A model with the previous crop gave a geometric mean squared error ratio of 0.91 against past means in 631 MASTREE+ series

We tested whether the previous crop helps forecasts of annual plant reproduction. We used 631 long, complete seed, fruit and cone series from MASTREE+ version 2. They cover 99 studies, 229 species and source years 1874–2022. The first two thirds of each series formed the initial training period. We predicted each later year with earlier observations only. A linear model with one lag and a fixed shrinkage penalty gave a geometric mean squared error ratio of 0.906 against the mean of all earlier crops. The 95% study bootstrap interval was 0.876 to 0.937. We gave each study equal weight. Error fell in 74 of 99 studies and 363 of 631 series. The ratio was 0.943 in seed-trap records and 0.940 without European beech. Repeating the last crop gave a ratio of 2.082. The model did not clearly beat the earlier median on absolute error: ratio 1.021, interval 0.975 to 1.069. Thus the forecast gain depends on the loss function. These historical tests assume that earlier annual values were available before the target crop. They do not identify a biological cause or establish gains in unmeasured populations. The package preserves the raw source, provenance, plan, code, individual forecasts, figures and clean-repeat checks.

plant reproduction · mast seeding · ecological forecasts · MASTREE+ · chronological validation · seed production · open data