For the last few years (since at least 2016), I’ve shared predictions for the next annual global mean surface air temperature (GMSAT) anomaly based on the long term trend and the state of ENSO at the start of the year. Generally speaking, this has been quite skillful compared to persistence or just the long term trend alone – the eventual anomaly was consistently within the predicted bounds. Until 2023.
As described in my original post on 538, I take a loess smooth for the GISTEMP long term trend (using roughly 20 year smoothing) and add a term based on the linear regression of the beginning of the year Multivariate ENSO Index (MEI2) (similar to Nino34) to the detrended anomalies (not including some big volcanic years). This makes sense since, historically, the interannual variations in GMSAT were largest in the first half of the year and dominated by the phase of ENSO (El Niño or La Niña). This pattern was important for recent record or near-record years like 2016 or 2020 which started with El Niño, as well as below-trend years like 2017, 2021 etc. that started with La Niña. The development of the ENSO phase in the latter part of the year (which peaks around December/January) generally has less of an impact because of the lag of ~3 months or so of its affect on global temperatures.
There are two main sources of uncertainty in this method, the variation of temperature not related to the prior ENSO, and the uncertainty in the DJF ENSO index from the Dec predictions. Thus the true prediction (made around Jan 1), is slightly more uncertain than the retrodiction (which knows the actual DJF ENSO value). As mentioned above, this technique has historically been quite skillful:
| Year | Prediction (ºC above 1880-1899, 95% CI) | Outcome |
| 2016 | 1.16 ± 0.13 | 1.24 |
| 2017 | 1.06 ± 0.13 | 1.15 |
| 2018 | 1.05 ± 0.14 | 1.07 |
| 2019 | 1.18 ± 0.15 | 1.20 |
| 2020 | 1.21 ± 0.14 | 1.24 |
| 2021 | 1.16 ± 0.14 | 1.07 |
| 2022 | 1.23 ± 0.14 | 1.11 |
| 2023 | 1.22 ± 0.14 | 1.37 ± 0.03* |
| 2024 | 1.38 ± 0.14 | TBD |
The RMS forecast error (not including 2023), is 0.07ºC, compared to 0.10ºC for persistence or smoothed trends. This year however was noticeably warmer than the prediction or retrodiction based only on DJF ENSO at the beginning of the year (which you will recall was a slight La Niña), falling well above the 95% CI.
This could be due to a real anomaly in the interannual variability that was outside the 95% expectation, a mis-specification in the statistical model (e.g. we could have included an autumnal ENSO state as an additional predictor, or taken predicted forcings (solar, aerosols, volcanoes) into account), or something extra that we just haven’t seen before.
But how are we going to find out? What happens in 2024 will be important. Does it go back to being predictable based on ENSO (in which case 2024 is expected to just be a little warmer than 2023), or does the positive anomaly persist? We will also be seeing more comprehensive estimates of the impact of the Hunga-Tonga eruption, and also of the impacts of the decreases in marine shipping emissions. It might be that the initial estimates of their impacts were underestimated. We will also see more in depth explorations of the spring to fall anomalies in the North Atlantic/North Pacific which contributed strongly to the temperature changes, but aren’t obviously related to El Niño.
If nothing else, 2023 reminds us that the climate system still has surprises for us, and that this would be a very bad time to our eyes off the ball.
Thank you Gavin for this article and the reminder of the modulation of global warming by ENSO. But may I suggest that the global mean surface air temperature (GSAT) is certainly an appropriate metric to underpin our ability to anticipate the future and speak a common language in international climate change negotiations, but that it is not necessarily the most appropriate metric for informing people of the tipping point at which they will probably have no choice but to leave their home or abandon their land, notably because of too frequent floodings or droughts.
GSAT gives us the double illusion of a gradual and predictable evolution of the world. Beyond the climate tipping points that may appear on a regional scale, we could give the floor to the human sciences to question the political tipping points which seem to appear here and there, partly fed by the feeling of powerlessness and despair over the proliferation of increasingly extreme events in the four corners of the globe.
