Evaluating Machine Learning and Econometric Models for Inflation Forecasting: Evidence from a Sub-Saharan African Panel

Policymakers in Sub-Saharan Africa increasingly face decisions regarding the adoption of machine learning methods for inflation forecasting, yet these approaches are rarely benchmarked against simple persistence models. This study constructs a reproducible, annual panel of macroeconomic indicators for eight Sub-Saharan African economies (1972–2024, N  = 424) to compare eight forecasting approaches: a naive random-walk benchmark, two linear econometric models (pooled OLS and country fixed effects), three machine learning models (Elastic Net, Random Forest, and XGBoost), and a recurrent neural network (LSTM) evaluated with and without early stopping. Using a chronological train/validation/test split and pairwise Diebold-Mariano tests, we evaluate out-of-sample predictive accuracy. The empirical results indicate that no estimated model statistically outperforms the naive persistence baseline. While Random Forest, XGBoost, Elastic Net, and the early-stopped LSTM achieve statistical parity with the random walk, traditional linear models and the uncorrected LSTM are significantly dominated by it. Although implementing early stopping substantially improves LSTM performance, it does not yield a competitive advantage over simpler tree-based ensembles. These findings align with the well-documented random-walk puzzle in global inflation forecasting, suggesting that structural breaks, high volatility, and modest sample sizes pose substantial challenges for highly parameterized forecasting architectures. Consequently, these results suggest that regional central banks and finance ministries should maintain simple persistence models as active baselines when evaluating investments in complex forecasting infrastructure. These conclusions are robust to a battery of additional checks: HAC-Newey-West/Harvey-Leybourne-Newbold-corrected and Holm-Bonferroni-adjusted Diebold-Mariano tests, a strictly lagged-only predictor respecification, expanding-window recursive re-estimation, per-country and subperiod decomposition, and LSTM input-sequence-length sensitivity analysis. JEL Classification: E31, E37, C53, C45, C33, O55