This paper studies spectrum trading in cognitive radio networks in which multIPle service providers (SPs) sell unused spectrum to multIPle unlicensed secondary users (SUs). Motivated by the nature of the problem with new considerations, spectrum trading is modeled as a multi-leader multi-follower expected Stackelberg game with two levels of competition. The SPs as leaders compete in offering subscrIPtion prices (upper-level subgame) and the SUs as followers compete in selecting service from the SPs (lower-level subgame). The lower-level subgame incorporates the time-varying spectrum availability as the external state so that the proposed scheme does not require knowledge of dynamic spectrum availability. To achieve self-organized network operation, we propose decentralized, stochastic learning-based algorithms for the game. The convergence properties of the proposed algorithms toward the Nash equilibrium (NE) are theoretically and numerically studied. The proposed scheme demonstrates good utility performance for the SUs as compared to other service selection schemes.