Past local deaths reduced squared forecast error by 4.4% against a historical baseline for 20 m cells at Barro Colorado, 2000–2015

Does a past local death map improve later forest forecasts? We tested seven open tree censuses at Barro Colorado Island, Panama. We fitted models on cell outcomes through 2000 and froze their coefficients. We then made three one-interval forecasts for 2000–2015, with updated historical inputs. The baseline used past species-size death probabilities and a past 100 m block residual. Adding the past 20 m cell residual reduced test mean squared error by 4.38% across 3,750 cell-interval records. A spatial block bootstrap interval was 2.60% to 5.84%, conditional on the observed times and fixed historical feature maps. RMSE fell from 4.401 to 4.304 percentage points. Error fell in each test interval. Larger spatial blocks, two smoothing checks and strict live statuses kept positive gains. For trees at least 10 cm in diameter, the gain was 0.33%, with an interval that included zero. Both models overpredicted the mean later death fraction. The test supports a small added forecast value for local history in this plot and baseline. It does not identify a cause, establish a carbon-loss forecast, or show transfer to other forests. The package preserves exact source columns, all forecasts, code, provenance, figures and checks.

tree mortality · forest dynamics · chronological prediction · spatial dependence · Barro Colorado Island · open data · reproducibility