Main comments 

This manuscript derives a new daily weather regime classification for April–July over the US and examines tornado frequency in those regimes. 

1. The regime classification methodology differs from standard ones (EOFs are not used, there is no time filtering, once per day hourly snapshots of 500 hPa heights are used, the choice of cluster number is subjective, there is no normalization of variance), and there is no indication how these methodological choices impact the results. Variance normalization is important because 500 hPa height anomalies have substantially greater variance in April than in July. The k-means clustering method minimizes within-cluster variance, and seasonality in variance might bias the results.   Consequently  it may well be the case that the weather regime frequencies vary with month (climatologically), which would confound any analysis with tornado frequency whose climatology also varies by month. Whether this is case or not is unclear because diagnostics such as the seasonality of regime frequency, variance explained, association with modes of large-scale variability, etc. are missing. 

2. The new regime classification is not compared with previous ones from the same authors for April and May and with year-round regime classifications from Lee et al., (2023) [data is in Zenodo for download]. Making connections to previous work would increase the value of the current work. The classification data (data needed to classify independent data and classification of the days in the study) should be provided. 

3. The dependence of tornado activity as well as the dependence of CAPE and shear on month does not seem to have been accounted for in the analysis. In both cases anomalies are computed with respect to the April–July average. Using anomalies with respect to the April–July average means that the anomalies of quantities with a seasonal cycle will appear correlated but might actually be unrelated after accounting for seasonality. 

4. Along with seasonality, ENSO may be another factor/alternative hypothesis to consider.

Detailed comments

Abstract. "Our study highlights the potential application of WRs for better seasonal prediction of tornado activity." The authors' previous regime/tornado study examine subseasonal prediction of weekly regimes and found that forecast skill was lost at about Days 7–13. Is there evidence that these regimes are predictable on seasonal time scales?

The Weather regime methodology differs substantially from that used commonly in the literature. There are no explanations provided why. The weather regime classification method lacks standard diagnostics and assessments of robustness. The classification data is unavailable which means the classification cannot be applied by others to independent data and cannot be compared with other classifications (e.g., Lee et al., 2023 which provides the data)

Line 106. "500H at 21 UTC was used to represent the daily circulation patterns." Previous weather regime classifications have used daily means and subsequently smoothed those in time, e.g., 10-day low-pass-filtered (Grams et al., 2017, 2020, Lee et al., 2023)

There is no EOF filtering which differs from previous work (Michelangeli et al., 1995, Grams et al., 2017, 2020, Lee et al., 2023)

Line 121. "the number of clusters was determined as five using the elbow method." 

From the reference cited, the elbow method "is a visual method. The idea is that Start with K=2, and keep increasing it in each step by 1, calculating your clusters and the cost that comes with the training. At some value for K the cost drops dramatically, and after that it reaches a plateau when you increase it further. This is the K value you want." This is not really an objective method. Lee et al., 2023 apply four objective, data-driven methods for determining the best number of clusters, including the classifiability and reproducibility indices of Michelangeli et al. (1995).

There is no variance normalization to account for seasonality of variance (Grams et al., 2017 Lee et al., 2023). During the April–July period examined, Lee et al., (2023) found that the domain averaged Z500 std varied from 80 m in April to 50 m in July. Removing the daily climatology does not account for seasonality of variance.

Because k-means cluster analysis minimizes the total within-cluster variance, seasonality in the variance of the data means that clusters might be biased toward the later months of June and July when variance is small and consequently within-cluster variance is easier to minimize. Consequently the resulting cluster centroids are likely to be skewed toward patterns that best represent June/July variability at the expense of other months. And indeed, cluster A shows reduced activity which would be typical of the June/July period. This bias is potentially a serious flaw for the application here since US tornado activity is much higher in April–May than in June–July. In other words, tornado activity might be substantially higher in a particular weather regime simply because that regime is more frequent during calendar months when tornado activity is climatologically higher. 

Whether this is the case, and the association between tornado activity and regime frequency is simply due to their having similar seasonal cycles, is impossible to say because the authors have as far as I can see failed to provide any indication of the seasonality of cluster frequency or how the variance explained depends on month. 

Overall there are essentially no diagnostics of the WRs such as variance explained. Also there is no assessment of how these regimes vary with large-scale modes of variability such as ENSO (known to be important for tornado activity), NAO, etc. 

