Abstract A simulation study was conducted to evaluate the performance of genomic random regression models for the continuous environmental descriptor temperature-humidity index (THI). Statistically innovative aspects of the study included the combined simulation of both longitudinal phenotypic data representing the same trait in the course of THI and genomic data. The longitudinal trait was simulated (phenotypically expressed) at 5 different values of THI. For a moderate heritability trait, heritabilities were 0.30, 0.35, 0.40, 0.40, and 0.35 for THI of 15, 30, 45, 60 and 75, respectively. In a consecutive run, low heritabilities of 0.05, 0.1, 0.15, 0.15, and 0.10 were simulated, respectively. On the genomic level, simulation combined high and low linkage disequilibrium with 5,000-, 15,000-, and 50,000-SNP chip applications to simulate different scenarios of genomic architecture. With regard to data analyses, 2 strategies were applied to evaluate the accuracy of genomic predictions across THI, with special focus on the extreme ends of the environmental scale. In the first strategy, 100, 80, 50, or 20% of phenotypes at THI 75 were deleted randomly and the remaining data set was used to predict the breeding value at THI 75 for non-phenotyped, but genotyped cows. In the second strategy, 1,600 cows had complete information (genotypes and phenotypes) and 400 cows were genotyped, but with missing phenotypes for all THI. For the first strategy and without phenotypic observations at THI 75, accuracies of genomic predictions were lower than 0.34. When only 20% of cows had phenotypic records at THI 75, accuracies increased (~0.60). Such a small proportion of phenotyped cows was sufficient to predict reliable genomic breeding values for cows without phenotypes for extreme THI. For the second strategy, also for low linkage disequilibrium combined with a low density 5,000-SNP chip, the average accuracy of genomic predictions was 0.52, which is substantially higher than accuracies based on pedigree relationships. From a practical perspective, genomic random regression models can be used to predict genomic breeding values for scarce phenotypes (e.g., novel traits) traits measured in extreme environments, or traits measured late in life, such as longevity.