Dear Authors,

I have accessed the paper and comments provided by the reviewers.
I have similar thoughts as reviewer 2.

I think it would be good, based on available papers on global hydrology, to say something on the (expected) source of skill in different parts Africa (see van Dijk et al. 2013, Candogan et al. 2013 and others) although you did not do this work yourself (maybe in the introduction)

I agree with reviewer 2 that using persistence makes much more sense than a seasonal mean (expecially because you focus on flood forecasting application) and requires little extra work
The choise of the verification metrics is unclear. I would also like to see reliability plots and maybe others. Besides the choice of the verification metrics, there is a problem with the one-year hindcast period which makes your statistics in essence meaningless and no claims in performance can be justified. Boostrapping procedures might give insight into the effect of limited sample sizes (see for instance Lopez Lopez etal., 2014 HESS)

The setup of the experiments are unclear. It is unclear what data is used when and for what. This has also to do with the structure and I advise to restructure the paper more inline with other articles Intro-Material & Methods (including data, model, experiments both calib/hind, verification both calib and hindcast), Results and Discussion and Conclusions) this will help to avoid restating which data set is used for what etc (this could be done in 1 table) because this is rather unclear and it will help to seperate results from the experimental approach etc.

For calibration it is ERAInterim +GPCP (page 5568):
For the hindcast in 2003 it is EARInterim+GPCP (page 5570) with some strange sentences: The initial hydrological conditions, i.e. all state variables, were computed for each forecasting date between 1 January and 31 December 2003 by running LISFLOOD with near real-time meteorological observations. During the hindcasting period these needed to be approximated by using the daily GPCP-corrected ERA-Interim; however, during real-time forecasting, the first day of each ECMWF deterministic forecast could be used.

However on page 5566 it is stated:
State variables are calculated for
15 each 0.1 pixel by forcing LISFLOOD with the near real-time meteorological observations over the forecasting period (here: 1 January–31 December 2003

This is also present in the conclusions where suddenly SRFE is being used (5578 line 17).

In the manuscript you refer to 5 data sources (which are not explained in 4.1):
Page 5565 :The five main data sources on which AFFS relies are: historical hydrological observations,historical as well as near real-time meteorological observations, real-time meteorological forecasts and an African GIS dataset. Specifications of these are given in Sect. 4.1.

In the end only ERAInterim/GPCP is used and it is suggested to use EMCWF first day as proxy for this ...????

If the same flow of data is not used operationally, it is needed to state in the conclusions and abstract that the operational results are probably less good than what is presented here in this paper

Page 5572: For the calculation of the CRPSS, discharges were normalized to remove possible systematic biases, while the seasonal mean was used as benchmark. This is stated without any explantion how it impacts the results and it is unclear for me and many other readers what the impact of this choice is.
I think it is good that some other simpler metrics like reliability plots or simple obs versus ens forecasts for various leadtimes at various locations this will give more insights for many folks instead of something that is rescaled.

I find the first part of the discussion and conclusion not up to HESS standards (5577/5578) which is a scientific journal and not a venue to do promotion please rewrite and keep it to the science (and I think the statements on page 5568 line 8-12 are not related to the FAR).

Minor:
I find the phrase "The predictive capability of the African Flood Forecasting System (AFFS) " reoccuring several times please adjust this it is not necessary to repeat this everytime, once should be enough etc

Dear Vera, Dear Authors,

In my comments I was not thinking about seasonal forecasting at all.
I think it is good to think about where the skill, if any, is coming from
and make this clear including some words on this in the introduction

Hence, also my comment about using persistence as reference forecast instead of the seasonal mean or moving average over 30 days (I don't understand the reply about using an ESP procedure). See for instance
A comparison between ensemble and deterministic hydrological forecasts in an op-
erational context Adv. Geosci., 29, 85–94, 2011 www.adv-geosci.net/29/85/2011/
doi:10.5194/adgeo-29-85-2011
and
The use of MOGREPS ensemble rainfall forecasts in operational flood forecast-
ing systems across England and Wales Adv. Geosci., 29, 77–84, 2011 www.adv-
geosci.net/29/77/2011/ doi:10.5194/adgeo-29-77-2011

where persistence is used as reference forecast. This would allow to see if the ECMWF-EPS improves skill for longer lead time or that hydrologic model/initial conditions is important. I think using Persistence gives probably better insight into the value of the ECMWF forecasts.This is especially important for flood forecasting.

I think you should state and make clear that the results are indicative as the record/nr of forecast and measurements for making statistically sound conclusion is not available (and this is the case in many forecasting studies)

Can you use persistence as reference forecast for the CRPSS results (and adjust discussion/text if needed) instead of the seasonal mean/moving average as suggested by reviewer 2 and myself? After which we can proceed with the review process?

Sincerely,

Albrecht Weerts

PS Figure 9 has no caption..

Dear Authors,

Two reviewers reviewed the adjusted manuscript. In general they are positive about the revised manuscript.
Reviewer 1 requires technical corrections and reviewer 2 requirers additional modifications and checks. I agree with reviewer 2 that these corrections and checks are necessary.
Point 2 of reviewer 2 was already raised last time and the reviewer raises another valid point regarding the use of the seasonal mean. So either use only the naive forecast/persistence as reference or use the climatology (and leave out the seasonal mean)
Point 3 needs serious attention. You would indeed expect that the CRPSS starts negative and (slowly) increases (1- CRPSforecast/CRPSref), CRPS=0 perfect forecast.

If any question exist please contact me to speed up the process.
Otherwise I expect a revised manuscript that needs only an editors review.

With kind regards

Albrecht Weerts

PS Yossef reference is missing, please check all the references.

Dear Authors,

Thanks for the corrections and improvements.

I have one final minor request to extend the discussion/conclusions based on
Figure 7 . This figures indicates that state updating (or even error correction) may improve the flood forecasts at least for day 0-2. Can you include this in the discussion/conclusion
If discharge data would be available or would be made available in real-time much more skillful predictions for shorter lead times could be issued.

Sincerely,

Albrecht Weerts