Zero-Shot Forecasting in Emerging Economies: A Comparison of Time Series Foundation Models and SARIMA on Ghanaian Macroeconomic and Climatic Data

Abstract In developing nations like Ghana, forecasting is essential for risk management, investment, and policy, especially in commodity industries that are vulnerable to market and climate fluctuations. While Time Series Foundation Models (TSFMs) such as TimeGPT and TimesFM offer cutting-edge zero-shot forecasting capabilities, traditional models (ARIMA, SARIMA) are affected by nonlinearity and structural breaks. This study uses datasets of macroeconomic indicators and climatic variables from Ghana to evaluate TSFMs against SARIMA. No prior study simultaneously benchmarks these four models. The findings demonstrate that SARIMA outperformed TSFMs in linear, stationary series such as exchange rates (RMSE: 1.262) and interest rates (RMSE: 2.642), while TimeGPT excelled in forecasting volatile inflation. Intermittent precipitation trends were successfully captured by TSFMs, indicating their potential usefulness in data-scarce resource-constrained settings. The study acknowledges key limitations including the absence of SARIMA hyperparameter sensitivity analysis and the lack of uncertainty quantification for TSFM outputs. To address these challenges, the study suggests improving TSFMs through multivariate approaches and domain-specific fine-tuning.