Review of Using seasonal forecasts to enhance our understanding of extreme wind and precipitation impacts from extratropical cyclones, by Maddison et al.

This paper proposes to use an ensemble forecast dataset as a surrogate to investigate the impacts wind and precipitation from extra-tropical storms. The paper uses a UK MetOffice product (GloSea6) and compares it the the ERA5 reanalysis. The paper focuses on the statistical analyses of impact indices related to precipitation and wind speed. The authors investigate the relation with the North Atlantic Oscillation (NAO).

The paper is well organized and the idea of using ensemble forecast data as surrogates of reanalyses is very appealing.

Major comments

Why not use the ensemble members of ERA5 (rather than the mean)?

The authors determine empirically the probability of exceeding the record (highest value) of ERA5 in the GloSea6 ensemble, after having verified that the two datasets yield similar probability distributions (Figures 3-6). In principle (B. Arnold et al., Records, Wiley, New York, 1998, Ch. 2), if the record in ERA5 is obtained in, say, N=75 years, then the probability to exceed this record (say in GloSea6) is 1/(N+1). This is close to the empirical estimates that the authors find. Therefore, an important finding of the paper (increasing the data size increases the probability of exceeding the record) is actually fairly trivial from the statistical point of view (i.e. one can get it from a paper-pencil computation).

What is not trivial, but undiscussed, is the strange behavior of SSI data in Finland, for which the GloSea6 distribution is much lower than the ERA5 distribution, although the core distributions look similar. Any idea?

The return level plots in Figure 9 are probably wrong (the curves should start from the same return period). What are the Pareto distribution parameters? Computing return level plots from a Pareto distribution fit is potentially tricky, especially because the SSI values are conditional to the occurrence of storms, and not on a time axis. This is where conditioning on an NAO index (for example) could be more useful. A clarification on how the GPD fits are obtained is necessary.

Minor comments

What the authors call “loss” is actually the value of wind or precipitation indices. This does not pertain to actual insurance losses, and might be misleading.

Figures 3-4 seem redundant with figures 5-6 (same information?).

L. 181: The NAO index on sub-daily increments might not be super relevant (it is generally used on monthly time scales) because of the spatial variance of the low and high pressure systems.

Sec. 5.3: The methods section should explain into more details what is done in Figure 8. How exactly are computed the ratios? Are there uncertainties?

L. 467: The cyclone tracks are produced by the TRACK algorithm of Hodges, and obviously by K. Hodges himself (cf. Acknowledgments). How can the algorithm be distributed from the authors, since K. Hodges does not distribute it? The cyclone tracks should be made available without a request to the authors.

Overall the paper describes some good work but I feel some of the emphasis is misplaced. It has a slight feeling of trying to sell something with claims about the significance for the insurance industry. Particularly on the windstorm aspect, I would like to see some clearer discussion on what the results show about the likelihood of 'unseen' storms and how that fits with existing understanding.

More specific comments: 

Line 36-37 seems dismissive of GCMs and seems to imply that the unseen dataset would not require downscaling or bias correction, which is not substantiated in the paper.

69-73 SSI discussion is brief and uses outdated references. Their limitations and alternatives could be better considered. Previous works include Pinto 2012 and Karremann 2014, then Moemken 2024 showed that while SSI or similar metrics are able to rank storms, they are not an estimator of loss in extreme cases.

110 - is the dataset resolution .7deg? This is quite low. Population weighted SSI at this resolution may not classify storms very well and will introduce biases towards core Europe.

170 - This seems a key step relegated to an appendix.

196 - 60 storms per season seems high if we are intending to focus on extremes

Fig 2 - could this analysis focus on the extreme events?

210-213 - The biases are small but coastal features etc are crucial in loss modelling.

Fig 3 - Again, I would like to see more focus on extremes, but acknowledge this comes later.

264/fig 5 - visually I don't find this a very helpful way of showing the extremes particularly eg UK. I'm not critical of the data shown here but this feels like a key result that could be illustrated better.

279 - reference to figure 5 I believe should be to figure 7?

310-313 - this feels well-reasoned and good demonstration of an existing hypothesis even if limited to a per-storm time window.

Fig 7 - Are some of these very high? 600 storms in 697 years in UK is around 1 per year. Lack of strong storms in AT is attributed to the gust quantile mapping in a mountainous country. What about Switzerland? It is mapped in figure 1 but does not seem to be included in the results. Also for Finland is this not a tracking issue with storms not persisting sufficiently far East?

Section 5.3 OK with all this. Good discussion.

371 & 374 - I understand what is being done but it reads to me that you state the 10 year RP is matched then claim that the agreement across most RPs is good which then seems obvious.

Fig 9 - particularly for UK this now seems to present a conclusion that is contradictory to figure 7. we now see that at eg 100 years RP all the GloSea6 storms are below the extrapolated ERA5 lines, whereas previously it was claimed that the dataset had a very large number of storms above the highest historic observation. I presume this is due to the 10yr RP scaling but this feels contradictory to the claims in the introduction that the dataset would not need re-calibrating due to biases.

396 - Can we, for example. estimate an RP for Daria from the dataset? This is an example of how the discussion could focus on how this dataset enhances or adds to existing understanding of extreme windstorm risk. Existing works include the XWS and C3S catalogues. Flynn 2024 (admittedly in pre-print) has some good review of these. Alongside the other works referenced around lines 57-64.

Appendix A: This seems quite brief, could it not just be in the text? I would like to see some more recent thinking on SSI included beyond the 2003 paper.