Flood early warning systems provide a potentially highly effective flood risk reduction measure. The effectiveness of early warning, however, is affected by forecasting uncertainty: the impossibility of knowing, in advance, the exact future state of hydrological systems. Early warning systems benefit from estimation of predictive uncertainties, i.e. by providing probabilistic forecasts. The present dissertation describes research in estimating the value of probabilistic forecasts as well as in skill improvement of estimates of predictive uncertainty. A framework for estimating the value of flood forecasts, expressed in flood risk, is proposed in Chapter 2. The framework includes the benefits of damage reduction through early warning as well as the costs associated with forecasting uncertainty. The latter manifests itself through instances of missed floods and false alarms. Application of the framework to a case study to the White Cart basin - a small river in Scotland - shows that probabilistic forecasts have higher value than deterministic forecasts. It also allows for deciding on an optimal warning lead time, where the combined benefits of damage reduction (which increase with increasing lead time) and costs of forecasting uncertainty (that also increase with increasing lead time) are most beneficial. Three post-processing approaches are investigated. The first approach (Chapter 3) comprises the statistical post-processing of meteorological forecasts and subsequent use thereof in hydrological forecasting. The analysis shows that while the quality of meteorological forecasts can be improved, the improvements do not proportionally propagate to the quality of the hydrological streamflow forecasts. It is believed that this is due to the inability of post-processing techniques to fully maintain the spatio-temporal correlations. The second approach comprises an exploration of potential improvements to the application of Quantile Regression as described by Weerts et al., 2011. These include the application of an explicit requirement for non-crossing quantiles, the exploration of the benefit of deriving the statistical models in Gaussian space and the derivation of multiple statistical models on several sub-domains of the predictor. The results indicate that the non-crossing quantiles comprise an improvement and that the other two potential improvements do not actually result in observable increase in forecast skill, hence that the post-processor may be simplified for use in operation practice without losing skill. The third approach explores the benefits - in terms of forecast skill - of a lumped post-processing approach versus separately addressing meteorological and hydrological uncertainties. The latter approach was found to yield sharper forecasts, but at the expense of reliability. Combined, this resulted in very similar skill scores with the source-specific approach offering more scope for improvement. The combined findings indicate that probabilistic forecasts have value and that there is scope for additional increase thereof. This is elaborated on in the Synthesis in Chapter 6. Also, recommendations for additional research are given. This includes research pertaining to value and skill of hydrological forecasts as well as to the use of forecasts in forecast, decision and response systems.