The manuscript submitted by Liu et al. to the journal Climate of the Past presents relevant quantitative climate reconstructions (mean temperature of the coldest and warmest months and moisture index calculated as the ratio of actual evapotranspiration to equilibrium evapotranspiration) at the global scale for millennial-scale climate variability in the 30-50 ka interval. These reconstructions are the first attempt to quantitatively document spatial patterns of atmospheric changes in temperature (winter and summer temperatures and hence seasonality) and humidity. These reconstructions are also crucial for testing the transient simulations of D-O events planned in the framework of the Palaeoclimatic Modelling Intercomparison Project. In my opinion, this manuscript should be accepted after the minor revisions listed below:

Abstract and Discussion and conclusion

Lines 18-20 and - The authors say : « The change in winter temperature was not significantly different from the change in summer temperature, and thus there is no evidence that the D-O events were characterised by a change in seasonality. ».

Lines 190-192 « There is no indication of a significant difference in the temperature change during winter and summer, and thus no indication of a large shift in seasonality as inferred from some site-based reconstructions (e.g. Zander et al., 2023). »

These sentences are not clear to me. The recent study by Zander et al. (2023, see also Denton et al., 2022) shows that during D-O cooling events, in particular during Heinrich events, there were no substantial changes in summer temperature in Europe, suggesting that abrupt millennial-scale events were defined by colder and longer winters and, consequently, strong seasonality. Reconstructions by Liu et al. show that during D-O warming events winter and summer temperatures change in a similar way. The results of Liu et al. do not necessarily disagree with those of Zander et al. Seasonality may decrease during D-O warming events as both winter and summer temperature increase in parallel, while during D-O cooling events winter temperature decreases strongly compared to summer temperature, leading to an increase in seasonality during D-O cooling compared to D-O warming. Therefore, there could be a substantial change in seasonality during D-O events.

Introduction

Lines 35-37 – Something to highlight here is the contribution of this work to document and understand the « regionalisation » of global warming, one of the major challenge of the IPCC.

Lines 45-51 – One of the first quantitative cllimatic reconstructions of well identified D-O cycles from two deep-sea pollen records based on the Modern Analogue Technique was published by Sanchez Goñi et al. in 2002. Please add this reference.

Methods

Lines 77-79 – Merging Quercus deciduous with Q. evergreen in a unique Quercus morphotype may have had strong implication for reconstructing the seasonality of precipitation in the Mediterranean region. Please add some discussion on that.

Line 79 – What does it mean « <10 occurrences » ? Less than 10% ? If this is the case, this threshold seems to me too high. For instance, Olea that it is an important climatic indicator hardly reaches 10%.

Lines 91-92 – Please replace « continental-shelf marine sites » with « deep-sea sites ». No one of the marine records included in ACER is located in the continental-shelf area as the sedimentation in this area only preserves the Holocene period.

Age modelling

Lines 123-124 – Maybe it is worth to say here that this allignment is supported by the global synchroneity of D-O warming events shown by the well-dated speleothem records (Corrick et al., 2020).

Line 212 – Replace « …this in necessary… » with «…this is necessary ».

Lines 220-221 – In contrast with author statement, there is a recent pollen record documenting the Indian monsoon D-O climatic variability by Zorzi et al. (2022). It would be interesting to apply the same methodology to this record that qualitatively show the increase of the Indian monsoon during the D-O warming events.

Line 233 – Please add the above one (Zorzi et al., 2022).

In the figures, I do not see the results (circles) of the quantitative climate reconstruction of the MD95-2042 and MD99-2331 cores. Both cores have a high mean temporal resolution of 300 and 250 years for the studied interval, respectively. Could you explain why?

Line 413 – Add a space between « (a) » and « from ».

References

Denton, G. H., Toucanne, S., Putnam, A. E., Barrell, D. J. A., and Russell, J. L.: Heinrich summers, Quaternary Science Reviews, 295, 107750, https://doi.org/10.1016/j.quascirev.2022.107750,2022.