Line 134. The tornado probability anomalies fail to account for seasonality because they are with respect to the April–July frequency. Consequently, substantial anomalies may occur simply because some regime are more frequent during April–May rather than June–July. I see no exploration in the text of this hypothesis. This issue applies to environments as well reports. 

Line 160. "These WRs have some spatial similarities to the year-round WRs found by Lee and Messori, (2024)." The more appropriate citation is Lee et al., (2023) which describes the classification in detail, and is not cited here. Also Lee et al., (2023) provide that classification data which means that authors here can make a more precision statement regarding the similarity of the classification. That is, with what frequency are the classifications the same. Also applying the diagnotic methods of Lee and Messori, (2024) to the regimes here would provide some evidence that regimes here are physically or dynamically meaningful. 

Line 165. "Composite anomalies of these parameters were calculated by subtracting the corresponding climatological mean." The same issue of using an inappropriate climatology applies to the MUCAPE and S06 anomalies, i.e., they will have seasonality both in their mean and variance. Because the April–July climatology is used, MUCAPE anomalies will tend to be positive in later months and negative in earlier months, and the opposite for S06. This means that any seasonality in the regime frequencies will project onto these anomalies, even if there is no relation when seasonality is taken into account. Moreover statistical significance tests used (e.g., Fig. 2) are inappropriate because they assume identically distributed but the variance of the data depends on month.

Line 136. Tornado data from period 1960–2022 is used and it is claimed (line 132) regarding well-known report trends that "this trend is not reflected in TDs" (tornado days). However, Miller et al., (2020) with two of the same three authors conclude that the period 1990–2019  "represents a compromise between data set length and an allowance for a significant fraction of the reports to have occurred during the Next‐Generation Radar era and thus have undergone some quality control." Moreover, Fig. 5a shows a very large, presumably secular shift in the number of tornado days, which is as large or larger than the year-to-year variability. The authors state later (line 294) that "modelled TDs are nearly out of phase with observations in the 1960s, when tornado reports are less reliable" which supports analysis on a shorter period. 

Line 210. "may be possibly linked to tropical cyclones (Figs. 1e and 2e)." This is an interesting point and should be verified using data here https://www.spc.noaa.gov/exper/tctor/ and https://www.spc.noaa.gov/publications/edwards/tctor.xls

Figs. 3 and 4 mention resampling to assess statistical significance without details. Depending on the design of the resampling procedure, the results may be incorrect if seasonality is not accounted for. For instance, in the case of a permutation test a possible way of taking seasonality into account is to compare tornado day frequency in say regime A with tornado day frequency on exactly the same calendar days when regime A did not occur. 

Fig. S4 "WR TD time series" it is unclear what this quantity is. If it is the number of tornado days in that weather regimen, then of course, there are fewer tornado days in that weather regime in years when that weather regime is less frequent (the very high correspondence between the blue and green curves). This would be the case even when there is no relation between tornado days and weather regimes. That being the case, reporting the correlation coefficient does not seem informative and might confuse some readers. (I think my interpretation of the green curve is correct because it is different in panels S4a-e.)

Line 290 and Fig. 5. "the empirical model fails to capture the observed decreasing trend or the decadal shift in the 1980s." The use of Spearman correlation here may obscure the extent to which empirical model fails to capture observed variability. A scatterplot would likely give a much more accurate and pessimistic picture. The association is stated to be statistically significant at the 5% level but visually is hard to see. Perhaps a bootstrap test might give a more credible  assessment of statistical significance. 

Line 312 and also the abstract. "the empirical model captures the interannual variability of TDs reasonably well" this seems an overly generous description. 

Conclusions. Line 3.43. "A year that includes a high number of WR-B days is likely to have an above average number of TDs and TOs." Is there analysis/figure in the manuscript that supports this statement?

Line 285. "The frequencies of persistent WRs also show changes across different multidecadal time periods (Fig. S4f" I really don't see any substantial changes in Fig. S4f and I question whether a t-test is suitable for a change in frequency, perhaps Fisher's exact test.

Grams, C., Beerli, R., Pfenninger, S. et al. Balancing Europe’s wind-power output through spatial deployment informed by weather regimes. Nature Clim Change 7, 557–562 (2017). https://doi.org/10.1038/nclimate3338

Grams, C. M., L. Ferranti, and L. Magnusson, 2020: How to make use of weather regimes in extended-range predictions for Europe. ECMWF Newsletter, No. 165, ECMWF, Reading, United Kingdom, 14–19, www.ecmwf.int/en/newsletter/165/meteorology/how-make-use-weather-regimes-extended-range-predictions-europe.

