Macroeconomic forecasting with machine learning in a small open economy

Macroeconomic forecasting in small open economies is challenging because samples are short, external shocks are important, and predictor sets are often high dimensional. This paper evaluates whether machine learning methods improve forecasts of Slovak industrial production relative to standard time series and factor based benchmarks. Using rolling pseudo real time forecasts at horizons of 1, 3, 6, 9, and 12 months, we compare regularized linear models, tree based methods, support vector machines, factor based approaches, and several forecast combination schemes. The analysis considers three settings: a pre-COVID hard data sample, a full sample hard data setting that includes the COVID shock and recovery period, and a sample based only on soft survey indicators. Across the hard data specifications, Lasso and Elastic Net consistently deliver the strongest out of sample performance in terms of both forecast accuracy and directional accuracy, while discounted mean squared forecast error weighting is the most competitive combination approach. These gains remain strong in the sample that includes the COVID period, indicating robustness under heightened volatility and structural disruption. By contrast, soft indicators used on their own provide only limited and unstable forecasting gains. The results show that, in a small open economy, sparse regularization offers a more effective forecasting strategy than broader shrinkage, factor compression, or standard univariate benchmarks.