the Creative Commons Attribution 4.0 License.
the Creative Commons Attribution 4.0 License.
Source-composition sensitivity in historical marine wind observations: a CLIWOC case study, 1750–1855
Abstract. Historical ship-logbook observations are widely used to reconstruct pre-instrumental marine climate, but aggregate trends can be sensitive to changes in the source composition of the underlying archive. We reproduce a recent article's principal annual aggregate from CLIWOC release 2.1 and find that this unstandardized, record-weighted series does not on its own identify a global atmospheric change. The observations contributing to it change materially in nationality, archival coverage, and – within Dutch material – format and documented processing history of wind-force values. BIC-based segmentation of the annual aggregate selects working boundaries at 1783, 1816, and 1833. For adjacent selected regimes, descriptive Oaxaca--Blinder decompositions show that the 1816 onset contrast is highly sensitive to nationality weights: under the stated reference convention, the nationality-weight component comprises a large share of the observed contrast, although a within-nationality component remains. The 1833 recovery contrast is likewise highly sensitive to whether Dutch records are pooled or classified by the operational numeric-versus-verbal 'AllWindForces' rule. These results demonstrate non-comparability risks in a pooled CLIWOC aggregate; they do not identify the causal effects of nationality, war, archival practice, or conversion pathway. The earlier 1783 decline is not accounted for by the nationality or geographic specifications examined here, and its cause remains unresolved. The appropriate conclusion is therefore methodological: a raw historical logbook aggregate requires an explicit target population, support and comparability diagnostics, and independent validation before it can support a global-atmospheric interpretation. The analysis provides a source-comparability diagnosis for this CLIWOC aggregate; it does not provide a replacement reconstruction of atmospheric circulation.
Status: open (extended)
- CC1: 'Comment on egusphere-2026-4956', Dagomar Degroot, 31 Oct 2026 reply
-
RC1: 'Comment on egusphere-2026-4956', Anonymous Referee #1, 12 Nov 2026
reply
Review of “Source-composition sensitivity in historical marine wind observations: a CLIWOC case study, 1750–1855” by Malte Rehbein
General comments
The author examines the sensitivity of an annual mean wind-speed series derived from the CLIWOC database to changes in the composition of the underlying historical ship-logbook observations. The study is mainly a response to Degroot (2026), who interpreted this series as possible evidence of changes in large-scale atmospheric circulation. Rehbein shows that the series is sensitive to several non-climatic factors, including changes in the nationality, geographical coverage and provenance of the observations.
The paper addresses the important issue of homogeneity in historical climate records. The analyses are generally well presented, and the author acknowledges that changes in source composition do not necessarily explain all variations in the annual wind-speed series.
My main concern is the scientific contribution of the study. That an unstandardized average of historical marine observations cannot be interpreted directly as evidence of global climate change is well known in climatology. The original CLIWOC researchers were fully aware of this problem.
CLIWOC was not merely a data rescue project. Considerable effort went into the interpretation and quality assessment of historical marine observations. This included reconstructing historical wind-force terminology, converting verbal descriptions into Beaufort values, investigating differences between countries and periods, comparing Beaufort-frequency distributions and analyzing wind anomalies relative to an ICOADS climatology.
An important distinction is that standardizing historical observations does not necessarily make them climatologically homogeneous. The original CLIWOC researchers recognized this and investigated several of its implications. The aim was to make historical observations available and useful for climate research, not to produce a homogeneous global wind-speed series that could be interpreted without further analysis.
Rehbein quantifies the sensitivity of Degroot's annual series to changes in source composition. This is useful, but the paper should explain more clearly what it adds to the original CLIWOC research. Two additional analyses, both closely related to work already carried out within CLIWOC, could help: calculating wind-speed anomalies relative to a gridded monthly climatology and comparing Beaufort-frequency distributions for the Dutch observations.
Major comments
1. Scientific contribution
My main question is what this study adds to our existing knowledge of the limitations of CLIWOC.
The original CLIWOC publications already discussed the problems associated with changing geographical coverage, observing practices and the conversion of historical wind-force terminology. They also investigated these problems using quality-control comparisons, Beaufort-frequency distributions and comparisons with ICOADS.
Rehbein now demonstrates quantitatively that changes in source composition affect the annual wind-speed series used by Degroot (2026). This is a useful result for evaluating that particular study, but the underlying problem is not new.
