The main vehicle for this study is the FSU global spectral model. The first part of this study entails a 30-member ensemble forecast with this model that includes the standard components of dynamics, physical parameterization, orography and normal mode initialization. The initial fields are subjected to EOF (Empirical Orthogonal Function) based perturbations. A thirty-member ensemble is constructed using this procedure. Following studies on singular vectors and error growth, we calculate an error variance field at day two of forecasts for the 30-member ensemble with respect to one member designated as the control run. These were the error variances of the precipitation forecasts at day two. These error variances were next correlated against model based initial and 12-hour forecast precipitation fields. Regions were identified which had the highest correlations. The adaptive TRMM-based precipitation data over these select regions were used to perform a physical initialization at a high-resolution global model. We show improvements in the precipitation forecast skill over the global tropics from this adaptive use of TRMM observations