Machine Learning in GDP Nowcasting: From Lagging Indicators to Leading Algorithms

This paper evaluates the performance of machine learning models for nowcasting U.S. GDP, comparing them to traditional econometric approaches in both fixed-horizon and rolling-horizon settings. We construct composite indicators from high-frequency macroeconomic data and assess models across varying forecast lead times. The results show that during periods of economic volatility, machine learning models—particularly a manually tuned Multi-Layer Perceptron—significantly reduce forecasting error compared to autoregressive benchmarks. In more stable periods, simpler models perform comparably. Horizon-stratified evaluation reveals that combining high-frequency indicators with autoregressive components delivers the most accurate forecasts in the critical 4–8 week window before official GDP releases. Our findings underscore the conditional value of machine learning techniques in improving real-time macroeconomic surveillance and decision-making.