Nowcasting GDP: a unified evaluation of high-dimensional models

We evaluate a suite of high-dimensional models for real-time nowcasting of US GDP growth using FRED-MD vintages. Our framework integrates mixed-frequency data through UMIDAS representations to compare econometric and machine learning methods under a unifying design. We assess model performance from the first quarter of 2015 to the last quarter of 2024. We find that, over the full evaluation window, sparse models deliver the best performance. However, if we exclude 2020, a year marked by a structural break due to COVID-19, dense alternatives yield the most accurate predictions. These results highlight the trade-off between stability and adaptability in macroeconomic forecasting: dense models are better suited to tranquil times, while sparse alternatives perform better under instability. Overall, we also find that forecast combination provides a consistently competitive alternative.