A local flood warning system typically serves a small community situated in a headwater area and exposed to slash floods or rapid riverine floods. Forecasts of such flood events are characterized by short lead times and large uncertainties. A Bayesian theory is formulated for a local flood warning system built of three functional components: monitor, forecaster, and decider. The theory offers a modeling framework and mathematical concepts necessary for (1) developing optimal decision rules for issuing warnings based on imperfect forecasts, (2) evaluating system performance statistically, and (3) computing the ex ante economic benefits from a system. ¿ American Geophysical Union 1993