Review of “Assessment of windstorm-damage relations in the complex terrain of Norway” by Jaison et al.

The authors assess the performance of different storm damage functions, that model the relationship between wind speeds and insured losses, for Norway. They make use of district-level insurance loss data from Norway, that span the time period 1985 to 2020, combined with regional reanalysis/hindcasts of wind speeds. The study nicely assesses the relative benefits of the different (previously developed) storm loss models.

The manuscript is overall very well written and provides a very clear overview of the topic, and insightful explanations and discussions. I only have a few minor comments for the authors to take into account before publication, and congratulate on this nice piece of work.

Line 59/60: Also the Donat et al 2011b study (already cited in the manuscript) estimated the losses at district level in Germany, regionally adjusting the loss function by Klawa&Ulbrich

Line 114: maybe specify that these hindcasts are “retrospective” forecasts, to avoid any ambiguity

Line 130-132: I note that the authors follow here a different order of operations than some of the studies they reference for the loss functions they use. E.g. the Klawa approach first calculates a loss index (i.e. the cubic exceedance of the wind speed threshold), and applies the population density weighting afterwards when spatially aggregating the losses. This is different to what the authors are doing here, as they seem to apply the population weighting already to the daily wind speeds. It would be good to (i) be explicit about this variation in approaches, and (ii) discuss/demonstrate the effects of these different orders of operations.

Line 146-150: I wonder how sensitive are the results to the specific choices of these testing and training samples?

Line 170: not clear if this statement that 82% of losses are recorded above the 95^th percentile is based on the Norwegian loss data analysed against NORA3 wind speeds? Also it is not clear if it refers to 82% of loss events (as count), or 82% of loss values?

Line 178: In my understanding Klawa developed the function for Germany-wide losses (not districts)?

Line 179: As mentioned further up, Donat et al 2011b was first calibrating the Klawa function at district level for Germany

Line 182: You should specify over which time period and which seasons the percentile threshold was calculated (e.g. annual percentile, or percentile over the winter storm seasons ~October to March)?

Line 197: It may be useful to clarify whether this fixed threshold of 9m/s applies to maximum gust or maximum wind speed from the NORA3 hindcast?

Line 276: should “spread” better be “distribution”?

Line 286: I think it is not really data-scarce, I understand that you have good/complete data but these tell you that there are only very few losses?

Line 320-325: Based on Table 1 it may be fair to say that seems to be the best function for most municipalities?

Line 361: insert “is” after “This”

Line 390: remove the “ ‘ “ after “functions”

Interactive comments for: Assessment of windstorm-damage relations in the complex terrain of Norway

Summary:

The authors present a well crafted calibration exercise for storm-damage functions. They present a clear decision tree for the chosen methods and assumption. They are using a storm-based approach to statistically fit historical losses to wind speed information using different models. The basis for the calibration are high resolution insurance loss data and wind speed data covering a relatively long time period. The authors present and discuss the results in detail focusing on the high impact events and on creating a damage classifier. In the end, the authors provide a short broader discussion.

General Comments:

• The insurance loss data and the modelled damages are obviously very skewed and mostly presented using logarithmic axis to focus on relative differenced / differences in the order of magnitude. But in the methodology, this is not incorporated as such. I would suggest the authors to change or at least expand their methodology at two points:
  • Section 2.4.5 Ensemble mean method: As another option instead of using the arithmetic mean, I would suggest to use the mean of all the logs, as in:
    
    mean_log  = 10 ^ ( 1/n * sum_i_n[ log( x_i ) ] )
    
    for a n loss estimates x.
  • Section 2.6 Model evaluation metric: I suggest to also calculate a metric that takes into account the very different order of magnitude. On option would be the mean absolute percentage error.
• In figure 1 the damage functions for only one municipality are shown. It is expected and written that there is a the variability of the calibrated damage functions over all municipalities, but it is not shown. It would be nice to either report the range of the calibrated parameters in a table in the supplementary material or even better reproduce a figure similar to Figure 1 showing not the points of the insurance losses but only all the calibrated damage funtions in one plot. This would provide a much better idea of the variability between the different municipalities.
• The data is split into testing and training set, which is a very important practice. I suggest to use a cross-validation approach, especially as the results in Table 1 are reported only on the unseen testing data, and municipalities have to be excluded from the evaluation due to lack of data.
• A short broader discussion is done in the section 4. “Conclusion”. I would suggest to also discuss the following aspects:
  • Discuss windspeed as explanatory value for damage model, is it able to represent the randomness of gust occurance? This is relevant for “low intensity” events, where damages are caused by infrequent and hard to predict stronger gusts. This is relevant for both damage classifier as well as estimating low intensity impacts)
  • The population distribution is not changing in the chosen model setup only the total number of people. Could it be that “where people live” did not only change in scale (represented in the model) but also in location (not represented). If yes, how would this influence the model performance?
  • What is the purpose of this calibration exercise, is there a foreseen application? If yes, it would be nice for the reader to know, especially what function would be the chosen for this specific application? Even better would be a general discussion of the metrics: Which function would be the best for which type of application?
  • Other readers might be needing to do a similar calibration exercise, but might lack the sample size used in this study or might face other shortcomings. It would be very interesting to discuss some learnings that could be generalized for other calibration exercises.

Specific comments:

Title: is the phrase “in the complex terrain” justified? The only terrain specific methodology used in this paper is the usage of a population density that follows topographic features. Would “taking into account heterogenic population density” be more describing?

L29: quick impact estimation for response planning directly after the event is also an important application (see Welker et al. 2021)

Figure 1 a) mark the threshold in the plot (compare with L170:”…,we chose  the 95^th percentile of the wind speed […] as the threshold.” If a modelled damage of zero would be assumed for events with a wind speed below the threshold, it would be nice to show a doted line at zero similar to Figure 1 b)

L172: “…are split into ten equally spaced bins…” I can only see 6 bins in Figure 1 a). I assume it is because 4 bins did not include the minimum of at least 5 loss days. If that is the case, I would be nice to state that for the reader. If this is not the case it would be even more important to state this.

L307ff: “The extreme loss class represents 31 and 9 ….”. The structure of this sentence makes it hard to understand. Please consider a revision.

Figure 3 b) and c). Please provide a reasoning for the skewed percentiles of the non-linear class boundaries. Why are there no 5^th and 10^th percentile?

References:

Welker, C., Röösli, T., and Bresch, D. N.: Comparing an insurer's perspective on building damages with modelled damages from pan-European winter windstorm event sets: a case study from Zurich, Switzerland, Nat. Hazards Earth Syst. Sci., 21, 279–299, https://doi.org/10.5194/nhess-21-279-2021, 2021.