Bulletin-gap estimation of Brazilian soybean harvest progress with bounded-increment reconstruction

This study evaluates pre-release estimation of weekly state-level soybean harvest progress reported by CONAB in Brazil. We analyzed 1,142 observations from 11 soybean-producing states spanning five crop seasons, using historical CONAB bulletins, NASA POWER meteorology, NASA HLS summaries, and a sensitivity analysis based on a MapBiomas soybean mask. Direct cumulative prediction was compared with bounded-increment reconstruction under LOSO, rolling-origin, and season-forward evaluation. With identical lag-only inputs and 1,021 matched observations, the direct and bounded-increment formulations achieved MAEs of 2.93 and 1.94 percentage points, respectively. Weather and HLS variables did not provide stable improvements over the lag-only model. For this delayed cumulative indicator, bounded-increment reconstruction improved accuracy more than adding the evaluated weather or HLS variables.