A one-year lag model gives an uncertain pooled forecast gain in 334 MASTREE+ records of at least 30 years

We tested whether the previous year's plant reproductive output improves a one-year forecast. We used 334 MASTREE+ v2 seed, fruit, and cone records with at least 30 consecutive years, from 40 studies and 17 countries. Each model started with 20 training years. We made 7,847 forecasts with past observations only. A linear model with one annual lag reduced the pooled scaled mean squared error by 1.46% relative to the expanding historical mean. The 95% study bootstrap interval was -1.37% to 6.10%. The lag model reduced error in 31 of 40 study means. The mean of the 40 relative study gains was 6.13% (2.09% to 10.21%); this differs from the ratio of pooled losses. A square-root fit and score gave a 7.06% gain (3.09% to 11.81%), which concerns a different prediction target. Persistence had about twice the error of the historical mean. The package preserves the raw source, selection rules, every primary forecast, code, figures, provenance, and an independent numerical check. The selected long records do not form a random sample of plant populations. The results do not test weather models, identify a resource mechanism, or establish a world-wide mean gain.

plant reproduction · MASTREE+ · mast seeding · ecological forecasting · rolling forecast · study bootstrap · seed production · open data