In plants, genomic prediction is employed to predict agronomically relevant traits, such as crop yield, phenology, and the concentration of minerals and metabolites, which possess varying genetic bases. Selecting the most appropriate model for a trait's genetic effects distribution and the population's allele frequencies is crucial. Linear regression models are typically preferred for genomic prediction given the number of genome-wide markers are usually much greater than the number of observations. However, additive models may not suit all genetic architectures and ancestral populations. Machine learning (ML) is suggested to enhance genomic prediction by modelling complex biology, such as epistasis, though has underperformed in past studies. This study evaluates (1) genomic prediction model sensitivity to trait ontology and (2) the impact of ancestral population-specific selection on prediction and model choice. Examining 35 quantitative traits in Arabidopsis thaliana with ~1000 observations across Europe, we assessed penalised regression, random forest, and multilayer perceptron performance. Regression models were generally most accurate, except in some biochemical traits where random forest excelled. We link this result to the biochemical traits having more simple genetic architecture, than macroscopic traits. Moreover, macroscopic traits, especially related to flowering and yield were strongly correlated with population structure and demography, while molecular traits required individual marker resolution. The study thus highlights the relevance of ensemble approaches for simple molecular traits in crop breeding and underscores the need to consider differential ancestral selection when designing association panels and training populations.