Please see the comment in the supplement.

General comments

This study deals with the relevant problem of establishing an early-warning system at a small regional scale with few observations. This problem is approached by testing different models with different inputs and tested for robustness. I think the study will be interesting to early-warning system developers as this system has been implemented in practice and therefore had to tackle many practical problems from how to establish thresholds to how to issue warning levels.

Generally, the manuscript is well written and organized. My main criticism is that the comparison/evaluation/validation is at times confusing and not as streamlined as other parts of the manuscript. Although I very much appreciate that much effort is put in the validation, I think section 2.3 (and 2.4 maybe) need some justification why so many different approaches are being taken. These should maybe also be presented more clearly later. You may have good reasons for choosing many different validation methods (leave-one-out for events, train/test split, different criterion metrics, etc.) but it’s not clear to me from the text and I think it will be confusing to readers. For example, there is leave-one-out for landslide events and train/test split. Wouldn’t it be simpler and similar added value to do leave-one-out with the recorded years (e.g. train with 2002-2018 and test with 2019)? Or why do you need the Brier skill score? Can’t you use the same skill score as for the other models and compare these? Anyway, I don’t think you go into details with the results of this part so it could be cut. To systematically compare the predictive power, you could also compute the area under the curve and assess if a model is better than random, as it is commonly done to assess predictive model skill.

I think if these issues could be solved, the manuscript will be more accessible to readers and that the authors are in a good position to solve this. Please find more specific comments below.

Specific comments

L17: Please specify «136 statistical models». When only reading the abstract I cannot imagine what that means. How do models differ?

L18: does that mean that your data is in daily resolution? If not, I would count the number of non-triggering rainfall events instead of the days.

L23-25: seems more like a conclusion to me. I would state this later in the abstract.

L34-72: the intro nicely shows that there is a need for a LEWS in this region but that it’s difficult to establish with current methods

L53-54: I think this sentence should end with a citation

L153: general question: what is the added value of the number of landslides if you don’t know where it’s going to happen?

L217: So 1-day antecedent precip is the total daily rainfall at the day of the landslide?

L230: Please give a citation for these equations

L235: Please integrate this sentence in another paragraph where you discuss this problem

L247: more intuitive than what?

L271: What is a chain?

L276-279: how are these criterions defined?

L275-289: why was not the same validation performed on each model?

L332.335: I agree about the inadequacy of accuracy in imbalanced datasets but ROC shows exactly true alarms and false alarms in relative terms.

Fig. 3&4: please make sure these figures have the same layout (axis limits, labels, font size, order of plotting lines, etc.) (same for fig. 5&6). As these figures look very similar, you could consider e.g. only showing figs 3&5 here and move the others to the supplement.

Fig. 7/8: blue stands for better and red for worse. Better and worse compared to what?

Fig 11: Generally a very nice figure. Some comments:

• Maybe the color scheme for the landslide probability could be optimized, e.g. change in color where you set your threshold (at values mentioned in L508)
• Why is the lower threshold not slightly higher where recall is still 1?
• When following this line from left to right, are you sure you can increase recall and precision at the same time? This is a univariate model, right? I can’t think of how this would happen. Since you have 5 events, should’t the steps be in intervals of 0.2 for precision?
• Caption: I would simply refer to the equations for definitions of recall and precision, but there you could mention the alternative names (e.g. precision=true positive rate)

L579: yes, compared to some shallow landslide thresholds this is rather low. Could it also be because some of the events are debris flows? The most predictive thresholds for runoff-triggered debris-flows can be at the 10-min timescale.

L610: please specify “hydrologic monitoring”. In this context, I assume soil moisture measurements.

L619: I would say with “few landslide events” instead of “without”. I don’t think you investigated threshold determination without triggering events.

L620-L625: you are of course right that negative events should be considered and in practice it may still be done only occasionally. However, this has been well-known for a while. The first ones I can think of are Staley et al. (2013, https://doi.org/10.1016/j.geomorph.2016.10.019) and Gariano et al. (2015, https://doi.org/10.1007/s11069-019-03830-x) and since then many others have adopted this procedure, some of them you cite earlier.

L639-644: isn’t this contradicting the earlier statement in L623-625 about the value of low precipitation totals? By using precision and recall you get rid of exactly these.