In this paper, we develop a cross-layer model for downlink interference in heterogenous DS-CDMA wireless cellular networks. In this model, interference is described as a function of application layer parameters (traffic characteristics) and physical layer variations (channel characteristics). We show that for a heterogenous service DS-CDMA network, downlink interference is a second-order self-similar process and thus has long-range dependence. We then use the predictive structure of total downlink interference to maximize non-realtime data throughput. We use fractional Gaussian noise (fGn) to model the self-similarity of downlink interference. In the proposed method, the base-station uses an optimal linear predictor, based on the fGn model, to estimate the level of interference. The estimated interference is then used to allocate power to users. To maximize data throughput, we use time domain scheduling. The simulation studies confirm the self-similarity of downlink interference and validate the fGn model. The simulation results also show a substantial performance improvement using the proposed predictive-adaptive scheme and confirm that the interference model is still valid after applying the proposed method.