When identifying all modules in a dynamic network it is natural to treat all node variables in a symmetric way, i.e. not having pre-assigned roles of `inputs' and `outputs'. In a prediction error setting this implies that every node signal is predicted on the basis of all other nodes. A usual restriction in direct and joint-io methods for dynamic network and closed-loop identification is the need for a delay to be present in every loop (absence of algebraic loops). It is shown that the classical one-step-ahead predictor that incorporates direct feedt-hrough terms in models can not be used in a dynamic network setting. It has to be replaced by a network predictor, for which consistency results are shown when applied in a direct identification method. The result is a one-stage direct/joint-io method that can handle the presence of algebraic loops. It is illustrated that the identified models have improved variance properties over instrumental variable estimation methods.