Attached you can find my comments.

“A 500-year runoff reconstruction for European catchments” by Sadaf Nasreen et al.

The manuscript “A 500-year runoff reconstruction for European catchments” by Sadaf Nasreen et al. shows the work and effort that has been done to create a new dataset of long-term runoff reconstruction for various European catchments. While reconstructions of meteorological variables such as temperature and precipitation were already available, this study closes the gap by providing open source runoff reconstructions. This is valuable information as it can provide historical context for upcoming studies, which are interested in assessing present and future extremes such as droughts. To create the runoff reconstruction, a semi-empirical hydrological model (GR1A) as well as two data-driven model (LSTM and BRNN) were used and tested for their suitability. An extensive data collection including precipitation, temperature, drought indices, natural proxy data and runoff observations over the period 1500 to 2000 were used to calibrate and validate the models. The data-driven models showed the most promising results, being able to correct for biases in the input data compared to semi-hydrological model. Furthermore, the separate analysis focussing on droughts showed that the reconstructed timeseries of these models correlated well with the historical documented droughts.

The paper was well written and included extensive information on the approach and validation of reconstructed runoff timeseries, especially regarding drought events. The main points for improvement are mainly focussing on additional clarifications regarding certain aspects of the methods. Therefore, I would like to recommend publication after minor revisions. The comments for improvement can be found below.

Major comments:

Section 3.3 Data-driven models: While there is an extensive general explanation on the LSTM model, the BRNN description falls short. More importantly, information necessary to be able to follow as a reader on how the data driven models were trained and the final model parameters are not given or not fully clear. Aspects on the training/testing data, like the type of splitting or how much was withheld for training/testing purposes are not clear. Furthermore, aspects on the input data (for example the gridded+proxy, gridded+PDSI, and gridded+lag) and its preparation (e.g. type of normalization, handling with outliers, etc) would be of interest as well. An additional table (could be in the Appendix) with the input parameters as they are used in the ML models would support the readers understanding. While the Appendix covers the LSTM model structure, similar information on the BRNN is missing. Additional suggestion is to add the final model parameters (e.g. amount of neurons) into the schematization (Fig A.1) as well.

Overall I think adding more specific information on the ML models will only improve the readers understanding and as the data-driven models show the most promising results highlighting and clarifying their use is important.

Section 3 Methods: A schematic overview of the data-preprocessing (3.1), the incorporation of all the different datatypes and sources in the different model types (3.2 and 3.3) as well as the postprocessing steps (3.4 and 3.5) would be a nice addition to this section to not only visualize the general approach of the study but also support the subchapters and the readers understanding.

Minor comments:

line 43: As an example: Hansson et al…. remove “of”

Fig 1 (and also Fig 5 and Fig 6): think about changing red or green to a different colour to ensure that colorblind people can follow your figures

Line 75: sentence not flowing, for example move comma in front of reference and remove ‘was done’

Fig 2 and Fig 3 same range and colorbar per evaluation metric (makes it easier to compare), list min and max values of scale bar for readability

Section 3.4: possibly some lines on the pros and cons of the GR1A model

Section 4.1 (and throughout the rest of the manuscript): be consistent in addressing GOF (now a mix between GOF and gof)

Section 4.2 the information on calibration and validation should be part of the Methods and the models. Furthermore, it would be nice to move a figure with the timeseries of one station as seen in Fig S1 from the supplementary to that paragraph to highlight calibration and validation periods.

Line 255: ‘greatly increased the performance (NSE from 0.2 to 0.62).’ Compared to the values mentioned in prior example of Basel Reinhalle, 0.62 is not listed Table 3 for BRNN(Gridded+PDSI) but 0.57

Table 3: highlighting the different performances is a nice feature and helps spotting important trends, however the darkest colours make it hard to read the values (same for tables in supplementary). Maybe also add a note in the table description what the colour indication means.

Table 3: both stations at Basel show higher correlation scores for validation than calibration. Ideas why this is the case?

Figure 7: Whitespace around the figure seems to be cut too narrow as the max value for station BaselRheinhalle-Rhine is cut off (130 instead of 1300)

Line 347 and Table 5: listing of years does not include 1724, which is also indicated in bold in Table 5.

Section 6: ’using the data set below’ move below

Appendix: add references to equations and in text (easier to follow in case chapter layout changes)