Concerning climate science, we could also emphasize the fact that most climate models struggle to reproduce some pbserved trends which could mean an even greater increase in the frequency and intensity of droughts (1,2,3) . At the moment more and more scientists are apparently having fun predicting the year, the month or the day when GSAT has already crossed or will definitively cross the threshold of +1.5°C compared to the pre-industrial level, At the same time, democracy is cracking in many countries, including in the Global North and the favorite of the Republican primaries in the USA is once again banking on the discontent of the most vulnerable, As scientists, it may be urgent to enrich the Paris Agreement by focusing on where models may be wrong or inaccurate and what the implications are for adaptation and mitigation policies.
(1) Allan, R.P., H.Douville (2024) An even drier future for the arid lands. P. Natl. Acad. Sci. USA commentary, 121(2), e2320840121, https://doi.org/10.1073/pnas.2320840121
(2) Douville H. and K. Willett (2023) A drier than expected future, supported by near-surface relative humidity observations. Sc. Adv., 9, eade6253, https://doi.org/10.1126/sciadv.ade6253
(3) Simpson, I.R. , McKinnon, K.A., Kennedya, D., Lawrence, D.M., Lehner, F., Seager, R. (2023) Observed humidity trends in dry regions contradict climate models. P. Natl. Acad. Sci. USA, in press
Have you considered including the state of the Atlantic Multidecadal Oscillation (AMO) as an additional predictor in your model? If not, why not?
[Response: I have not seen any compelling evidence that independent oscillations in the Atlantic (other than that driven by forcings) has any imprint on the global mean SAT anomalies. – gavin]
Contrary to popular opinion, the AMO time-series has just as much higher frequency content as multi-decadal, as the multidecadal variation only emerges after severe low-pass signal filtering is applied. In actuality, the raw AMO shows almost as much inter-annual (and faster) variation as the NAO index. So whatever is happening with the AMO, it is just not simply a multidecadal effect on top of a faster cycling, but more likely this faster cycling is superposing to create a multidecadal variation.
A good geophysics comparison is with the Earth’s LOD, which is essentially a measure of the Earth’s rotation rate variation. The LOD is clearly primarily composed of relatively fast tidal variations ranging from semi-diurnal to annual in period. Yet there are also long-term variations in LOD, best revealed after low-pass filtering, that curiously also seem to match the multidecadal AMO.
It’s a secret known only to tidal oceanographers that the main long-period (i.e. anything greater in period than the well-known diurnal and semi-diurnal cycles) tidal factors Mf and Mm work together to create a strong nonlinear tidal factor Mt that shows a remarkable interaction with the annual cycle. It just so happens that the Mt tidal cycle forms an almost perfect beat with the annual cycle but not quite commensurate with a year. This means that any integrated response of the annual cycle interacting with the Mt cycle will show a massive multidecadal variation. So while the Mf and Mm and other tidal factors will contribute to intra-and-inter-annual variations to an inertial response, the Mt cycle will appear superimposed on this as a strong multidecadal response as its relatively weaker forcing is compensated by long time intervals where it is constructively interfering with the annual cycle (and then long intervals where it is reversing that).
Importantly, this will happen with the LOD (a solid-body inertial response) as well as with the AMO (a fluid inertial response). The complication is that the fluid response is mathematically more complex and will likely impact the solid response, which explains why disentangling the slow responses is not the easiest challenge.
Is this information not getting out because of all the schisms in earth sciences — meteorology vs climate science vs atmospheric science vs ocean science vs geophysics? Who is tieing all this stuff together? NASA should be leading the way on this, but the machine learning folks will likely get there sooner.
Perhaps the above seems like a long shot in terms of capturing an erratic response such as AMO or ENSO or PDO, but consider that the lauded neural networks that are in vogue in climate research circles is even more of a bleeding-edge approach. I’m leaning towards a rationalization that the nonlinear NN training is likely capturing some of the aspects of the nonlinear fluid dynamics response formulation that I derived and published 5 years ago. That gives me a 5 year head-start, with the nonlinear response automatically “reverse engineered” — which is the most difficult part of applying a neural network — in other words: “the ANN works, but why?”
Thanks, Gavin!
Presumably you meant that the [cooling] effects of the Hunga-Tonga eruption were *overestimated,” in contrast to the [presumably warming] effects of cleaner marine shipping being underestimated?
https://agupubs.onlinelibrary.wiley.com/doi/10.1029/2023GL104634
Or am I missing something here?
