We propose a program of research aimed at clarifying what kinds of decadal variability are generated by climate models, what are their mechanisms, and what connection, if any, such climate model decadal variability has to reality. This research requires developing statistical techniques for optimally detecting predictability, since model runs on decadal time scales are expensive and hence limited, and traditional space-time filtering methods are sub-optimal in the sense that they do not take advantage of space-time correlations that could be used to enhance the signal-to-noise ratio. We propose two new methods for detecting predictability. The first method, called optimal persistence analysis, identifies projections of the data with as much power concentrated at low frequencies as possible. The second method, called predictable component analysis, is a generalized form of discriminant analysis that identifies projections of the data whose forecast distribution differs as much as possible from the climatological distribution. Both methods can be interpreted as forms of “fingerprinting” techniques, except adapted to the problem of isolating either low-frequency or predictable variability. The two methods also can be combined by using one set of projections as the initial basis set for the other. We propose applying these methods to long control simulations from the CMIP3/IPCC AR4 data set to document the kinds of decadal variability that are generated by climate models. Furthermore, the space-time evolution of this variability will be documented to facilitate study of the physical mechanisms of the variability. To compare decadal variability between models and observations, we propose a new orthogonal basis set for representing data that allows the spatial scale and location to be controlled. This new basis set will allow model derived structures to be projected onto observations so that the connection between decadal predictability in models and observations can be investigated. Attempts will be made to account for uncertainty in the decadal variability by drawing on the connection between optimal fingerprinting, the total least squares problem, and Bayesian theory. This research will clarify the degree of predictability that can be expected from near-term climate projections that are initialized close to the present day climate.