In this paper, the late part of a room response is modeled in the frequency domain as a complex Gaussian random process. The autocovariance function (ACVF) and power spectral density (PSD) are theoretically defined from the exponential decay of the late reverberation power. Furthermore we show that the ACVF and PSD are accurately parametrized by an autoregressive moving average (ARMA) model. This leads to a new generative model of late reverberation in the frequency domain. The ARMA parameters are easily estimated from the theoretical ACVF. The statistical characterization is consistent with empirical results on simulated and real data. This model could be used to incorporate priors in audio source separation and dereverberation.