[Response: The impact of HT is a bit different from other volcanoes in the record because of the large amount of water vapor relative to SO2, so there are two (competing) effects: warming from strat WV, and cooling from SO2–> sulphates. There were initial signs that the WV effect would dominate, but I don’t think that is how it’s actually played out. – gavin]
Re Gavin in comment to Kevin: “There were initial signs that the WV effect would dominate, but I don’t think that is how it’s actually played out”
– is there a temporal aspect to that – i.e. was the lack of WV domination from the beginning, or only after some time? The latter would suggest longer residence time of the SO2, the former probably that the plume didn’t get high enough so the residence time of WV was so short that WV didn’t stay long enough to dominate SO2…
[Response: The WV and the sulphate are in slightly different layers now, and that will increase as the sulphate particles gravitationally settle (while the WV anomaly is just advected with the flow). So the sulphate will fall out before the water vapor is circulated out. But there are nuances here that people are still looking at… – gavin]
Re: Gavin’s: “ So the sulphate will fall out before the water vapor is circulated out”
so how do people reconcile this with your earlier observation::
the initial signs that the WV effect would dominate, but I don’t think that is how it’s actually played out ?
Wouldn’t the SO2 falling out and WV staying up- result in the … “WV domination” ?
[Response: Eventually – but there are nuances related to the heights and temperatures.- gavin]
Ah yes, those nuances. Exactly. People still looking at those nuances. We’ll get back to you. :-)
Gavin: ” Eventually [yes] – but there are nuances related to the heights and temperatures.”
Ned Kelly: Jan 11: “ Ah yes, those nuances. Exactly. People still looking at those nuances.”
N. Kelly dismissing nuanced thinking. Playing to your wheelhouse, Ned ? ;-)
Ned Kelly, I’m very sorry the real world is too complicated for you. Let us know how you fare when you find a simpler world to live in.
Gavin, thanks for clarifying.
+++++
It appears to me that nucleation from nSO2 molecules would remove several to many times n more water molecules by the time it rained out. It also appears that the processes are devilishly complex and not fully understood.
“The actual sulfuric acid vapor concentration necessary for a considerable nucleation rate in the atmosphere is estimated to be in the order of 10-4 ppm by volume or 3 × 109 molecules per cm3.”
https://www.sciencedirect.com/science/article/abs/pii/0021850279900326
“Stratospheric aerosols are usually sulfuric acid solutions in the 75 wt % H2SO 4 range [.lunge, 1954, 1963], while the tropospheric sulfate aerosol may occur as sulfuric acid or as neutralized solution”
https://agupubs.onlinelibrary.wiley.com/doi/pdf/10.1029/96JD03064
” The online measurement technique developed by Eisele and Tanner (1993) in the last century led to the identification of the role of SA in NPF through field measurements. However, classic nucleation theory predicts a binary H2SO4-H2O nucleation rate far below the measured ambient nucleation rate.”
https://www.sciencedirect.com/science/article/abs/pii/S100107422200122X
“Currently, the governing mechanisms for atmospheric nucleation under sulfur-rich environments are not fully understood. However, both field and laboratory measurements clearly demonstrate the critical roles of neutral sulfuric acid dimer formation during NPF.[NanoParticleFormation]”
https://www.mdpi.com/1660-4601/19/11/6848
I understand that this year it will be important for finding out if we are missing something with ENSO. But how definitive will next year’s warming be for settling debates regarding ECS and the Earth’s Energy Imbalance?
James Hansen seems to think it would be definitive, writing, “By May the 12-month running-mean global temperature relative to 1880-1920 should be +1.6-1.7°C and not fall below +1.4 ± 0.1°C during the next La Nina minimum. Thus, given the planetary energy imbalance, it will be clear that the 1.5°C ceiling has been passed for all practical purposes.” He also argues that the progression of next year’s warming will decide whether or not his “Global Warming in the Pipeline” paper is accurate. Do you agree?
