The predictability of weekly averaged circulation anomalies in the Northern Hemisphere, and diabatic heating and moisture convergence anomalies in the Tropics, is investigated in a linear inverse model (LIM) derived from their observed simultaneous and time-lag correlation statistics. Throughout the year, the model's forecast skill at Week 2 (Days 8 to 14) and Week 3 (Days 15 to 21) is comparable to that of a comprehensive global medium range forecast (MRF) model developed at the National Centers for Environmental Prediction (NCEP). Its skill at Week 3 is actually higher on average, partly due to its better ability to forecast tropical heating variations and their influence on the extratropical circulation. The geographical and temporal variations of forecast skill are also similar in the two models. This makes the much simpler LIM an attractive tool for assessing and diagnosing atmospheric predictability at these forecast ranges.

The LIM assumes that the dynamics of weekly averages are linear, asymptotically stable, and stochastically forced. In a forecasting context, the predictable signal is associated with the deterministic linear dynamics, and the forecast error with the unpredictable stochastic noise. In a low-order linear model of a high-order chaotic system, this stochastic noise represents the effects of both chaotic nonlinear interactions and unresolved initial components on the evolution of the resolved components. Its statistics are assumed here to be state-independent.

An average signal to noise ratio is estimated at each grid point on the hemisphere, and is then used to estimate the potential predictability of weekly variations at the point. In general, this predictability is on the order of 50% higher in winter than summer over the Pacific and North America sectors; the situation is reversed over Eurasia and North Africa. Skill in predicting tropical heating variations is important for realizing this potential skill. The actual LIM forecast skill has a similar geographical structure but weaker magnitude than the potential skill. Results are similar for forecasts of diabatic heating and moisture convergence anomalies. Both predictability and skill are lowest in fall.

In this framework, predictable variations of forecast skill from case to case are associated with predictable variations of signal rather than of noise. This contrasts with the traditional emphasis in studies of shorter-term predictability on flow-dependent instabilities, i.e. on the predictable variations of noise. In the LIM, the predictable variations of signal are associated with variations of the initial state projection on the growing singular vectors of the LIM's propagator, which have relatively large amplitude in the tropics. At times of strong projection on such structures, the signal to noise ratio is relatively high, and the Northern Hemispheric circulation is not only potentially but also actually more predictable than at other times.