RealClimate logo

Unforced Variations vs Forced Responses?

Guest commentary by Karsten Haustein, U. Oxford, and Peter Jacobs (George Mason University).

One of the perennial issues in climate research is how big a role internal climate variability plays on decadal to longer timescales. A large role would increase the uncertainty on the attribution of recent trends to human causes, while a small role would tighten that attribution. There have been a number of attempts to quantify this over the years, and we have just published a new study (Haustein et al, 2019) in the Journal of Climate addressing this question.

Using a simplified climate model, we find that we can reproduce temperature observations since 1850 and proxy-data since 1500 with high accuracy. Our results suggest that multidecadal ocean oscillations are only a minor contributing factor in the global mean surface temperature evolution (GMST) over that time. The basic results were covered in excellent articles in CarbonBrief and Science Magazine, but this post will try and go a little deeper into what we found.

Until recently, the hypothesis that there are significant natural (unforced) ocean cycles with an approximate periodicity of 60-70 years had been widely accepted. The so-called Atlantic Multidecadal Variability index (AMV, sometimes called the AMO instead), but also the Pacific Decadal Variability index (PDV) have been touted as major factors in observed multidecadal GMST fluctuations (for instance, here). Due to the strong co-variability between AMV and GMST, both, the Early 20th Century Warming (1915-1945) and the Mid-Century Cooling (1950-1980) have been attributed to low-frequency AMV variability, associated to a varying degree with changes in the Atlantic Meridional Overturning Circulation (AMOC). In particular, the uncertainty in quantifying the human-induced warming fraction in the early 20th Century was still substantial.

Fig. 1: Matches of modeled temperature to the observations since 1850. Upper graph shows the global response model with ENSO (bold green) compared to HadOST (bold black). Lower graph is the same as above but with lowess smoothed observational data. The response model results (green thin lines) represent the parameter uncertainty for an associated TCR of 1.6K. The dashed thin line is the upper and lower (reasonable) bound for the effective aerosol forcing for 2017 (-0.5 and -1.0 W/m2), in contrast to the best estimate of -0.75 W/m2 used in the response model. The grey area indicates the 5-95th percentile of the total uncertainty. The two graphs are offset by 0.9°C without a particular baseline. Response model and observations are aligned for the 1901-2000 period.

In contrast to those earlier studies, we were able to reproduce effectively all the observed multidecadal temperature evolution, including the Early Warming and the Mid-Century cooling, using known external forcing factors (solar activity, volcanic eruptions, greenhouse gases, pollution aerosol particles). Adding an El Niño signal, we virtually explain the entire observed record (Figure 1). Further, we were able to reproduce the temperature evolution separately over land and ocean, and between Northern and Southern Hemispheres (NH/SH). We found equally high fractions of explained variability associated with anthropogenic and natural radiative forcing changes in each case. Attributing 90% of the Early Warming to external forcings (50% of which is due to natural forcing from volcanoes and solar) is – in our view – a key leap forward. To date, no more than 50% had been attributed to external forcing (Hegerl et al. 2018). While there is less controversy about the drivers of the Mid-Century cooling, our response model results strongly support the idea that the trend was caused by increased levels of sulphate aerosols which temporarily offset greenhouse gas-induced warming.

What does this mean?

Some commentators have used the uncertainty in the attribution for the Early 20th Century warming as an excuse to not accept the far stronger evidence for the human causes of more recent trends (notably, Judith Curry). This was never very convincing, but is even further diminished given a viable attribution for the Early Warming now exists. Despite a number of studies that have already provided evidence – based on a solid physical underpinning – for a large external contribution to observed multidecadal ocean variability, most prominently the AMV (e.g. Mann et al., 2014; Clement et al., 2015, Stolpe et al. 2017), ideas such as the stadium wave (Wyatt and Curry, 2014) continue to be proposed. The problem is that most studies that argue for unforced low-frequency ocean oscillations do not accommodate time-varying external drivers such as anthropogenic aerosols. Our findings highlight that this non-linearity is a crucial feature of the historic forcing evolution. Any claim that these forcings were/are small has to be accompanied by solid evidence disproving the observed multidecadal variations in incoming radiation (e.g. Wild 2009). On the contrary, our findings confirm that the fraction of human-induced warming since the pre-industrial era is bascially all of it.


Fig 2. The residual observed variability in the NH. Model minus 30 year smooth observations (red). A revised AMV index is shown in black. Note that the rhs y-axis labels for the AMV SSTs is different.

We conclude that the AMV time series (based on the widely accepted definition) almost certainly does not represent a simple internal mode of variability. Indeed, we think that the AMV definition is flawed and not a suitable method to extract whatever internal ocean signal there might be. Instead we recommend the use of an alternative index which we think will be closer to the internal signal, called the North Atlantic Variability Index (NAVI). It is essentially the AMV relative to the NH temperature (Figure 2). The resulting timeseries of the new NAVI index is a good representation of the AMOC decline, arguably the true internal component (although also forcing-related) in the North Atlantic. This implies that while the AMOC is an important player (see for instance, Stefan’s RC post), it is not driving alleged low-frequency North Atlantic ocean oscillations. The AMV should therefore not be used as predictor in attribution studies given that the multidecadal temperature swings are unlikely internally generated. Though we note that the projection of AMV on GMST is small in any case.

