A rule that used the last crop status reduced mean Brier loss by 1.9% against past frequency in 433 MASTREE+ records, with equal study weights

We tested whether the previous year's low crop status improves the probability forecast for the next year. We used 433 quantitative seed, fruit and cone records from MASTREE+ version 2. The records cover 62 source studies, 187 source taxon names and 21 countries. Each record has at least 25 consecutive years. Low crop means output at or below the 25th percentile of the first 15 years. This threshold stays fixed. A rule that uses the previous label, with a fixed weight toward past frequency, made 10,686 forecasts from earlier data only. Mean Brier loss fell from 0.21372 for past frequency to 0.20961 for the rule. With equal source-study weights, the relative reduction was 1.92%, with a nominal 95% study bootstrap interval of 0.11% to 3.78%. The rule improved 37 of 62 study means and 214 of 433 records. Its gain over a baseline that uses the last ten years was uncertain: 1.76%, with an interval of -1.52% to 4.88%. Taxon resampling and the log score also gave intervals that included no gain. Two binary lags did not improve the mean Brier score. The result is a small average gain for a fixed retrospective benchmark. It does not establish reliable warnings for each population, a biological mechanism or a global mean benefit. Relative crop labels do not measure ecological food shortage. The package gives the unchanged source data, origin and licence, the plan, all forecasts, scores, bootstrap draws, figures, code and independent checks.

plant reproduction · MASTREE+ · ecological forecasting · low crop · Brier score · probability forecast · time series · open data