A new recursive attitude prediction method is derived for spacecraft equipped with attitude sensors, such as star cameras. The attitude predictor model is given in an autoregressive form (finite-difference model) and its model parameters are estimated by the prediction error approach with a quadratic norm. In this approach, concerns about system uncertainties (such as unknown external torque, inertia matrix, or initial conditions) are avoided by utilizing a kinematics-based model. Angular velocity is estimated as a by-product of the attitude estimation algorithm. An example illustrates that the recursive attitude predictor estimates attitude trajectories accurately, for the case of a slowly rotating spacecraft, for a significant time interval without attitude measurements, which allows attitude to be predicted during intervals following occasional data dropout.