Lee, S. H., M. K. Tippett, and L. M. Polvani, 2023: A New Year-Round Weather Regime Classification for North America. J. Climate, 36, 7091–7108, https://doi.org/10.1175/JCLI-D-23-0214.1.

Michelangeli, P., R. Vautard, and B. Legras, 1995: Weather Regimes: Recurrence and Quasi Stationarity. J. Atmos. Sci., 52, 1237–1256, https://doi.org/10.1175/1520-0469(1995)052<1237:WRRAQS>2.0.CO;2

The authors focus on the characterization of weather regimes that favor tornado outbreaks, and the persistence of such regimes. While an interesting presentation, their work does tread a similar vein to a number of recent studies, which aren’t fully acknowledged in the introduction. There are several points where it is difficult to follow the methodological tests taken, and the applications of statistical significance tests are not clear as written. A few of the figures also need to be adjusted for accessibility. Finally, while the authors claim that these results will help with seasonal prediction, I’d actually like to see a little more of a connection as to how they envisage this improving on the existing paradigm. While these issues exist, they are not overly burdensome, which leads me to recommend a set of minor revisions.

Minor Comments:

Introduction, First Paragraph: I’m not convinced the approach taken by the authors here really fits with the manuscript. Talking about trends and variability does not connect well with their primary focus on the occurrence of weather regimes favoring tornado outbreaks at least as presented. The authors could stand to connect their work to analysis of trends and identifying what the gaps may be, which would help the flow of the text.  

Line 48: The literature cited here portrays an incorrect precedence of the understanding derived. Arguably, any such reference to ENSOs influence on tornadoes should include Cook and Schaefer (2008), which was the first study to show a wintertime signal, and a reference to Allen et al. (2015), which showed the first statistically robust connection to the environment in springtime, 500 hPa geopotential height and inferred cyclonic track, and observations, and developed the first seasonal prediction algorithm – which is directly relevant to the analyses presented here.

Paragraph beginning Line 73: It would seem appropriate to reference that the European community has long used weather regimes to look at relationships with significant severe weather events. See Punge and Kunz (2016) and references therein.

Paragraph beginning Line 62: It seems strange to me that the literature cited for weather regimes does not include Tippet et al. (2024). The subsequent references to this paper also seem to neglect that study considering variability.

Line 108: It is not clear what is meant by the authors by the statement ‘and to avoid too many near-zero CP values in a daily mean’, suggest clarifying.

Line 116: What mean long term trend was used for the detrending procedure exactly? Given you are interested in variability, to what extent is the trend removed influenced by the ending year?

Paragraph beginning Line 154: It would seem more appropriate to compare to the weather regimes of Tippett et al. (2024), given that these are tornado focused.

Line 221: How is the significance t-test performed here, and how is it indicated on the diagram and for what specifically? This is unclear.

Line 362: The extent to which the WR would help with skillful seasonal forecasts is an interesting one. Given that existing seasonal prediction models do not incorporate weather regimes, and offer skillful forecasts, particularly when climate variability is strong (e.g. Allen et al. 2015, Lepore et al. 2018), I would encourage the authors to discuss what advantages applying the WR approach would achieve over existing models, and how that may contribute to more skillful forecasts rather than the current abstract statement.

Figures: There is some colorblindness suitability issues with the figures, particularly green and red overlapping contours. Please address this in the revision.

References:

Allen, J. T., Tippett, M. K., & Sobel, A. H. (2015). Influence of the El Niño/Southern Oscillation on tornado and hail frequency in the United States. Nature Geoscience, 8(4), 278-283.

Cook, A. R., & Schaefer, J. T. (2008). The relation of El Niño–Southern Oscillation (ENSO) to winter tornado outbreaks. Monthly Weather Review, 136(8), 3121-3137.

Lepore, C., Tippett, M. K., & Allen, J. T. (2017). ENSO‐based probabilistic forecasts of March–May US tornado and hail activity. Geophysical Research Letters, 44(17), 9093-9101.

Punge, H. J., & Kunz, M. (2016). Hail observations and hailstorm characteristics in Europe: A review. Atmospheric Research, 176, 159-184.