Fig. 4: Model performance by simulated rate of thermal tolerance enhancement.

Comparison of the 13 bleaching prediction models (points), each with different rates of thermal tolerance enhancement and resulting bleaching heat stress (annual maximum DHW–degree heating weeks), shows: (a) model parsimony, as the inverse difference in deviance information criterion (DIC) between each model and the most parsimonious model; (b) model prediction skill; and (c) model prediction error, broken down into over-predictions (solid line) and under-predictions (dashed black line). Correct predictions, over-predictions and under-predictions are assigned by comparing predicted and observed bleaching severity scores. To achieve this, each predicted severity value (fitted values ranging from 0 to 1) is transformed to the closest severity score (none, mild, moderate, or severe, corresponding to values 0, 0.333, 0.667, or 1), and then compared against the corresponding true bleaching severity score of the initial observation. The most-likely historic rate of thermal tolerance enhancement from the most parsimonious model is highlighted in all panels (vertical dashed grey line). Source data are provided as a Source Data file.