A method is proposed to calculate measures of forecast skill for high, medium and low temporal frequency variations in the atmosphere. This method is applied to a series of 128 consecutive 1 to 10-day forecasts produced at NMC with their operational global medium-range-forecast model during 1 May–5 September 1988. It is found that over this period, more than 50% of the variance in observed 500 mb height fields is found at periods of 18 days or longer. The. intuitive notion that the predictability time of a phenomenon should be proportional to its lifetime is found to be qualitatively correct; i.e., the low frequencies are predicted (at a given skill level) over a longer time than high frequencies. However, the current prediction skill in low frequencies is far below its potential if one assumes that for any frequency the predictability time scale ought to be equal to the lifetime scale. In the high frequencies, however, the current prediction skill has already reached its potential; i.e., cyclones are being predicted over a time comparable to their lifetime; i.e. 3 to 4 days. We offer some speculations as to why the low frequency variations in the atmosphere are so poorly predicted by our current state-of-the-art models. The conclusions are tested, and found to hold up, on a more recent dataset covering 10 December 1988–16 April 1989.