In this paper we introduce the notion of association bundle as a new pattern for association analysis. Different from association rule which has the one-way association from rule antecedent to rule consequent, an association bundle has the two-way association between each pair of bundle elements. Association bundles present a more compact association pattern than association rules, and have meaningful interpretations in marketing applications including shelf arrangement and cross-selling by discounting. The identification of association bundles can be transformed into the problem of maximal clique enumeration, and very efficient algorithms can be applied. Our experimental results show that the time cost is less than two minutes on three real world data sets.