Nowcasting GDP with Machine Learning — the Swiss National Bank's Head-to-Head Test Against 1,100+ Indicators

The Swiss National Bank (SNB) publishes real-time GDP forecasts ('nowcasts') to inform its quarterly monetary-policy decisions, traditionally relying on time-series models such as autoregressive and factor models.

In Working Paper 2024-06, 'Nowcasting GDP: What Are the Gains From Machine Learning Algorithms?', SNB economists Milen Arro-Cannarsa and Rolf Scheufele systematically tested whether newer machine-learning methods do better than the traditional benchmarks.

The study built a large mixed-frequency dataset of more than 1,100 time series, applying publication-lag adjustments and blocking to align monthly indicators with quarterly GDP. It benchmarked eight machine-learning approaches -- LASSO, ridge and elastic-net regression, bagging, random forests, gradient boosting and support vector regression (SVR) -- against three standard benchmarks: a univariate model, a forward-selection algorithm and principal components regression (PCR).

For the demanding post-Great-Recession sample, every machine-learning technique tested beat the univariate benchmark, with gains of up to 28% in out-of-sample root-mean-square error (RMSE). Ridge regression, elastic net and SVR were the strongest performers, significantly outperforming PCR, the most common benchmark used in central-bank nowcasting.

The paper is a rigorous, published backtest rather than a live operational deployment: the SNB has not stated that these specific models now drive its official nowcasts, and the authors frame the results as evidence for further methodological adoption rather than a finished production system.