Past bleaching gave no clear log loss gain in fixed models for 2,615 Reef Check surveys, 2012–2019

We tested whether a previous bleaching event improved the estimated risk at a later Reef Check survey. We used Global Coral-Bleaching Database version 2 and matched past surveys by site, depth and season. Models used only earlier years to estimate risk in 2,615 surveys at 672 sites in 34 regions from 2012 to 2019. An event meant at least 10% reported bleaching. A fixed logistic model used current heat, depth, season, latitude, year, region, past heat and elapsed time. Adding the previous event indicator changed site-weighted mean log loss from 0.261434 to 0.261424. The difference was -0.000009. Its nominal 95% interval from paired region and year-block draws was -0.01012 to +0.00728. The Brier score difference was +0.000410, with an interval of -0.002454 to +0.003373. All six planned sensitivity intervals for log loss included zero. A simple running-rate reference had lower point loss than the fitted environment models. Thus, the test did not show a clear gain from this additive binary history term. It does not establish equal performance or the absence of ecological memory. The intervals hold fitted models fixed. Survey selection, few positive events, surface heat data and model form limit inference. An independent implementation matched all source survey groups, history links and primary scores. The package preserves source data, provenance, code, all test probabilities, results, figures and offline repeat instructions.

coral bleaching · ecological memory · Reef Check · repeat surveys · temporal validation · forecast benchmark · open data · null result