Methods for genetic risk prediction have been widely investigated in recent years. However, most available training data involves European samples, and it is currently unclear how to accurately predict disease risk in other populations. Previous studies have used either training data from European samples in large sample size or training data from the target population in small sample size, but not both. Here, we introduce a multi-ethnic polygenic risk score that combines training data from European samples and training data from the target population. We applied this approach to predict type 2 diabetes (T2D) in a Latino cohort using both publicly available European summary statistics in large sample size and Latino training data in small sample size. We attained a >70% relative improvement in prediction accuracy (from R2=0.027 to R2=0.047) compared to methods that use only one source of training data, consistent with large relative improvements in simulations. We observed a systematically lower load of T2D risk alleles in Latino individuals with more European ancestry, which could be explained by polygenic selection in ancestral European and/or Native American populations. Application of our approach to predict T2D in a South Asian UK Biobank cohort attained a >70% relative improvement in prediction accuracy, and application to predict height in an African UK Biobank cohort attained a 30% relative improvement. Our work reduces the gap in polygenic risk prediction accuracy between European and non-European target populations.