The feasibility of reactor noise analysis by autoregressive (AR) modeling is studied from the viewpoint of system identifiability. A condition is derived in which only a part of the identified model becomes meaningful. A practical checking method termed “RRV checking” is proposed, with which the occurrence of the condition is recognized a posteriori for the estimated AR model. This method is applied to AR models obtained by processing the experimental data from the Japan Power Demonstration Reactor II. These models would have been discarded from a conventional viewpoint, since some parts of the model showed physically implausible characteristics. It is verified that the RR V checking method and the empirical evaluation of the usability of the model resulted in the same conclusion about the acceptability of the parts of the models. The processes evaluated to be identifiable from the reactor noise are the response of the fuel temperature to the neutron density and the response of the steam control valve to the reactor pressure. The present method is particularly useful if a priori knowledge about the dynamics of the objective process is limited before the identification experiment.