Ten conceptually different models in predicting discharge from the artificial Chicken Creek catchment in North-East Germany were used for this study. Soil texture and topography data were given to the modellers, but discharge data was withheld. We compare the predictions with the measurements from the 6 ha catchment and discuss the conceptualization and parameterization of the models. The predictions vary in a wide range, e.g. with the predicted actual evapotranspiration ranging from 88 to 579 mm/y and the discharge from 19 to 346 mm/y. The predicted components of the hydrological cycle deviated systematically from the observations, which were not known to the modellers. Discharge was mainly predicted as subsurface discharge with little direct runoff. In reality, surface runoff was a major flow component despite the fairly coarse soil texture. The actual evapotranspiration (AET) and the ratio between actual and potential ET was systematically overestimated by nine of the ten models. None of the model simulations came even close to the observed water balance for the entire 3-year study period. The comparison indicates that the personal judgement of the modellers was a major source of the differences between the model results. The most important parameters to be presumed were the soil parameters and the initial soil-water content while plant parameterization had, in this particular case of sparse vegetation, only a minor influence on the results.

In practice, the catchment hydrologist is often confronted with the task of predicting discharge without having the needed records for calibration. Here, we report the discharge predictions of 10 modellers – using the model of their choice – for the man-made Chicken Creek catchment (6 ha, northeast Germany, Gerwin et al., 2009b) and we analyse how well they improved their prediction in three steps based on adding information prior to each following step. The modellers predicted the catchment's hydrological response in its initial phase without having access to the observed records. They used conceptually different physically based models and their modelling experience differed largely. Hence, they encountered two problems: (i) to simulate discharge for an ungauged catchment and (ii) using models that were developed for catchments, which are not in a state of landscape transformation. The prediction exercise was organized in three steps: (1) for the first prediction the modellers received a basic data set describing the catchment to a degree somewhat more complete than usually available for a priori predictions of ungauged catchments; they did not obtain information on stream flow, soil moisture, nor groundwater response and had therefore to guess the initial conditions; (2) before the second prediction they inspected the catchment on-site and discussed their first prediction attempt; (3) for their third prediction they were offered additional data by charging them pro forma with the costs for obtaining this additional information. Holländer et al. (2009) discussed the range of predictions obtained in step (1). Here, we detail the modeller's assumptions and decisions in accounting for the various processes. We document the prediction progress as well as the learning process resulting from the availability of added information. For the second and third steps, the progress in prediction quality is evaluated in relation to individual modelling experience and costs of added information. In this qualitative analysis of a statistically small number of predictions we learned (i) that soft information such as the modeller's system understanding is as important as the model itself (hard information), (ii) that the sequence of modelling steps matters (field visit, interactions between differently experienced experts, choice of model, selection of available data, and methods for parameter guessing), and (iii) that added process understanding can be as efficient as adding data for improving parameters needed to satisfy model requirements.