Temporal Robustness in Cross‐Country Gross Domestic Product Forecasting: Evidence From Classical, Regime‐Switching, and Ensemble Models

ABSTRACT This study evaluates whether the relative performance of gross domestic product forecasting models is temporally robust across economies and changing macroeconomic conditions. Using annual real gross domestic product data from the World Bank for 177 separately reported economies and territories, we conduct a common‐vintage pseudo‐out‐of‐sample evaluation over 1990–2024. Seven transparent benchmarks are compared under a recursive one‐step‐ahead protocol: a Naive no‐change forecast, ARIMA, exponential smoothing, two‐ and three‐state Gaussian Hidden Markov Models, a simple average, and an adaptive ensemble. Both combinations use the same five‐model pool, including the Naive benchmark. To prevent the scale of large economies from dominating the comparison, the principal results use economy‐specific RMSE and MAE ratios relative to the corresponding Naive forecast, together with medians, geometric means and rank distributions. The adaptive ensemble records the lowest geometric‐mean relative RMSE (0.762) and relative MAE (0.641), whereas the simple average records the lowest median relative RMSE (0.813), the best mean rank (2.52), and the largest top‐two share (58.8%). After Benjamini–Hochberg adjustment, the simple average and adaptive ensemble significantly outperform the Naive benchmark in 64 and 47 economies, respectively, at the 5% level. In a balanced sample of 149 economies, adaptive weighting minimizes geometric‐mean relative RMSE in 1990–2019, while the simple average leads in 2015–2024 and in the strictly non‐overlapping 2020–2024 window. Formal break tests show lower relative losses for all non‐Naive models in break‐affected economies. The two combinations nevertheless remain the strongest average performers, and the identity of the best individual model changes across stability groups. The evidence favors model combination as protection against model‐selection uncertainty, while also confirming the forecast‐combination puzzle: estimated adaptive weights provide only modest gains over a simple average.