DFROG — De Nederlandsche Bank's Machine-Learning-Tested Nowcasting Model for GDP Growth

De Nederlandsche Bank (DNB) operates DFROG (Dutch Forecasting model for Real-time Output Growth), a dynamic-factor-model nowcasting tool combining roughly 70 monthly economic and financial indicators into a single real-time estimate of current-quarter Dutch GDP growth, published monthly on DNB's website as a 'mechanical projection' distinct from the bank's official forecasts.

Provide a real-time, model-based nowcast of current-quarter GDP growth alongside DNB's official June and December forecasts.

DNB researchers evaluated DFROG's accuracy over 2013 Q3–2023 Q3 against professional forecasters (Consensus Economics), naive benchmarks (random walk, autoregressive model), and the Bańbura–Rünstler dynamic factor model (Working Paper 819). A related earlier paper (754, peer-reviewed in De Economist and AStA Advances in Statistical Analysis) tested machine-learning methods (random forest, LASSO, elastic net, random subspace regression) against DFROG's approach.

Excluding the three most volatile COVID-19 quarters, DFROG beat Consensus Economics backcasts by a 46% RMSFE margin and naive benchmarks by up to 38%; it was 11% and 16% more accurate than the Bańbura–Rünstler benchmark in the first and second month of the quarter respectively (though statistically indistinguishable over the full COVID-inclusive sample). A random forest model modestly outperformed the factor model at the nowcast horizon (0.88 vs 0.90 relative RMSFE) and one quarter ahead (0.92 vs 0.97).

DNB is candid that DFROG is a mechanical statistical tool, not an official forecast, and that professional forecasters' judgement still beats it during exceptional distress like COVID-19; despite ML methods showing modest gains, DFROG's simpler factor-model design remains DNB's production choice.