see attached PDF

This study by Yu et al. considers short-term (12h) mean forecast errors (biases) in North-Atlantic extratropical cyclones using ERA5 reanalysis data. The authors distinguish between cyclones with strong and weak latent heat release during the time of strongest intensification. The results show a low bias of cyclone intensity, more prominent so for cyclones with strong latent heat release, as well as structural differences between the strong and weak latent-heating groups.

The study has a clear focus and the presentation of results in the text and by the figures is very good. The manuscript is fine from the perspective of reporting observed patterns. Unfortunately, however, it is not clear to me what we learn from these observations. I expand on this issue below.

Before publication, the manuscript needs at least to discuss a fundamental caveat of the design of the study, which is not acknowledged in the current version. While I do not have many comments, I recommend major revisions before potential publication.

Kind regards

Main issue: Conditional verification

The authors consider conditional verification, i.e., they condition their examination of forecast errors on the existence of a cyclone in the analysis. A cyclone constitutes a (strong) anomaly from the climatological mean. Forecasts tend to underestimate analysis anomalies, not because of inherent biases in the forecast system but as an inherent feature of conditioning the verification on analysis anomalies! Supposedly, this effect of conditional verification is well known in the verification community, but less so in academia (Mark Rodwell, personal communication). Unfortunately, I am not aware of a reference in the literature of this effect. The effect seems plausible when considering forecasts that have lost all skill. An average of such forecasts represents a climo state. Forecasts without skill thus evidently underestimate on average any anomaly from the climo state when conditioned on anomalies existing in the analysis. Forecasts with less than 100% skill hence exhibit the tendency to underestimate anomaly amplitude in conditional verifications. This effect may be accentuated in the current study by the authors’ choice to condition on maximum intensification rate.

Most of the signal that the authors find is a low bias in intensity. As frequently noted by the authors, most of the other features they discuss are consistent with this low bias. The low bias in turn may be a mere artifact of the authors’ conditional verification. Hence, what do we learn in this study about the role of latent heat release in cyclone forecasts? Other than the well-known effect of latent heat release to additionally intensify cyclones. How can we then distinguish between structure bias that is consistent with low intensity bias and additional biases due to latent heat release?

And as a corollary of the above discussion: To what extent are the bias structure differences between the weak and strong heating groups mere reflections of the structure differences between cyclones with weak and strong latent heat release?

Other non-minor issues:

Section 2.1.: It remains unclear to me what the forecast data is that is examined by the authors.

Confusion of (unbiased) forecast errors vs. bias and error source vs. amplification:

To be clear upfront: I have no doubt that (mis)representation of diabatic processes induces forecast biases.

The theme of this study are biases. The introduction, however, also discusses the role of (moist) diabatic processes in the amplification of (mostly unbiased) forecast errors. This discussion can be interpreted such that this contribution to error growth is due to deficiencies in the representation, i.e., the model’s parameterizations of the diabatic processes. By no means, however, such deficiencies need to dominate or even make a prominent contribution to error growth. Cyclone amplification with strong latent heat release is a highly nonlinear process and one may expect error growth in such a situation even with perfect representation of latent heat release, as e.g., seen in spread growth in ensembles or in perfect model experiments.

Similarly, the introduction contains a discussion of diabatic processes as sources of forecast errors, which states that “forecast biases and errors in extratropical cyclones have increasingly been attributed to deficiencies in the representation of diabatic processes” (L44ff). At least two of the references cited by the authors to this end (Lamberson et al. 2016 and Pickl et al. 2023) conclude the opposite in their abstracts: Lamberson et al. attribute the main differences in the cyclone evolution to initial condition uncertainty in the upstream trough, and so do Pickl et al. (Berman and Torn (2019, 2022(?)) found similar results). In addition, Pickl et al. consider latent heat release in warm conveyor belts as amplifier of forecast uncertainty – as diagnosed in an ensemble – hence this amplification does not rely on “deficiencies in the representation of diabatic processes”.

In L173 the authors note that “This bias might be attributable to error sources from both microphysics and dynamics.“ and subsequently attempt to disentangle the distributions. While disentangling is very difficult, the study would benefit at least from a clearer discussion of the background knowledge in the introduction.

Minor comments

L25: Is there a distinction between systematic forecast errors and biases? I suggest rephrasing, e.g., leaving out „systematic“.

L35: I suggest adding a reference to Davis, Stoelinga, and Kuo 1993, which in my view is a seminal paper on quantitative diagnostic of the impact of latent heat release on cyclones.

End of introduction: I’d appreciate a brief outline of the paper here.

L105: It is not obvious to me that the distribution of mesoscale features aligns with the direction of motion. Can you provide a brief explanation/ illustration. Is this a heuristic or based on theory/ conceptual understanding?

L123: I do not understand why the *analysis* cyclone needs to be re-centred. Please explain.

L128ff: More than 20% of the data is lost for the weak heating group but less than 5% for the strong heating group. Why is this difference so large? On a related note: Did the authors consider also an EOF/ PCA approach to better understand the modes of variability of the systematic errors/ biases. I recommend considering such an approach, which can be applied without loss of data.

L138: I would not say intensity underestimation is restricted. Rather, the maximum of underestimation is in the center, whereas the maximum of underestimation in the strong-heating group is at larger radii, ahead and left of motion.

L149: I do not understand this potential explanation. Please clarify.

L187ff: I do not understand why this is consistent. Please expand your argument for clarity.