Assessing the incremental value of VIIRS nighttime lights for prefecture-level GDP growth in China

Earth-observation nighttime lights are widely used to map subnational economic activity, yet whether strong cross-sectional fit translates into reliable measurement of annual local growth remains uncertain. Here we develop a forward, loss-based validation framework across 296 Chinese prefectures (2014–2024) using two VIIRS-derived series—EOG Annual VNL (with its continuation) and NASA Black Marble VNP46A4—to separate within-year spatial representation, stand-alone forward temporal transfer, and conditional incremental value beyond four prior-year GDP-history benchmarks. While log light levels strongly track nominal GDP (\(\:{R}²=0.730\--0.831\)), annual light changes align weakly with certified real GDP growth (\(\:{R}²=0.0005\--0.247\)). In pre-2020 forward evaluations, stand-alone light calibrations generally underperform GDP-history benchmarks. During 2016–2019, all eight pooled conditional incremental-skill estimates are negative (-0.0362 to -0.0016), with province-block mean-loss confidence intervals spanning zero, indicating no detectable positive gain from adding light to historical GDP. Conversely, conditional point estimates are positive across all benchmarks during the 2020–2021 stress period. In the 2022–2024 extension on common 296-prefecture support, pooled conditional skill is positive for the EOG-derived series (0.0319–0.1097) but negative for Black Marble (-0.2240 to -0.1127), a divergence concentrated in 2022 and robust across verified GDP-source definitions. These findings demonstrate that socioeconomic proxy validation must be tailored to the target estimand, conditioning information set, and temporal domain.