We develop a probabilistic forecasting methodology through a synthesis of belief-network models and classical time-series analysis. We provide methods for constructing, refining, and performing inference with these dynamic network models. From the perspective of investigators trained in statistical forecasting, our belief-network formulation allows us to generalize common assumptions of independence in traditional time-series analyses. From the perspective of investigators studying the automation of uncertain reasoning, we extend static belief-network models to more general, dynamic forecasting models by providing (1) a representation that integrates temporal and contemporaneous dependencies, (2) techniques for iteratively refining the extent of the temporal relationships, and (3) efficient means for performing forecasting and inference with the network models. We conclude with a discussion of how our approach addresses several limitations found in traditional time-series analyses.