Identification of individuals at highest risk of coronary artery disease (CAD)-ideally before onset-remains an important public health need. Prior studies have developed genome-wide polygenic scores to enable risk stratification, reflecting the substantial inherited component to CAD risk. Here we develop a new and significantly improved polygenic score for CAD, termed GPS, that incorporates genome-wide association data across five ancestries for CAD (>269,000 cases and >1,178,000 controls) and ten CAD risk factors. GPS strongly associated with prevalent CAD (odds ratio per standard deviation 2.14, 95% confidence interval 2.10-2.19, P < 0.001) in UK Biobank participants of European ancestry, identifying 20.0% of the population with 3-fold increased risk and conversely 13.9% with 3-fold decreased risk as compared with those in the middle quintile. GPS was also associated with incident CAD events (hazard ratio per standard deviation 1.73, 95% confidence interval 1.70-1.76, P < 0.001), identifying 3% of healthy individuals with risk of future CAD events equivalent to those with existing disease and significantly improving risk discrimination and reclassification. Across multiethnic, external validation datasets inclusive of 33,096, 124,467, 16,433 and 16,874 participants of African, European, Hispanic and South Asian ancestry, respectively, GPS demonstrated increased strength of associations across all ancestries and outperformed all available previously published CAD polygenic scores. These data contribute a new GPS for CAD to the field and provide a generalizable framework for how large-scale integration of genetic association data for CAD and related traits from diverse populations can meaningfully improve polygenic risk prediction.