A fitted small tree mixture reduced squared forecast error but increased total variation error for 2015 species shares at Barro Colorado Island

We tested a simple forecast of tree species shares in one 50-hectare forest plot in Panama. Small trees had diameters from 1 to less than 10 cm. Large trees had diameters of at least 10 cm. We fitted one mixture fraction with 1985–2000 counts. We then forecast 2015 large tree shares from the 2000 small and large tree shares. The fitted fraction was 0.0355. Relative to no change, squared error fell by 3.68%, but total variation error rose by 3.41%. Their 95% spatial strip resampling ranges were 0.59–6.49% lower and 2.04–6.07% higher. Squared Hellinger error rose by 60.2%. The squared and total variation error reversals remained after each single species removal. A mixture fitted to training total variation selected no change. A reference that repeated the earlier change reduced later total variation error by 31.0%. Restricting small trees to 5–10 cm changed the result. Only 2.94% of measured small trees in 2000 reached the large class by 2015. Their species shares differed from those of the initial small pool. This result extends earlier work on size distributions with a later forecast test. It is a retrospective test of one mixture rule and one forest interval. It does not test all regeneration forecasts. Spatial resampling does not estimate temporal forecast uncertainty. The package has compact source records, source hashes, code, forecasts, resampling draws, figures, and independent and repeat checks.

forest composition · ecological forecasting · tree size · Barro Colorado Island · held-out test · error measures · regeneration · reproducibility