It is hard for me (and I assume other layman following this debate) to parse how important 2024 will end up being for deciding if Hansen is write to posit a much faster rate of warming. It intuitively seems like it will be but I do not really understand some of the fundamentals of the debate e.g. are radiative forcing numbers given in the recent paper, “Robust acceleration of Earth system heating observed over the past six decades” [https://www.nature.com/articles/s41598-023-49353-1] more consistent with the mild acceleration of warming rates predicted by the CMIP6 models (adjusted down to account for the “hot model problem”) or are they more consistent with Hansen’s paper?
In summary, I can clearly see why 2024 is an important year to watch for “surprises.” I am unclear on the implications.
HMD: – “I am unclear on the implications.”
James Hansen tweeted on Jan 5:
https://twitter.com/DrJamesEHansen/status/1742934284217008623
See also my comments at:
https://www.realclimate.org/index.php/archives/2023/11/unforced-variations-nov-2023/#comment-815896
Geoff, are you, like Hansen et al, feeling isolated and ignored?
Thanks for this explanation of your global temperature predictions – I’m very curious to see how the analysis of the 2023 global temperature anomaly unfolds. One clarifying question: is the 2023 prediction value listed in the table for (1.22 +/- 0.14) the same as the 2023 prediction using ENSO plotted in blue on the graph? The graph prediction looks quite a bit lower than the YTD value (brown), but the upper limit of the table prediction C.I. bounds is technically still within the window of 1.37 +/- 0.03.
It seems like a decent chunk (probably not half) of the model error for 2023 might be “explainable” by the temporary nature of whatever caused the 2021 and 2022 predictions to be too high. If 2021 and 2022 had held to form, the loess extrapolation to 2023 would have made for a warmer forecast.
Along the same lines, the cumulative error over the past 3 years is positive, despite the exceptionally warm and unpredicted 2023. The evidence is pretty thin at this point that there’s been a systematic change in climate system behavior causing additional warming, just as a year ago the evidence was thin that there had been a systematic change reducing the warming rate.
Dr. G.S. says: “But how are we going to find out? What happens in 2024 will be important.”
Golf: Nearest the pin tees off.
Darts: Nearest the bullseye throws first.
Cards: High card denotes the dealer
Monopoly: Highest dice throw starts first.
This new featured story may not be as self-indulgent and self-serving as it appears. ChatGPT could be wrong.
Please reflect on whether you are using your limited time online to maximum efficiency.
These actions lack practicality, fail to contribute positively to anyone’s well-being, and may often be regarded as unproductive or even detrimental. They may consume valuable resources like time, energy, or attention without providing any meaningful benefit or outcome for individuals or society as a whole.
Actions with no practical benefit to anyone typically refer to behaviors or activities that don’t serve a purpose, or provide value to individuals, communities, or society in general.
Thankfully, this possibly isn’t the case here. There must be some usefulness to be gained.
I believe that you have described how I feel after reading your comment.
I was wondering why the SAT data for the NOAA station in Liberia, Costa Rica was adjusted downward at least 5 degrees for pre-1998 data and not after 1998, and why all the data for the year 1998 was removed from that station set. Was it found that there was something wrong with the thermometer at that station?
Thanks Gavin. Deniers are trying to blame the extreme 2023 temperature on el Niño, but my analysis (like yours) indicates that the net effect this year from ENSO has been negative not positive, so the extremity of this year’s anomaly is made yet more extreme.
I’ll disagree with John N-G, because the data, when corrected for my estimate of the impact of ENSO and volcanic eruptions and solar variations, show that when it comes to recent acceleration of global warming, the evidence is not at all thin:
https://tamino.wordpress.com/2024/01/05/global-warming-picks-up-speed/
Tamino –
I don’t think we’re in disagreement. Your post provides evidence for an acceleration in the rate of warming over the past twenty or so years, while I was questioning the strength of evidence that the jump in 2023 represents a sudden acceleration due to events in the past 1-3 years that should be expected to persist.
Indeed, that is a different question. I’m looking at that now … I’ll keep y’all posted.
I’m not a mathematician so I can ‘illegitimately’ use curve-fitting to project into the future. I got tired of seeing straight lines being fitted to geodata , that were patently obviously straying from straight lines, global SLR, global temp,CO2, CH4 etc. No error bars with this but I can check on the class of acceleration in the existing data and hopefully see trends a bit into the future.
