https://en.m.wikipedia.org/wiki/Nash–Sutcliffe_model_efficiency_coefficient

I had to look up NSE. This suggests that the values you are getting are inadequate at 0.1-0.13, to be used as a predictor. That model appears not to be working for the circumstances. I would agree, an r2 of probably 0.5, is probably the minimum you may want to consider using as for any degree of prediction. And in using one that low, you would want to be able to state what it does good at and what not good. It might do well at baseflow levels, perhaps more helpful for projecting drought, but not so good with flood extremes or storm responses.

Most experimental hydrological catchments or gauging stations want as a minimum of ten years of data, before making comparisons for paired treatment studies, or modeling. If you can find any long term meteorology, stream gaging, well levels, etc., in the same vicinity or physiographic area that may help augment the data you have. Long term stations of at least 25 years are more apt to have had some extremes to assess. And most rainfall and stream gauging suggest some limits in recording frequency. The USGS for subwatersheds, watersheds, etc typically uses 15 minute recording intervals. For small ephemeral catchments, readings might be taken more frequently to catch hydrograph response. So one might ask is the physical data collected sufficient for the desired modeling.

There are quite a few hydrologic models to choose from, and actual data could be used to help validate or modify for conditions. If there have been validated models applied to your area or similar circumstances, these might be worth considering. Since critical design and management decisions could be based on hydrological models or advice, it is important to include assumptions made and estimate levels of uncertainty, limits, etc. Review and possibly signature in support by licensed engineer or hydrologist might be desired to help address the potential liability issues to avoid if published models or measures are later found to be inadequate or misleading, especially for example topics as flood frequency, severity and design.

I probably did not answer your questions. I have always preferred to use the USGS stream gauging data and along with local data I collected when needed. Being able to read the channel relative to bankful elevation is also helpful (more info at www.wildlandhydrology.com). I hope as professionals we can do better than a good guess, but I know with the past information that I have had access to (USGS, experimental hydrologic stations USFS, precip data network, well records, etc.) is likely more than most have to work with. Some of the earliest experimental watershed studies I believe were in Germany, since I saw your University location). In the USA, there are some unbelievably excellent works from many decades ago. In some ways, they seemed to invest more back then, than they do now. Many long term stations have been abandoned due to cost savings. I know very little about the Germany studies if that is the area of interest. Land use, soils, geology, topography, channel types, incipient conditions and hydrologic modifications are among the factors that need to be considered in hydrologic responses. I do note that the USGS flood frequency models typically have a 40% variation, so this level of uncertainty from data based models across physiographic areas or regions should be expected. By selecting the more localized stations to an area of interest, it is likely the variation can be reduced. Often the physiographic area or regional approach to prediction of flood frequency uses a regional skew, sometimes urban or land use adjustment coefficient, and does not normally make adjustment for rainfall intensity and frequency differences such as those in Technical Paper 40 (US Weather Bureau, NOAA). Based on my work in the Chattooga River, it appears that the high rainfall in this area has to be considered when applying regional or physiographic area equations for flood prediction. In the upper part of the Chattooga watershed, a 40% increase in the 100 year flood estimates appear to be justified, based on the rainfall frequency map when comparing this local area to the average rainfall of the many stations, and also based on storm response data actually collected and compared to other station data. In addition, past failure of road sections were due to culvert failure or erosion bipass. In some cases, leaf fall was also contributing to culvert plugging or diversion, and the engineers decides to just oversize replacement culverts. But in my response, I have not been very impressed with the few models I have tried. I have seen some results of individuals that suggest they have had good, even impressive results with Mike-She, SWAT, etc. I have not been very impressed in past trials from the SCS-CN method that I have tried or been exposed to. In a couple of instances, I found and somewhat impressed that the Rational Method was in the ballpark on small catchment estimates in comparing with results from the USGS equations. It seems important to validate model with data, or data based equations from vicinity of interest.

Which models are you using for both approaches? I work with TopModel and Transfer function models. NSE emphasises the fit to peak flows so it would seem that neither approach is reproducing the peaks well. What is the purpose of your modelling?