHET in South Africa and Cape Verde, and PROPHET-MLP in Morocco for longer-term forecasts, while in Nigeria, Botswana, and Libya, hybrid models combining classical and machine learning techniques performed best across all horizons.