INTRODUCING SATELLITE DATA INTO WITHIN-YEAR NOWCASTING OF ANNUAL GDP GROWTH

This study examines whether satellite derived controls can improve within-year nowcasts of annual GDP growth across countries when traditional high frequency macroeconomic indicators are limited or not consistently available. Using a five-country sample consisting of the United States, China, Turkey, South Africa, and Ethiopia, I combine nighttime lights, atmospheric NO₂, and lagged structural controls within an expanding-window recursive forecasting framework to update estimates of current-year GDP growth in purchasing power parity terms. The analysis compares several approaches, including a lag-only benchmark, lag plus controls, lag plus satellites, a full model that combines controls and satellites, a factor model using GAUL Level 1 satellite data, and a spatiotemporal graph neural network (ST-GNN). The results show that most conventional approaches do not improve on the lag-only benchmark. Models that add structural controls, national level satellite variables, or satellite factors generally increase error across the full sample, even when they provide modest gains in selected cases such as the United States. The ST-GNN delivers the strongest overall pooled performance, producing the best results across the full sample. It records the lowest overall error and performs especially well in China, South Africa, and Ethiopia, while also improving on the lag-only benchmark in Turkey. Overall, the findings suggest that satellite data is most useful when modeled in a framework that preserves subnational spatial relationships rather than relying on national aggregation or dimensionality reduction through factors.