Forecasting Quarterly GDP Growth and the GDP Deflator in Albania under Data Scarcity: A Comparative Evaluation of Statistical and Machine Learning Models

The purpose of this paper is to evaluate forecasts of Albania’s Real GDP Growth and GDP Deflator under real-time operational constraints by conducting an operational backtest using the leakage-free rolling origin approach and mixed-frequency predictors. The main difficulty is the small number of observations in the quarterly data set, partially observed predictors, and a 90-day publication delay for GDP. In this study, we evaluate 19 models grouped into six families: baseline univariate rules, classical time-series models, dynamic regression/SARIMAX specifications, regularised linear models, bridge/factor-augmented models, and selected nonlinear/probabilistic machine-learning models. These approaches were applied individually to Real GDP Growth and the GDP Deflator to generate point forecasts and forecast intervals for 1-, 2-, and 4-quarter-ahead horizons, using metrics such as MAE, RMSE, interval accuracy and coverage, interval width, and CRPS. Robustness was assessed by conducting Diebold-Mariano tests, applying the block bootstrap, and treating the period after 2016 as an out-of-sample period. The three highest-ranking models are univariate rules that converge toward the expanding-window in-sample mean. This suggests mean reversion in GDP growth and persistence in the GDP deflator. Among the structured models, Bridge_PCA_Ridge is the most consistent option at the 1- and 2-quarter horizons, while SARIMAX is the strongest structured specification at H=4. Selected nonlinear machine-learning models did not improve upon the regularised linear alternatives in this sample.