Control of wave energy converters requires knowledge of some seconds of the future behavior of certain physical quantities, in order to approach optimality. That is why short time prediction of the oncoming waves is a crucial problem in the field of wave energy, whose solution could bring great benefits to the effectiveness of the devices and to their economical viability. This study is proposed as a preliminary approach to cope with this necessity, where wave forecasts are computed on the basis of past observations collected at the prediction site itself. Working on single point measurements allows the treatment of the wave elevation as a pure time series, so that a wide range of well established techniques from the stochastic time series modelling and forecasting field may be exploited. Among the proposed solutions there are some cyclical models, based on an explicit representation of the a priori knowledge about the real process. It is then shown how a lot simpler and more effective solution can be obtained through classical AR models, which are shown to be able to implicitly represent the cyclical behavior of real waves. As a comparison with AR models some results obtained with neural networks are also provided.