Time series with systematic misses occur often in practice and can be modeled as amplitude modulated ARMA processes. With this as a motivating application, modeling of cyclostationary amplitude modulated time series is addressed in this paper. Assuming that the modulating sequence is (almost) periodic, parameter estimation algorithms are developed based on second- and higher order cumulants of the resulting cyclostationary observations, which may be corrupted by any additive stationary noise of unknown covariance. If unknown, the modulating sequence can be recovered even in the presence of additive (perhaps nonstationary and colored) Gaussian, or any symmetrically distributed, noise. If the ARMA process is nonGaussian, cyclic cumulants of order greater than three can identify (non)causal and (non)minimum phase models from partial noisy data. Simulation experiments corroborate the theoretical results.