Floods are the most common natural disaster, directly affecting billions of people with a disproportionate impact in developing countries. Early warning systems allow people time to prepare for floods, saving lives and property. A major challenge is a lack of streamflow data -- hydrological models must be calibrated using data records in each location where they are used for forecasting. Machine learning allows us to train a model using all streamflow data available globally, resulting in a model that can be applied anywhere. This allows us to provide significantly more accurate flood forecasts for locations in the world that lack observation records.