Dear Authors,
your replies address all the points raised by the three referees (who did indeed an excellent and very constructive job!)and you propose to implement modifications that should adequately solve their main concerns, and therefore I warmly invite you to submit a revised version, amending the paper as you already suggest in your Replies.

The major issue is certainly that of the fairness of the comparison in forecasting mode as far as data assimilation/updating is concerned. As the Authors themselves propose, in their reply to Ref1, it is indeed necessary to replace the current forecasting experiment with one where the actual observed outputs are not used in any way.

Secondly, the calibration procedures should be better described, with details in the text, in addition to Table 3), with particular attention paid to the use of different Objective Functions, especially in order to guarantee a fairness in the models comparison: the use of the same Obj Function(s) would be strongly preferable.

Also the proposed deeper analysis/interpretation of the performances (maybe using the integrated criterion you suggest in reply to Ref3) between catchments and between models (and in particular putting them in relation to model structure), even if certainly not easy, would further improve the paper significance.

Additional remarks:
- I agree with you that it is not necessary here to make new analyses with actual forecast ensembles and extending your justification (as already described in your reply to Ref1, RC1-2) should be enough.
- I agree with Ref2 that section 1.3 should be removed and replaced by a shorter paragraph citing only the papers that are closely related to your work (same models or/and same catchments) and relating their results to yours in the final discussion (when, as you propose in the reply to Ref2-RC1, you will refer to literature results for corroborating your findings).
- The reply to Ref3-RC2 is incomplete.

Thank you for having submitted your interesting work to HESS.

Best regards,
Elena Toth

Dear Authors,
thank you indeed for resubmitting the paper and for the detailed reply letter, where you addressed the points raised by the Referees and by myself.

I report below some observations on the revised manuscript, which I invite you to address in the final manuscript.

Looking forward to receiving the last version,
best regards,
Elena Toth

1) ll. 79-99 this part is still a bit long and the close connection with the applied models is not so clear
2) ll. 309-310: the part referring to teleconnections is still not necessary (nor very clear…)
3) l. 379: rephrase the English
4) ll 399-401: the graph (Figure 6) does not show what happens for the criterion C2M since it does not identify the single criteria.
5) ll. 429-431: add that the quantile-quantile post-correction method is used only in the SIM model and explain why for SIM the performances (especially the POD) not only do not improve but actually deteriorate.
6) l. 460: I would repeat/specify that ‘lower’ refers to the increasing lead-time.
7) l. 463: I would repeat/specify ‘lower than 1 IN FORECASTING MODE’
8) l 497: specify that Figure 13 refers to 7-days ahead lead-time.
9) ll. 637-641: this sentence is not clear (and, accordingly, also ll. 65-74 might be rephrased): it does not exist a distinct class of ‘low-flow forecasting models’ nor a class of ‘hydrological models’ separate from the first one (all models used to obtain simulation/forecast of any hydrological variable are hydrological models). And even referring, correctly, to ‘rainfall-runoff’ models, the models proposed by Demirel and Booij and some of the models listed as ‘simple modelling approaches’ in Section 1.2 (e.g Campolo et al., 1999) are models that are fed by rainfall and evapotranspiration estimates, that is the same input variables used to fed your models and they represent therefore just a different type of rainfall-runoff models. The difference is in the structure of the rainfall-runoff model, either data-driven/systemic (see neural networks) or conceptual, but also a neural network model is a rainfall-runoff model that may be used in real-time forecasting as an alternative to a conceptual model (see, for example, the comparisons in Coulibaly et al., 2000, or Toth and Brath, 2007) and they may of course be fed by meteorological forecasts and not necessarily by no-rainfall scenarios, as, for example, in Coulibaly (2003) or in the recent paper by Demirel et al., 2014. I would suggest rephrasing (or removing since it is not strictly needed) ll 637-641 and I would rephrase ll. 65-74.
Coulibaly P, Anctil F, and Bobee B (2000) Daily reservoir inflow forecasting using artificial neural networks with stopped training approach. Journal of Hydrology 230: 244–257.
Coulibaly P., Impact of meteorological predictions on real-time spring flow forecasting, Hydrological Processes, 17 (18) (2003), pp. 3791–3801.
Demirel, M. C. and Booij, M. J. and Hoekstra, A. Y. The skill of seasonal ensemble low flow forecasts for four different hydrological models, Hydrol. Earth Syst. Sci. Discuss., 11, 5377–5420, 2014.
Toth E and Brath A (2007) Multistep ahead streamflow forecasting: Role of calibration data in conceptual and neural network modelling. Water Resources Research 43: W11405.
10) In most of the captions of figures/tables it would be useful to add if they refer to ‘simulation’ or ‘forecasting’ mode.

Dear Authors,
thank you indeed for your careful revision.

I suggest two final additional technical corrections:
1) adding the bisector line in Figure 6.
2) l. 443: “SIM tends to underestimate low-flows when the POST-CORRECTION method is not used”

Thank you for having submitted your interesting work to HESS and I look forward to see the published manuscript.

All the best,
Elena