Expectiles have received increasing attention as coherent and elicitable market risk measure. Their estimation from heavy-tailed data in an extreme value framework has been studied using solely the Weissman extrapolation method. We challenge this dominance by developing the theory of two classes of semiparametric Generalized Pareto estimators that make more efficient use of tail observations by incorporating the location, scale and shape extreme value parameters: the first class relies on asymmetric least squares estimation, while the second is based on extreme quantile estimation. A comparison with simulated and real data shows the superiority of our proposals for real-valued profit-loss distributions.