1.5ºC and all that

The Paris Agreement temperature limits are a little ambiguous and knowing where we are is tricky.

The desire to keep global temperature rises since the pre-industrial, ideally below 1.5ºC and definitely below 2.0ºC, is a little bit complicated by the lack of definition in what constitutes the pre-industrial, uncertainties in what the temperature was in the pre-industrial, and an ambiguity in what counts as exceeding these limits.

These are old questions that were tackled in prior IPCC reports (SR15 and AR6), but there are new elements that have come in to the equation since then and, of course, the real world has got closer to exceeding 1.5ºC and so there is additional salience. There is a big collective effort going on to provide some clarity on these broader questions (that has just been submitted), but I thought it would be interesting to review some of the more technical questions here.

First off, when is the ‘pre-industrial’? It’s not really definable from history without going back to periods when we didn’t have any (or much) instrumental data, and as we’ve discussed, anthropogenic impacts on climate might date back to the dawn of agriculture. So, for practical reasons, people have settled on the 19th Century as ‘close enough’. But even there, we have issues. The early 19th C was abnormally cold because of a series of big volcanoes (incl. Tambora in 1815), so that shouldn’t be included if you want to highlight anthropogenic changes. In any case, the instrumental data sets for temperature are mostly only good for the global mean (with some relevant uncertainties) from 1850 onward (though there are some good efforts to push this back further e.g. Lundstad et al. (2025)). And since you need a few decades to smooth out the internal variability, people have been using various multi-decadal averages around then. While there are still a few holdouts, most folks have followed the IPCC lead and are now using 1850-1900 as the baseline for ‘pre-industrial’ in practice.

There are now at least four data sets that are regularly maintained and that go back to at least 1850: HadCRUT (currently v5.1), NOAA (v6), Berkeley Earth and, a relatively new effort, DCENT. They use different (though overlapping) raw data, different methods, and different interpolations, and thus (unsurprisingly) give different magnitudes of change since 1850-1900. With respect to their own baselines, 2024 was 1.45-1.65ºC above 1850-1900. If they are aligned in the modern period (when the differences between the methods/data are minimal), there is clearly a variation in both inferred interannual variability and mean change in the ‘pre-industrial’ period (see fig. 1). How should this be interpreted? It’s not the full structural uncertainty (since we are not really sampling all of the issues – particularly in the SST products), but it is perhaps a lower bound on that uncertainty. Ensembles that sample the methodological uncertainty are also useful, of course.

Other datasets such as GISTEMP or JMA, or the more modern reanalyses (ERA5, JRA-3Q etc.) that don’t extend that far back, can still be useful because they add to our understanding of the structural uncertainty in the periods where they overlap. The WMO uses a mix of these records (in 2024 it used an average of HadCRUT, NOAA, Berkeley, GISTEMP, JMA-3Q, and ERA5) when they are available to create a composite record. But how do we get the change since the pre-industrial?

As we discussed earlier this year, one way would be to baseline each long record to their own 1850-1900 data, and then add in the shorter records by tying them to one master record (but which one?) or an average of the longer ones. However, if you plot this out it gives the impression that all the uncertainty is in the modern period. Using a modern period to cross calibrate the different records (as in figure 1) but then imposing a constant offset to translate from the modern to the pre-industrial allows the uncertainty to be clearly associated with the past (not the present) (as in figure 2). But how should the offset be calculated? We could either assume that the average of the four long records should be zero over the pre-industrial period or that the WMO average should be zero (or even something else entirely).

How much does this matter?

First, the baseline issue. With respect to 1850-1900, using an average of the 4 records mentioned above, 1880-1900 is 0.01ºC cooler and 1880-1920 (which Jim Hansen uses) is ~0.02ºC cooler. These are small numbers on average, but the spread across the 4 records is large (±0.08ºC) indicating that this is not very robust and could change in future. Second, the difference between setting the WMO average, or the average of the 4 long records, or the average of the records that were used in AR6 to zero, can make a ~0.04ºC difference.