Sanchez Goñi, M.F., Cacho, I., Turon, J.-L., Guiot, J., Sierro, F.J., Peypouquet, J.-P., Grimalt, J.& Shackleton, N.J.(2002). Synchroneity between marine and terrestrial responses to millenial scale climatic variability during the last glacial period in the Mediterranean region. Climate Dynamics 19: 95-105

Zorzi, C., Desprat, S., Clément, C., Thirumalai, K., Oliveira, D. Anupama, K., Prasad, S., Martinez, P. (2022) When Eastern India Oscillated Between Desert Versus Savannah‐Dominated Vegetation. Geophysical Research Letters, 49(16) e2022GL099417

This paper presents global pollen-based quantitative temperature and moisture reconstructions for DO events during MIS3 using a novel reconstruction technique the authors called fxTWA-PLS v2. One of the main findings is an absence of difference between winter and summer changes, which challenges some recent hypotheses. Overall, the manuscript is nicely written, and everything in it is easy to follow and understand. I particularly appreciated the data treatments presented in sections 2.3 and 2.4. This was very nice. My main criticism concerns all the elements of the analysis that have been omitted from the manuscript, rendering its evaluation quite challenging. I detail these below.

In summary, this paper could make a substantial contribution to the field. However, I must ask for major revisions to ensure that the results are as robust as the paper claims.

All the DO results discussed in the paper are derived from data that needs to be explicitly presented. I need to see the raw data used to generate this article's main conclusions. Pollen-based climate reconstructions can be pretty noisy when not off-target if improperly controlled. I am confident the authors did all the necessary checks, but I still want to see the data. Note: Yes, they are available in that GitHub repository, but the paper should be convincing by itself. And it is short enough to take on a few more details.

I also wonder what the impact of amalgamating all the pollen data at the genus level is. I am thinking here about some of the euro-Mediterranean taxa, such as Quercus, which would have a broadleaf and an evergreen version. Amalgamating these two will undoubtedly impact seasonality reconstructions. Have you considered refined the classification of some key taxa that may have a direct impact on the reconstructed seasonality?

I find Fig.1 misleading. Like the two South African sites, many sites are included here, even if they are not used subsequently. It almost looks like a way to beef up the study's numbers: “based on 73 sites” instead of my roughly estimated “based on 25-30 sites.” Pollen records documenting DO events are rare, and that’s okay. Once this list is reduced according to the study, it would also be more ethical to cite the original reference of all the datasets used. ACER is an excellent archive/repository but can only exist because other colleagues have generated the data. ACER contains information about these original publications. Please use it.

I want more details about the calculation of the anomalies. This problem is not trivial and fundamental to this study since the DO results are presented as climate anomalies. Anomalies to what? How were they calculated? A particular emphasis on how these were calculated for marine records is warranted if any of these are kept in the final 25-30 records.

Unfortunately, I have major concerns about the reconstructions themselves, although I am sure these doubts will be lifted when more details are provided. At the moment, I understand that the authors did not regionalise their calibration dataset and used the same transfer function based on all their modern samples to analyse data from Australia, Asia, Africa, North America, and Europe. This methodological choice needs to be more challenged in the paper, and the authors should justify why they think this is not a problem (it most likely isn't!). Using as much calibration data as possible is not standard with regression techniques, even if all the novelties added to the standard WA-PLS should give it more flexibility. Based on this global mixing of samples and taxa amalgamation, I am not overly surprised that seasonal differences do not appear in the results. These specific points make me seriously question the main findings.

I am also worried about the RMSEs presented in Table 2. They are enormous, with a mean error of 6.5°C and 3.7°C for winter and summer temperatures, respectively. An R2 has little meaning when dealing with so much data, so I wouldn't give this too much importance. Regionalising the pollen data will undoubtedly shrink these to more adequate values. But this is not just a game of optimising errors. These errors are of the order of, if not larger than, the signal reconstructed (Fig. 3). This is highly problematic because your key results may be background noise. You need to demonstrate how that’s not the case.

More specific comments:

L19: Here and elsewhere, Europe is part of Eurasia. So, it is either Eurasia or Europe and Asia.

L21-22: “broadscale features of temperature change” is too vague. Please specify which features you have in mind here.

L108-109: Clarify how you define two sites as “geographically and climatically similar”. What thresholds do you use?

L115-116: You mention some sample-specific errors. The method is appropriate, but these errors are not used in any of the subsequent analyses. There is potential to do better on that front.

 L136: Interpolating all records at 25 years is a bit ambitious. I do not think any of the selected records match this resolution. This artificial high-resolution may bias statistical tests performed on the data since the errors behave in ~1/n, n being the number of points. Please adjust to using an interval more consistent with the studied data.