A procedure for identifying classes of effective event patterns in the sequential prediction task was applied to responses to 1 fully random and 13 semirandom schedules of binary events. The assumption that subjects respond to homogeneous and heterogeneous event patterns of fixed length led to more accurate prediction of behavior on 13 of the 14 schedules than the assumption that they respond to homogeneous event patterns of variable length. The event patterns responded to were a function of maximum run length; there was a regression toward simpler patterns as task complexity increased and a tendency to encode patterns as combinations rather than as permutations; and the fully random schedule did possess discriminative properties.