A model with one annual lag reduced squared forecast error by 7% in 432 MASTREE+ records with equal study weights

Past seed output can help forecasts of forest resources. We tested simple models in 432 MASTREE+ V2 seed, fruit, and cone records with at least 25 consecutive years. The records cover 187 plant taxa, 62 source studies, and 21 countries. We used the first 15 years for the initial fit. We predicted each later year from earlier observations only and updated the fit each year. This gave 10,675 test forecasts. A linear model with one annual lag (AR1) reduced mean squared error by 7.2% relative to the mean of all earlier years (95% source-study bootstrap interval 3.6% to 10.8%). Each source study had equal weight. The median record gain was 1.0%; 240 of 432 records had lower error. The gain persisted with a 20-year initial fit, square-root output, equal taxon weights, and removal of each source study. The planned primary model used two lags (AR2). Its mean error reduction was -0.4% (-6.6% to +5.0%), so this test did not establish a gain. One unstable two-lag fit caused a large error. These results support AR1 as a modest history-only benchmark. They do not test weather models or show that two-year biological memory is absent. Source studies can share weather and taxa, so study bootstrap intervals may be too narrow. The package preserves the unchanged CC BY 4.0 source data, 10,675 forecasts, selection records, code, provenance, checks, and figures.

mast seeding · plant reproduction · ecological forecasting · MASTREE+ · seed production · out-of-sample prediction · autoregression