Single-Model-Bootstrap Applied to Neural Network Rainfall-Runoff Forecasting

Most neural network hydrological modelling has used split-sample validation to ensure good out-of-sample generalisation and thus safeguard each potential solution against the danger of overfitting. However, given that each sub-set is required to provide a comprehensive and sufficient representation of both environmental inputs and hydrological processes, then to partition the data could create limited individual representations that are in some form or other deficient with respect to fitness-for-purpose. To address this issue a comparison has been undertaken between neural network rainfall-runoff models developed using [a] conventional stopping conditions and [b] a continuous single-model-bootstrap. The results demonstrate marginal improvements in terms of greater accuracies and better global generalisations – but substantial advantages in the form of automation and diagnostic capabilities.