SummaryPredicting restoration outcomes requires an understanding of the natural variability of ecosystem properties. A hierarchy of predictability has been proposed that ranks measures of restoration success from most‐to‐least predictable in the following order: vegetation structure > taxonomic diversity > functional diversity > taxonomic composition. This hierarchy has not been tested empirically, and the location within the hierarchy of trait‐based measures, such as community‐level trait means and variances, is not well understood.Our objective was to test the hierarchy of predictability in one of the longest running ecological restoration experiments in the westernUSA. We used linear mixed effects models to analyse changes in herbaceous biomass, species richness, two functional diversity (FD) indices, community‐weighted mean (CWM) traits and taxonomic composition among experimental restoration treatments from 1992 to 2014 in a ponderosa pine‐bunchgrass ecosystem. Restoration treatments included combinations of light or heavy tree thinning and no fire or repeated prescribed fire every 4 years to release the herbaceous understorey from overstorey competition.Herbaceous biomass and species richness were the two most predictable and least variable measures of success, whereas taxonomic composition exhibited the highest variability among plots through time. Trait‐based measures of FD tended to be more predictable and less variable than CWM trait values in this experiment. Both CWM trait values and FD were less variable among plots than taxonomic composition.Synthesis and applications. Ecosystem properties that are intrinsically more variable over space and time will often be the least predictable restoration outcomes. Restoration practitioners can expect vegetation structure, species richness and functional diversity to be more predictable and less variable than taxonomic composition, which can exhibit dynamic responses to restoration treatments over time. Monitoring dominant native and invasive species will always be important, but given the functional redundancy that can occur within communities, strict targets based on composition may rarely be met. Trait‐based metrics that integrate taxonomic composition into their calculation are less variable and potentially more meaningful for evaluating ecosystem responses. The hierarchy of predictability should be tested in a range of ecosystems to determine its generality.