I also wonder whether some of the statistical analyses are unnecessarily complicated for the questions being addressed. The manuscript uses random forests, BIC-based breakpoint detection and Oaxaca–Blinder decomposition.
In particular, I am not convinced of the added value of the random forest analysis. That wind speed is associated with geographical position, nationality and historical period is hardly surprising. What does this analysis tell us that could not be established from simpler descriptive statistics and the existing CLIWOC literature?
The breakpoint analysis also raises questions. The identified breakpoints are based on an unstandardized series with changing geographical and source composition. Although the author recognizes that these are not necessarily climatic discontinuities, I wonder what is gained by identifying relatively precise breakpoints in such a series.
The Oaxaca–Blinder decomposition is more directly relevant to the research question. It provides a useful way to quantify how changes in group proportions and group-specific mean wind speeds contribute to differences between periods. In this application, it is essentially a decomposition of weighted means. The results are descriptive, however, and depend on the selected groups and reference periods. They do not explain the causes of the observed changes.
I would suggest reconsidering whether all the statistical analyses are needed. Some of the results might be presented more clearly using relatively simple climatological and descriptive methods.
More fundamentally, the author should clarify the purpose of the paper. If it is primarily intended as a critical reassessment of Degroot (2026), a shorter methodological comment or replication study might be more appropriate. If the aim is to make a broader contribution to historical climatology, the paper needs to show more clearly what new insights it provides.
2. Climatological interpretation of the annual wind-speed series (Figure 2 and Section 2.7)
The annual wind-speed series in Figure 2 is based on observations with strongly varying geographical and seasonal coverage. The author recognizes this and investigates geographical sampling effects in Section 2.7. However, these analyses do not account simultaneously for geographical and seasonal differences in the climatological wind field.
I suggest adding an annual series of wind-speed anomalies to Figure 2. For each CLIWOC observation, the corresponding monthly climatological mean wind speed could be subtracted using a gridded reference climatology, for example derived from ICOADS for 1991–2020. Annual averages of these anomalies would reduce the influence of changing geographical and seasonal sampling. It would also be useful to examine the sensitivity to the weighting of individual ships or voyages.
A related approach was already used in the original CLIWOC project. The CLIWOC Final Report (2004, Section 2.4, Figure 8) describes the calculation of monthly anomalies of the zonal and meridional wind components (u and v) relative to an ICOADS climatology for 1961–1990, using a 2° × 2° grid. Although this differs from the scalar wind-speed anomalies proposed here, the underlying idea of accounting for geographical and seasonal differences in the climatological wind field is similar.
The geographical standardizations presented in Section 2.7 explain relatively little of the early decline in wind speed. It would therefore be interesting to see whether this decline, particularly around 1783, also appears in an anomaly series.
Such an analysis would not remove uncertainties related to historical observing practices or Beaufort conversions. A modern reference climatology also introduces its own limitations. Nevertheless, it would provide a straightforward climatological test of the results and a useful comparison with the original series.
I realize that a new climate reconstruction is not the purpose of the paper. However, this relatively simple analysis would help establish whether the observed changes remain after accounting for differences in geographical and seasonal sampling.
3. Interpretation of the Dutch NL-Numeric and NL-Verbal observations (Section 2.8, Tables 5–6 and Appendix E)
The distinction between NL-Numeric and NL-Verbal observations is one of the more interesting results of the paper. The difference in mean wind speed between these groups, approximately 3.1 m/s, is substantial and deserves further investigation.
I suggest comparing the normalized frequency distributions of wind force for both groups, preferably using the original Beaufort classes rather than the derived wind-speed equivalents. This would show whether the difference is caused by a general shift in the distributions or by differences in the frequencies of particular Beaufort classes.
Koek and Können (2005, Section 5, Figures 2 and 3) compared Beaufort-frequency distributions for Dutch CLIWOC observations before and after 1800 with Dutch ICOADS observations from 1854–1938, using a common North Atlantic region. Interestingly, they found that the Beaufort values assigned by the nineteenth-century compilers of the Dutch extract logbooks agreed better with the ICOADS distribution than the values derived from verbal descriptions during CLIWOC. They concluded that there might be room for improving the CLIWOC conversion of Dutch wind-force terminology. This earlier finding seems particularly relevant to the large difference between Rehbein's NL-Numeric and NL-Verbal groups.
