Machine learning methods in macroeconomic forecasting: Preliminary results

The paper summarizes machine-learning (ML) methods most relevant to macroeconomics and assesses their performance in forecasting and nowcasting key macro indicators. Despite rapid methodological progress and a surge of publications over the past 25 years, gains in forecast accuracy with traditional statistical (economic, financial, and survey) data remain modest. ML models often outperform naïve and standard econometric benchmarks, but improvements are not always statistically significant and, when they are, may be too small to matter for practitioners once implementation costs are considered. We highlight several tasks where ML is already useful even with traditional data and stress that ML becomes indispensable with “big” and unstructured data.