In elections with three or more candidates, coordination among like-minded voters is an important problem. We analyse the trade-off between coordination and learning about candidate quality under different temporal election systems in the context of the U.S. presidential primary system. In our model, candidates with different policy positions and qualities compete for the nomination, and voters are uncertain about the candidates' valence. This setup generates two effects: vote splitting (i.e. several candidates in the same policy position compete for the same voter pool) and voter learning (as the results in earlier elections help voters to update their beliefs on candidate quality). Sequential voting minimizes vote splitting in late districts, but voters may coordinate on a low-quality candidate. Using the parameter estimates obtained from all the Democratic and Republican presidential primaries during 2000-12, we conduct policy experiments such as replacing the current system with a simultaneous system, adopting the reform proposal of the National Association of Secretaries of State, or imposing party rules that lead to candidate withdrawal when prespecified conditions are met.