Water management decisions from the short-range to the climate scale commonly rely on hydrologic analysis and predictions that must provide both strategic information over large domains and high-quality information at local watershed scales. Over the last decade, new paradigms for hydrologic prediction have emerged, as well as new modeling frameworks, earth observations and tools to support these paradigms. This presentation describes the evolution of flood forecasting from a manual, forecaster ‘in-the-loop’ paradigm for local to regional areas toward a more centralized, large-domain, ‘over-the-loop’ paradigm that can take multiple forms. Examples of this latter paradigm include coupled NWP-based forecasting, intermediate scale ensemble forecasting, high-resolution process mode-based forecasting, and the new heterogenous modeling approach taken by the US National Water Center’s ‘Nextgen’ system. Yet these new paradigms often struggle to produce skillful, actionable predictions, in part due to the omission of key methodological elements that notably gave rise to the traditional ‘in-the-loop’ paradigm. More recently, there has been a rapid rise of systems based entirely on machine learning, an approach that intrinsically includes these elements and is likely to become a dominant form of short-range flood forecasting for water management in the next five years. Although this talk includes mainly examples from the US. This presentation illustrates these paradigmatic developments with examples drawn mainly from the US, providing a general overview on trends that are advancing in multiple continents.