Review of gmd-2024-98 manuscript ‘Robust handling of extremes in quantile mapping - “Murder your darlings”’ by Berg et al.

General comments:

The paper gives an interesting insight into the special problems of extreme-value extrapolations/bias corrections. The paper identifies a real problem in the MIdAS method which is also important for efforts not based on that specific software. A series of tests were performed and new approaches are suggested. 

Important paper that should be published just to help users of MIdAS. The ideas in the paper will also provide food for thought for general practitioners in the field of climate-change projection; especially at the extremes end of the distributions.

Specific comments:

The paper is set in a language attuned to people who use MIdAS, which is fine for that community, as well as accessible for people in the relevant areas.

I’d like to see a sentence or two more on the subject in Line 71 ‘conservatively remapped’ - how?

Technical corrections:

Line 22: ‘a reference data’ -> perhaps just ‘reference data’

Line 23: ‘in worst case’ -> ‘in the worst case’

Line 31: ‘will act differently’ - ‘will’? - perhaps true. Seems a strong statement.

Line 81: ‘main figures’ -> ‘main Figures’

Figure 1 caption: ‘absolute values’? Do you mean numerical absolute, or a more general phrase relating to ‘levels’ (as opposed to anomalies)?

Line 87: ‘due to the ensemble mean’ … elaborate that thought, perhaps?

Figure 2 legend in both panels, but especially the lower panel, obscures the graphs.

Line 103: ‘forces the’ -> ‘force the’, as ‘experiments’ is plural.

Line 108 was hard for this reviewer to parse.

I liked this paper. It is well written, succinct, has a catchy title, and makes an important point. My one complaint – and it is a big complaint – is that this paper only considers linear extrapolation. My intuition is that you wouldn’t need to “murder your darlings” if you fit an extreme value distribution to the tail of the data. I would strongly recommend adding the results from such an implementation to Fig. 2 and Fig 3. I’d be ok accepting the paper without this addition, however, since it already stands on its own as a cautionary tale.

One other big-picture thought is that the efficacy of your extrapolation method at reproducing the real world can be assessed by removing the top 5 percent of your training data and confirming that your extrapolation method correctly predicts that top 5% during the training period. Fortuitously, your Fig 2b shows this in the sense that MIdAS minus origGCM (teal line) by construction handles the top 5% properly during the training period and R5T5 (red line) is what happens when you throw out the top 5% of the data and try to extrapolate it instead. The difference between the teal line and the red line represents the failure of your extrapolation to produce the correct extreme values. It doesn’t look to me like your method is doing a very good job; again, I’m curious whether an extreme value distribution fit would perform better.

Other than these big-picture complaints, I have only minor complaints about wording:

• L14: missing word: “for each calendar month, WHICH further aggravates this issue”
• I think it would be useful if you define quantile mapping in the introduction. This will make your paper accessible to people who are just learning about bias correction.
• L33-34: “…will not be the same method as that which was calibrated…” Could you rewrite this sentence? I don’t understand it at all.
• L35: “One can force the bias adjustment to apply its full range of methods only by…” – I don’t think “methods” is the right word?
• ~L47: I think it would be useful to write out the equation for your linear spline implementation.
• L52: “31 d are used to build the distribution of the reference and model data” – I think you mean 31 days times 29 years of samples, right?
• L65: Since you always use Theil-Sen regression for the top 5% of the data, you could remove the “R5” part of your T%R5 names if you want.
• L73: put an “s” on the end of “unique combination”
• Figs 1-3: in panel b, it would be useful if the y label could be clear that this is a difference in annual maxima.
• L87: “interannual variability is reduced due to the ensemble mean for the model data” – it may be worth explicitly clarifying that reduced variance in no way indicates that the quantile mapping is wrong – it is due to averaging realizations once the remapping is done.
• I’d like to know a bit more about why the original MIdAS approach causes bias corrected future predictions to look ~identical to the original data. It seems like this would only happen if the slopes you’re using for extrapolation are ~identical for the model data and the observation data, which doesn’t seem assured. Could you say more about this?
• ~L63: you could be more clear that you are doing a *single* Thiel-Senn regression using all data from the top 5% of the remaining data.
• On L123 you say that changed slope is a “negative impact” but the following sentence says that whether it is good or bad is up for debate. This is inconsistent.
• L131: The second sentence of this bullet (“To create a robust sample size the bias should be assessed over an ensemble of simulations and/or over a larger set of gridcells”) is a separate thought and should have its own bullet.