Articles | Volume 374
https://doi.org/10.5194/piahs-374-159-2016
https://doi.org/10.5194/piahs-374-159-2016
17 Oct 2016
 | 17 Oct 2016

Comparison of cross-validation and bootstrap aggregating for building a seasonal streamflow forecast model

Simon Schick, Ole Rössler, and Rolf Weingartner

Abstract. Based on a hindcast experiment for the period 1982–2013 in 66 sub-catchments of the Swiss Rhine, the present study compares two approaches of building a regression model for seasonal streamflow forecasting. The first approach selects a single "best guess" model, which is tested by leave-one-out cross-validation. The second approach implements the idea of bootstrap aggregating, where bootstrap replicates are employed to select several models, and out-of-bag predictions provide model testing. The target value is mean streamflow for durations of 30, 60 and 90 days, starting with the 1st and 16th day of every month. Compared to the best guess model, bootstrap aggregating reduces the mean squared error of the streamflow forecast by seven percent on average. Thus, if resampling is anyway part of the model building procedure, bootstrap aggregating seems to be a useful strategy in statistical seasonal streamflow forecasting. Since the improved accuracy comes at the cost of a less interpretable model, the approach might be best suited for pure prediction tasks, e.g. as in operational applications.

Download
Short summary
In water resources management, planning at the seasonal time scale is confronted with large uncertainties. Key variables are often unknown or hard to forecast, e.g. precipitation of the next three months. In the present study, we try to highlight some aspects concerning the development of a model faced with these uncertainties. Using the example of statistical streamflow forecasts, the results of the study indicate that the forecast accuracy is improved by the combination of several models.