Given that its forecast skill is comparable with operational coupled GCMs and it reproduces observed spatio-temporal statistics, the much simpler LIM is useful for diagnosis of predictability, which may be determined from its forecast signal-to-noise ratios. The state-dependence of potential LIM skill is assessed and shown to compare well both with the LIM and operational model realized skill. Further analysis is performed based on the fact that the eigenvectors of the LIM dynamical evolution operator separate into two distinct, but nonorthogonal, subspaces: an “internal” space governing the nearly uncoupled subseasonal dynamics, and a “coupled” space governing the strongly coupled longer-term dynamics. These subspaces arise naturally from the LIM analysis; no bandpass frequency filtering need be applied. The internal space eigenmodes typically have much shorter periods and e-folding time scales than the coupled space eigenmodes. Additionally, the MJO mostly lies in this internal space, whereas ENSO mostly lies in the coupled space. Anomalies that project onto the coupled space are shown to more predictable than those projecting onto the internal space, even for relatively “short” Weeks 3-4 leads. That is, ENSO is the majority contributor to overall tropical skill even on the Weeks 3-4 time scale, and provides almost all of it on longer S2S forecast leads.