MIdAS- MultI-scale bias AdjuStment

Abstract. Bias adjustment is the practice of statistically transforming climate model data in order to reduce systematic deviations from a reference data set, typically some sort of observations. There are numerous proposed methodologies to perform the adjustments – ranging from simple scaling approaches to advanced multi-variate distribution based mapping. In practice, the actual bias adjustment method is a small step in the application, and most of the processing handles reading, writing and linking different data sets. These practical processing steps become especially heavy with increasing model domain size and resolution in both time and space. Here, we present a new implementation platform for bias adjustment, which we call MIdAS (MultI-scale bias AdjuStment). MIdAS is a modern code implementation that supports features such as: modern Python libraries that allow efficient processing of large data sets at computing clusters, state-of-the-art bias adjustment methods based on quantile mapping, "day-of-year" based adjustments to avoid artificial discontinuities, and also introduces cascade adjustment in time and space. The MIdAS platform has been set up such that it will continually support development of methods aimed towards higher resolution climate model data, explicitly targeting cases where there is a scale mismatch between data sets. The paper presents a comparison of different quantile mapping based bias adjustment methods and the subsequently chosen code implementation for MIdAS. A current recommended setup of the MIdAS bias adjustment is presented and evaluated in a pseudo-reference setup for regions around the world. Special focus is put on preservation of trends in future climate projections, and it is shown that the cascade adjustments perform better than the standard quantile mapping implementations, and often similar to more advanced trend preserving methods. The code is available from Berg et al. (2021).



Summery
In this paper, Berg et al. 2022 suggested a modern code set-up that allows for flexible bias adjustment and compared it to different methods based on quantile mapping. This set-up allows for 1-day-of-year-bias 2-cascacade adjustments to prevent from discontinuity and variance inflation in the data. The paper culminates in discussions about the skill of different methods and future directions for advancing MIdAS code implementation.

General Comment
The paper is very well-written. It was quite easy to understand and enjoyable to read the paper. I found the story about King Midas and the effort to relate the story to bias adjustment, quite cool. My only issue was 'the extent of discussion' about some matters. I was expecting a little bit more of explanation (e.g., about CDF-t method, why distribution-based methods are not covered or around L286 to 291). I understand that the authors might deliberately opted out of thorough discussions because of the nature of the paper, but in my opinion, such discussions strengthen this paper and make the whole bias adjustment process clearer.
Given the quality of this paper, I suggest minor revisions for this paper.

Specific Comment
L3: I would remove 'distribution based' from this sentence. There are some advanced multivariate methods that are not distribution based.

L14-L16:
This whole part is a bit unclear to me.
-Can you please clarify: what do you mean by 'spatial focus is put on preservation of trends'?
-What do you mean by more advanced trend preserving method? Do you consider QDM or CDF-t as the advanced method? To me Midas might be as advanced as QDM (simply because I have worked with QDM but have not implemented multiscale bias adjustment). Thus, isn't advanced a bit subjective here?
Please also consider naming some of the advanced methods.

L23:
What are some of the side effect adjustment? Please consider naming some.

L47: multi-variate features
L50: This sentence is unclear to me. please consider re-writing it. What do you mean by stress test of methods?
L77: I would clearly state that why only QM-based methods are selected to be compared to MIdAS.
L147: this sentence needs to be rephrased. Ny probably needs to be changed to no L153: This part needs more clarification. Why distribution-based methods are not favored?
Coming from hydrological community, maybe I am biased but among us distribution-based methods are highly favored. This also comes naturally, as distribution-based smoothing is applied in many hydrological studies to smoothen outliers. In fact, in some studies, at least for temperature, Gaussian distribution seemed to perform reasonably well. Note for example Räty et al. (2018).
L170: I would prefer a little bit more explanation of the theory of this method as it is the most intricate one.
Change and to an L232: I don't understand why this part (method intercomparing) is located in result section? Doesn't it fit better in the method section? With e.g., experiment protocol subheading?
and why the order of describing variables, is changed (in section 4.1 first temperature is explained while in section 4.2.1 first precipitation is described).
Please consider modifying this section.
L286-287: this part seems like a very important part of discussion. However, it is not entirely clear to me what do you mean by different methods for mapping. By 'such methods' which methods are you referring to? Please consider rephrasing.