For individual months, there is a secondary issue – should you adjust the climatological baseline on an annual basis or on a monthly basis? Different months have warmed differently since the 19th Century (Jan/Feb/Mar have warmed by about 0.17ºC more than Jul/Aug/Sep). That is, do you consider the anomaly for Oct to relative to the climatological Oct (which seems sensible) or to the climatological annual mean? (which is slightly easier). For Berkeley Earth, October 2025 was at 1.52º above pre-industrial Octobers, or 1.57ºC above if baselined annually. Winter months are affected oppositely. Depending on the month it is an effect of ±0.08ºC. Note this is only an issue for the monthly anomalies w.r.t. to a different baseline than the native baseline (usually modern) for any particular product.

Finally, given that all of these approaches rely on moving targets (which records are being maintained, raw data being added through data rescue efforts, updates to method versions etc.), one has the choice of either updating these calculations every year (which means you need to explain why things might change from previous years), or sticking to a canonical calculation (such as the one in AR6) for consistency. The best estimate of the annual offset from 1981-2010 to the 1850-1900 period was estimated as 0.69ºC in AR6, but following an analogous recipe now would give 0.73ºC (mainly because DCENT has a colder 19th C than the older products, and updates to HadCRUT and Berkeley Earth have shifted things slightly).

The offset to tie the shorter records to the longer ones also varies over time if you keep the same method. For GISTEMP, I’ve been calibrating to the other records over the 1880-1900 period. Last year, that gave an offset of -0.028ºC to go from 1880-1900 to 1850-1900, but this year (with the addition of DCENT and minor updates to the raw data), it gives an offset of -0.01ºC. Copernicus uses a fixed 0.88ºC offset (from AR6) to go from a 1991-2020 baseline in ERA5 to 1850-1900, but following an analogous recipe and adding in DCENT, you’d end up with 0.92ºC.

Last year was the first in which we “likely” exceeded 1.5ºC in the annual average (the WMO average was 1.55ºC), and the assessed uncertainty in this (arising from all the mentioned issues) is about ±0.13ºC (90% CI). With updates to the records, the WMO average would now be 1.54ºC. But if you added an offset to the 1981-2010 baselined data so that the average of the four long records was zero, 2024 would be at 1.58ºC. Adding an offset so that the WMO average was zero over the baseline takes you back to 1.54ºC.

On a monthly basis we have been exceeding 1.5ºC in the individual records (briefly) since the El Niño event of 2016 (maybe 2017 and 2019, and then again in 2020). Since 2023, we have exceeded it on a monthly basis more often and that has been sustained in 2024 and 2025 (figure 3).

Knowing when we’ve exceeded the limit with respect to a longer term average is trickier again except in hindsight. If we want to know if we’ve gone past the mid-point of the first twenty year period above the threshold, that involves some forecasting for the next ten years – which adds to the uncertainties. We have many forecasts – from CMIP, initialized projections, statistical fits, even machine learning – but there are many uncertainties (in the projected forcings, the structure of the fit, the appropriateness of training data). So this too will be something that is subject to (annual?) revision, and the precise answer might not be available for a while. Whether it matters if it turns out (in a decade or so) to have been 2028 or 2030 or another year is not obvious to me.

Summary

There are some irreducible uncertainties in defining where we are with respect to the pre-industrial at any one moment (day, month, year, decade), and so one shouldn’t expect to know this precisely, and one can expect a bit of ‘jitter’ in the assessments. Right now, while we are hovering around the 1.5ºC level, differences in method can move the value slightly above or below the threshold, but it should be understood that these jitters are not scientifically meaningful. The long term trends are.

References

1. E. Lundstad, Y. Brugnara, D. Pappert, J. Kopp, E. Samakinwa, A. Hürzeler, A. Andersson, B. Chimani, R. Cornes, G. Demarée, J. Filipiak, L. Gates, G.L. Ives, J.M. Jones, S. Jourdain, A. Kiss, S.E. Nicholson, R. Przybylak, P. Jones, D. Rousseau, B. Tinz, F.S. Rodrigo, S. Grab, F. Domínguez-Castro, V. Slonosky, J. Cooper, M. Brunet, and S. Brönnimann, "The global historical climate database HCLIM", Scientific Data, vol. 10, 2023. http://dx.doi.org/10.1038/s41597-022-01919-w