Clustering of functional magnetic resonance imaging (fMRI) time series--either directly or through characteristic features such as the cross-correlation with the experimental protocol signal--has been extensively used for the identification of active regions in the brain. Both approaches have drawbacks; clustering of the time series themselves may identify voxels with similar temporal behavior that is unrelated to the stimulus, whereas cross-correlation requires knowledge of the stimulus presentation protocol. In this paper we propose the use of autocorrelation structure instead--an idea borrowed from geostatistics; this approach does not suffer from the deficits associated with previous clustering methods. We first formalize the traditional classification methods as three steps: feature extraction, choice of classification metric and choice of classification algorithm. The use of different characteristics to effect the clustering (cross-correlation, autocorrelation, and so forth) relates to the first of these three steps. We then demonstrate the efficacy of autocorrelation clustering on a simple visual task and on resting data. A byproduct of our analysis is the finding that masking prior to clustering, as is commonly done, may degrade the quality of the discovered clusters, and we offer an explanation for this phenomenon.