We explore the prediction of individuals' phenotypes for complex traits using genomic data. We compare several widely used prediction models, including Ridge Regression, LASSO and Elastic Nets estimated from cohort data, and polygenic risk scores constructed using published summary statistics from genome-wide association meta-analyses (GWAMA). We evaluate the interplay between relatedness, trait architecture and optimal marker density, by predicting height, body mass index (BMI) and high-density lipoprotein level (HDL) in two data cohorts, originating from Croatia and Scotland. We empirically demonstrate that dense models are better when all genetic effects are small (height and BMI) and target individuals are related to the training samples, while sparse models predict better in unrelated individuals and when some effects have moderate size (HDL). For HDL sparse models achieved good across-cohort prediction, performing similarly to the GWAMA risk score and to models trained within the same cohort, which indicates that, for predicting traits with moderately sized effects, large sample sizes and familial structure become less important, though still potentially useful. Finally, we propose a novel ensemble of whole-genome predictors with GWAMA risk scores and demonstrate that the resulting meta-model achieves higher prediction accuracy than either model on its own. We conclude that although current genomic predictors are not accurate enough for diagnostic purposes, performance can be improved without requiring access to large-scale individual-level data. Our methodologically simple meta-model is a means of performing predictive meta-analysis for optimizing genomic predictions and can be easily extended to incorporate multiple population-level summary statistics or other domain knowledge.