Abstract
<title>Abstract</title> <p> This study evaluates whether machine learning and deep learning models offer a genuine forecasting advantage over classical econometric benchmarks for Ghana’s headline (year-on-year) inflation, a question of direct relevance to a central bank operating an inflation-targeting regime in a small, commodity-exposed, currency-volatile economy. Using 280 months (January 2000–April 2023) of Bank of Ghana monetary and macro-financial data, we construct a strictly one-step-ahead forecasting matrix in which every predictor reflects information available one month before the forecast target, and compare nine models (a random walk and seasonal naive baseline, an AIC-selected ARIMA benchmark, two regularized linear models [Ridge, Lasso], two tree ensembles [Random Forest, XGBoost], a kernel method [Support Vector Regression], and a shallow feed-forward neural network) on a chronological train/validation/test split whose 42-month test window (October 2019–April 2023) deliberately spans Ghana’s 2022–23 inflation surge. Contrary to the presumption that more flexible models dominate, the ARIMA(3,1,0) benchmark achieved the lowest test-set RMSE (2.41 percentage points), narrowly ahead of the random walk (2.65) and Lasso (3.41). However, Diebold–Mariano tests show that ARIMA is not statistically distinguishable from either of these two closest competitors, while its advantage over every other model is statistically significant. Every nonlinear machine learning model underperformed both simple baselines out of sample, with several models (notably XGBoost and the neural network) showing a large gap between near-perfect training fit and poor test performance: a textbook overfitting signature accentuated by the regime shift embedded in the test period. We argue this is a genuine and policy-relevant finding rather than a failure of specification: headline inflation is highly persistent, and models that best exploit that persistence, rather than those with the greatest functional flexibility, forecast best through a structural break. The study details a leakage-free feature-engineering protocol, reports full hyperparameter-tuning and diagnostic results, and discusses the conditions, such as larger samples, richer nonlinear feature sets, or regime-aware architectures, under which machine learning might be expected to add value for inflation forecasting in a market such as Ghana’s. <bold>JEL Classification:</bold> C53, C45, E31, E37, E52, O55 </p>