This paper does a good job of taking the reader through  bias corrections for three regional climate models which overestimate the number of rainy days and underestimate the precipitation amounts compared to annual and monthly observations.  They demonstate that random forest regression outperforms thresholding for station comparisons where the model-measurement difference exceeds 5%.  The figures and discussion are clear, and I have only minor comments:

1) I had hoped for some insight into why the random forest regression succeeded when thresholding failed.  Perhaps  a paragraph or two in the discussion could cover next steps for learning which of the 5 feature set variables mattered most, and how to use that information for model improvement?

2)  On line 152 the authors state that they used scikit-learn for the random forest model, but as far as I can see the zenodo archive contains only R code?

3) The captions for Figures 5 and 6 say that the model and rf data are red and orange, but in the figures they are green and yellow

4) For the Q-Q plot in figure 5, standard case 48 thresholding does especially poorly -- any idea why this gridcell is an outlier?

5) the column headers for Table 2 and Table 3 use different labels for the same quantities -- case 0, case  1, "standard cases", "extreme deviation case" "div > 5%" - why not just adopt the unambiguous labels of Table 3?

6) it is stated that GridSearchCV was used to establish "the optimal set of hyper-parameters".  It would be useful to state what the hyper-parameters were, along with the feature set varaibles

Hi editor,

Please feel free to check the attachment.