Calibration and model prediction is always affected by uncertainty in the forcing data, response data and structural error in the model. The storm-dependent parameters are believed to be able to capture these errors and improve model prediction. The goal of this research will be to further investigate and develop the storm-based approach and compare it to more traditional approaches, including the use of static (time- invariant) parameters and the use of GLUE to capture model errors. The hypothesis in this paper is the ran- dom variation of storm-dependent parameters can capture the model error and improve the prediction. The storm-based method will apply a sensitivity analysis to identify the storm-dependent parameters that are most likely to vary by storms. The variation of storm-dependent parameters will give large changes in model performance which is measured by Nash–Sutcliffe efficiency. In the storm-based method, parameters will be calibrated storm by storm in the calibration period. Then in the validation period, streamflow will still be predicted storm by storm through picking each parameter set from the calibrated parameter sets. Besides, the effect of dryness and the optimal threshold for identifying the storm epochs on the model performance will also be explored in the storm-based method. The results obtained by the storm-based method will be compared with the method using static parameters and GLUE method. Six case studies calibrating a con- ceptual rainfall-runoff model with for parameters with daily data illustrate the improvement of prediction obtained by the storm-based method for dry basins. The extent of variation of storm-dependent parameters is very random in each case which indicates there is error in the model. Moreover, the extent of variation of storm-dependent parameters has no relation with initial water storage, rainfall characteristic and basin characteristics. Although the parameters cannot be predicted deterministically, they can be predicted prob- abilistically with the histogram or fitted distribution for the calibrated parameter sets in the future work. By making storm-dependent parameters vary with storms and other parameters constant, the storm-based method performs better for drier basins while worse for wetter basins compared to GLUE method and tradi- tional method. The logscore value obtained in storm-based method(e.g. -0.68 for one of the dry basin A) is larger than those obtained in the traditional method (e.g. -1.23 for basin A) and GLUE method (e.g. -0.67 for basin A). Additionally, the RMSE values for total flow obtained in the storm-based method are all smaller than those obtained in the traditional method, and GLUE method for dry basins. This suggests the storm-based method is more applicable for dry basins and this method should be better developed for wet basins. What is more, the extent of variation of storm-dependent parameters has no relation with basin characteristics but the mean and variation of the storm-dependent parameters can be obtained. Hence the extent of variation of parameters can be described probabilistically.