There is also an important historical distinction. In the nineteenth-century Dutch extract logbooks, numerical Beaufort values were often derived from originally verbal wind-force descriptions. The distinction between NL-Numeric and NL-Verbal therefore does not necessarily mean that the original observations were numerical in one group and verbal in the other.
The author acknowledges that the classification is based on the format of the AllWindForces field and does not identify the original processing history. Nevertheless, I think the historical background deserves more attention.
Ideally, the frequency distributions should also be compared for overlapping geographical areas, seasons and periods. This would help determine whether the large difference between the two groups persists when differences in sampling are reduced.
A comparison with the original CLIWOC findings might also help establish whether the observed differences reflect known conversion problems or reveal something new about the historical observations.
Minor comments
1. Breakpoints (Section 2.3)
The breakpoints around 1783, 1816 and 1833 are derived from an unstandardized series. Although the author recognizes that these are working boundaries rather than climatic discontinuities, this distinction should remain clear when discussing possible historical explanations.
2. Historical Beaufort conversions (Section 2.1)
Most CLIWOC wind-speed values are derived from verbal wind-force descriptions rather than instrumental measurements. The conversion from historical terminology to Beaufort values, and subsequently to wind-speed equivalents, introduces uncertainties that may differ between countries and periods. This could be explained more clearly.
3. Wind-speed units and consistency (Section 2.8, Table 5)
The CLIWOC W field is expressed in tenths of meters per second. For clarity, I suggest reporting wind speeds in m/s or explicitly stating the units in the table captions.
There also appears to be an inconsistency between the text in Section 2.8, which gives mean values of 77.18 and 49.94 for NL-Numeric and NL-Verbal, respectively, and Table 5, which reports 78.85 and 47.60. Please check these values.
4. Statistical independence of observations
Observations from the same ship or voyage are not statistically independent. The author uses ship-cluster bootstrap procedures for some analyses but record-level bootstrap confidence intervals for several decompositions. It would be useful to explain this choice and its possible effect on the reported uncertainty.
5. Conclusions
The conclusions should distinguish clearly between limitations of CLIWOC itself and limitations of the particular application examined here. The results do not imply that CLIWOC observations are unsuitable for climate reconstruction. Rather, they illustrate the importance of taking account of geographical coverage, observing practices and historical data processing, issues that were already central to the original CLIWOC project.
Citation: https://doi.org/10.5194/egusphere-2026-4956-RC1
Interactive computing environment
Software Open CLIWOC Aggregate Analysis Authors/Creators Malte Rehbein https://doi.org/10.5281/zenodo.21951456
Viewed
Since the preprint corresponding to this journal article was posted outside of Copernicus Publications, the preprint-related metrics are limited to HTML views.
| HTML | XML | Total | BibTeX | EndNote | |
|---|---|---|---|---|---|
| 293 | 0 | 3 | 296 | 0 | 0 |
- HTML: 293
- PDF: 0
- XML: 3
- Total: 296
- BibTeX: 0
- EndNote: 0
Viewed (geographical distribution)
Since the preprint corresponding to this journal article was posted outside of Copernicus Publications, the preprint-related metrics are limited to HTML views.
| Country | # | Views | % |
|---|
| Total: | 0 |
| HTML: | 0 |
| PDF: | 0 |
| XML: | 0 |
- 1
At the same time, this analysis seems to be the product not just of statistical analysis and literature reviews undertaken by frontier AI systems - as the article acknowledges - but also of writing that appears almost entirely AI-generated. It reads like the (human) author not only asked a frontier LLM to use everything statistical method in the book to challenge work that relied, in part, on an LLM system from a couple years ago, but also asked that LLM to write up the results. This may be where scholarship is headed now, but I think the article would be stronger if it was written in the author's voice. Among other things, that would avoid many of the repetitive methodological qualifications and cautions.
In addition, it's worth noting that "Recovering Climate's History with Artificial Intelligence" provided an initial, speculative reconnaissance of what LLMs might be able to do with a big dataset, rather than firm conclusions. This article rightly cautions against firming up those conclusions without considering the nationality of logbook authors, for example - but that is not the same as challenging established findings.