Still the projection is increasing month on month since June 2023.
First projected monthly occurance of 1.5 degC above pre-industrial temperatures comes out as April 2026 (2027 was the ‘official’ projection for this earlier in 2023).
First projected monthly occurance of 2 degC above pre-industrial temps comes out as in 2037.
Using data outputed on 14 Dec 2023
GLOBAL Land-Ocean Temperature Index
GHCN-v4 data
https://data.giss.nasa.gov/gistemp/tabledata_v4/GLB.Ts+dSST.txt
Using offline javascript curvefitter at statpages.info
and curve types linear (no acceleration), exponential (increasing acceleration), quadratic (constant acceleration), indicial (falling acceleration)
best fit,by R^2, being indicial again like most of the geodata plots, processing the whole 251 point dataset from 2003.0
y= 0.556158 + 0.002464*x^1.681316
( if ^ gets mangled in transit then exponentiation symbol)
where x is year minus 2000, for post-Pinatubo 2003.0 start, avoiding
the 10 years of recovery of much geodata after the Mt. Pinatubo eruption, 6 figures retained for comparison by anyone else repeating/checking this exercise.
for data to Nov 2023 with public output 14 Dec 2023
For 2050 , 2.326863 deg C minus 1.0536 for year start 2023.5
of that curve , so +1.273 deg C above present
For 2100 , 6.235169 deg C minus 1.0536 , so +5.182 deg C
for data to Oct 2023 with public output 16 Nov 2023
For 2050 , 2.184512 deg C minus 1.0394074 for year start 2023.5
of that curve , so +1.145 deg C above present
For 2100 , 5.487198 deg C minus 1.0394 , so +4.448 deg C
Earlier results removed for brevity
First occurance of 1.5 deg C above pre-industrial
Determining the change of reference offset, from late 20C to pre-industrial. The NOAA table is structured as anomaly rather than absolute.
Taking an accepted value (www.climate.gov) of pre-industrial 1.06 degC rise to 2022.5 , then from the NASA table, mid 2022 taking average of the calendar year was 89.75 centi-Celcius cC, or 0.8975 deg C, giving a reference offset of 1.06- 0.8975 = 0.162 deg C.
The target temp in that listing would be 100*(1.5- 0.162) = 134 cC.
For the last 5 years in that table, the positive outlier wrt the average of its year, averaged over the 5 years gives a +outlier value of 0.180 or 18 cC.
So what year does my curve give 1.34 -0.180 = 1.16 deg C,
and that comes out as in April 2026.
For 2 deg C ,what year gives 1.66 deg C from that curve and that is in 2037.
Your math is interesting, but don’t forget to check (by t test or partial F test) whether the additional terms are meaningful or not. Also, try to limit the number of significant digits you show.
Repeating this exercise for 12 Jan 2024 public output, all getting rather surreal. And not just because of 6 figure resolution retained for anyone else repeating this execise. Roll on a return to La Nina
to bring it back down again, for this 2100 projection, a previous minimum of 2.8 degC above the then present.
First month’s occurance of 1.5 degC above pre-industrial temps comes out as in Dec 2025, so possible for the next La Nina to push this event further into the future.
First projected monthly occurance of 2 degC above pre-industrial temps comes out as in 2036.
Best fit,by R^2, being indicial again, processing
the whole 252 point dataset from 2003.0
y= 0.561973 + 0.001831*x^1.777910
for data to Dec 2023 with public output 12 Jan 2024
For 2050 , 2.481986 deg C minus 1.082664 for year start 2024.0
of that curve , so +1.399 deg C above present.
For 2100 , 7.146254 deg C minus 1.082664 , so +6.064 deg C
Projection to 2100 increasing by near enough 1 deg C per month of successive data releases this year it would seem. Since June 2023 these virtual monthly increments have been +0.16degC, +0.38, +0.49, +0.53, +0.75, +0.91 .
The same exercise but for SST data
Data downloaded 13 Jan 2024 from
https://www.metoffice.gov.uk/hadobs/hadsst4/data/download.html
selecting
HadSST.4.0.1.0_monthly_GLOBE.csv
SST anomaly wrt the 1961 to 1990 average.
