Real-time computation of data streams over affordable virtualized infrastructure resources is an important form of data in motion processing architecture. However, processing such data streams while ensuring strict guarantees on quality of services is problematic due to: (i) uncertain stream arrival pattern; (ii) need of processing different types of continuous queries; and (iii) variable resource consumption behavior of continuous queries. Recent work has explored the use of statistical techniques for resource estimation of SQL queries and OLTP workloads. All these techniques approximate resource usage for each query as a single point value. However, in data stream processing workloads in which data flows through the graph of operators endlessly and poses performance and resource demand fluctuations, the single point resource estimation is inadequate. Because it is neither expressive enough nor does it capture the multi-modal nature of the target data. To this end, we present a novel technique which uses mixture density networks, a combined structure of neural networks and mixture models, to estimate the whole spectrum of resource usage as probability density functions. The proposed approach is a flexible and convenient means of modeling unknown distribution models. We have validated the models using both the linear road benchmark and the TPC-H, observing high accuracy under a number of error metrics: mean-square error, continuous ranked probability score, and negative log predictive density.