The authors address the problem of measuring asynchronous transfer mode (ATM) traffic on the cell, burst and activity level. With the aid of a traffic sampling mechanism it is shown that efficient data can be collected. These data are then used to estimate the spectral density on each level. It is shown that a forward-backward least square spectral estimator with prefiltering gives an accurate estimation even at low frequencies and that the power spectral density at frequency zero and the load are the most dominant parameters for the queue length and hence also for loss.>