Using 252 datapoints from 2003.0 to Dec 2023 to avoid the 10 year recovery period of geodata post Mt Pinatubo eruption.
Using offline javascript curvefitter at statpages.info
and curve types linear (no acceleration), exponential (increasing acceleration), quadratic (constant acceleration), indicial (falling acceleration)
For data to Dec 2023
Best fit quadratic again and 6 figures retained for anyone else repeating/checking this processing
y=0.419916 -0.006279*x +0.000993*x*x
where y is Hadley SST anomaly in deg C and x is year minus 2000
Projected to 2050, +2.588 deg C
Dec o/p projected to 2100, +9.722 deg C
For the month on month difference for 2100 projection, from Jun/Jul 2023, +0.67 degC, +0.86, +0.88, +1.60, +0.34, +0.20, so having peaked.
For data to Nov 2023
Projected to 2050, +2.550 deg C
Nov o/p projected to 2100, +9.518 deg C
y=0.417237 -0.005692*x +0.000967*x*x
ps for the El Nino refs below.
I find it difficult to believe the massive lack of sea-ice for Antartica last year, suddenly switching 3 times the area of France from reflecteing energy back out to space, to the southern oceans absorbing it, had no effect, even if recouped now
nick cook: – “First month’s occurance of 1.5 degC above pre-industrial temps comes out as in Dec 2025, so possible for the next La Nina to push this event further into the future.”
What do you mean by “First month’s occurance of 1.5 degC”?
The global monthly average surface temperature exceedance of the +1.5 °C threshold already occurred last year, per ERA5 data, from July through December 2023.
https://twitter.com/EliotJacobson/status/1735658781986693145
Or per Berkeley Earth, the global monthly average surface temperature exceedance of the +1.5 °C threshold occurred in March, and July through December 2023. And in some earlier years in January through March.
https://berkeleyearth.org/wp-content/uploads/2024/01/SeasonalWrap-2023.png
Per Copernicus, the global daily average surface temperature exceedance of the +2.0 °C threshold occurred for the first time in the instrumental record on 17-18 Nov 2023.
https://climate.copernicus.eu/sites/default/files/custom-uploads/Global%20Climate%20Highlights%202023/fig3_GCH2023_PR_daily_global_temperature_increase_above_preindustrial_2023.png
The Earth System is already nudging the global 12-month average surface temperature threshold of +1.5 °C (relative to the 1850-1900 baseline).
https://twitter.com/LeonSimons8/status/1745897129636143470
Evidence/data I see suggests the global multi-year average surface temperature threshold of +1.5 °C is likely to be exceeded sometime in the 2020s.
https://twitter.com/EliotJacobson/status/1733629769915457578
On 12 Jan 2024, James Hansen, Makiko Sato and Reto Ruedy published their latest communication titled Global Warming Acceleration: Causes and Consequences. It included these statements (bold text my emphasis):
https://mailchi.mp/caa/global-warming-acceleration-causes-and-consequences
I think it would be foolish to bet that Dr Hansen & colleagues are significantly wrong on this issue.
Hi Gavin,
Greetings from Australia! El Niño is of course well underway here but so far it has been unusual with the large amount of rain we have received even with the temps being higher.
This did remind me though that there was a delay of nearly 1 & 1/2 months between the WMO and our own BOM declaring the official onset of El Niño, so I wondered as well if there might be something different about the pattern this year that would also be contributing to the discrepancy with your prediction.
With regard to El Nino, the extraordinary difference between the two most frequently used indices is striking: While the Nino34 index, which is relatively local and limited to sea surface temperatures (SSTs), already indicates a fairly strong El Niño, the MEI index, which includes a much larger region and also atmospheric variables (see explanation), has only just started to rise above the El Niño threshold. This is in contrast to the previous strong events in 1982/83, 1997/98 and 2015/16, when the MEI index showed an abrupt and strong change to the El Niño state already in early summer. And the difference between MEI index and Nino34 values is much larger (about 1 instead of <0.5). Could this indicate an unusual state of atmosphere and/or ocean over the greater central/northern Pacific for a strong El Niño event?