What did our results rely on?

Fig 3. NH response model results from 1500-2017 (bold red). NTrend proxy in orange and a subset of individual NH proxy reconstruction (thin brown lines). HadCRUT4 and Berkeley in grey and black for the 1850-2017 period. The response model is baselined to the initialisation data which corresponds to 1500 A.D.

There are three novelties that led to our conclusions: (1) We differentiate between forcing factors such as volcanoes and pollution aerosols with regard to their transient climate response (TCR). For example, anthropogenic aerosols are primarily emitted over NH continents, i.e. they have a faster TCR which we explicitly account for in our analysis. (2) We use an updated aerosol emission dataset (CEDS, Hoesly et al., 2017, also used in CMIP6), resulting in a substantially different temporal evolution of historic aerosol emissions compared to the older dataset (Lamarque et al. 2010). The effective aerosol forcing is based on the most recent estimate by [9]. (3) The final change is related to the observational data. The HadISST1/2 (Kennedy et al. in prep) ocean temperature dataset (SST) has never been used in conjunction with land data. We have combined HadISST2 with Cowtan/Way over land (using air temperature over sea ice) and filled the missing years after 2010 with OSTIA SSTs (due to it being preliminary only). In addition, it has been known for quite some time now that there is a bias in virtually all  SST dataset during the 2nd world war (Cowtan et al., 2018, and see also Kevin’s SkS post). We correct for that bias over ocean (1942-1945), which, in conjunction with warmer HadISST2 SSTs before the 1930s, significantly reduced previous discrepancies related to the Early Warming. Lastly, the fact that the model is initialised in 1500 A.D. ensures that the slow response to strong volcanic eruptions is sensibly accounted for (Figure 3), as it has shown to be important on centennial timescales (e.g. Gleckler et al. 2006).

What about overfitting?

In order to address this issue, we would like to point out that not a single parameter depends on regression. TCR and ECS span a wide range of accepted values and all we did is to estimate TCR based on the best fit of the final response model result with observations. We concede that the fast response time and the effective aerosol forcing are difficult to pin down given there is a wide range of published estimates available. However, it is worth mentioning that the results are not very sensitive to variations in both parameters (see thin lines in Figure 1). Instead, the overall uncertainty is dominated by the TCR and GHG forcing uncertainty. The story is more complex when it comes to the NH/SH and land/ocean-only results as we need to account for the different warming-ratios. Guided by climate model and observational data, we introduce a novel method that objectively estimates the required TCR factors.

Conclusions: It was us.

The findings presented in our paper highlight that we are now able to explain almost all the warming patterns since 1850, including the Early Warming period. We achieve this by separating different forcing factors, by including an updated aerosol dataset and by removing notable SST biases. We have avoided overfitting by virtue of a strict non-regression policy. We ask the different research communities to take these findings as food for thought, particularly with regard to the Early Warming. We most definitely believe that it is time to rethink the role of the AMV and recommend using our newly introduced NAVI definition instead. This will also help to understand contemporary AMOC changes and its relation to climate change better, and perhaps provide guidance as to which climate models best approximate internal ocean variability on longer timescales.


  1. K. Haustein, F.E.L. Otto, V. Venema, P. Jacobs, K. Cowtan, Z. Hausfather, R.G. Way, B. White, A. Subramanian, and A.P. Schurer, "A Limited Role for Unforced Internal Variability in Twentieth-Century Warming", Journal of Climate, vol. 32, pp. 4893-4917, 2019.
  2. G.C. Hegerl, S. Brönnimann, A. Schurer, and T. Cowan, "The early 20th century warming: Anomalies, causes, and consequences", Wiley Interdisciplinary Reviews: Climate Change, vol. 9, pp. e522, 2018.
  3. M.E. Mann, B.A. Steinman, and S.K. Miller, "On forced temperature changes, internal variability, and the AMO", Geophysical Research Letters, vol. 41, pp. 3211-3219, 2014.
  4. A. Clement, K. Bellomo, L.N. Murphy, M.A. Cane, T. Mauritsen, G. Radel, and B. Stevens, "The Atlantic Multidecadal Oscillation without a role for ocean circulation", Science, vol. 350, pp. 320-324, 2015.
  5. M.B. Stolpe, I. Medhaug, and R. Knutti, "Contribution of Atlantic and Pacific Multidecadal Variability to Twentieth-Century Temperature Changes", Journal of Climate, vol. 30, pp. 6279-6295, 2017.
  6. M.G. Wyatt, and J.A. Curry, "Role for Eurasian Arctic shelf sea ice in a secularly varying hemispheric climate signal during the 20th century", Climate Dynamics, vol. 42, pp. 2763-2782, 2013.
  7. M. Wild, "Global dimming and brightening: A review", Journal of Geophysical Research, vol. 114, 2009.
  8. R.M. Hoesly, S.J. Smith, L. Feng, Z. Klimont, G. Janssens-Maenhout, T. Pitkanen, J.J. Seibert, L. Vu, R.J. Andres, R.M. Bolt, T.C. Bond, L. Dawidowski, N. Kholod, J. Kurokawa, M. Li, L. Liu, Z. Lu, M.C.P. Moura, P.R. O'Rourke, and Q. Zhang, "Historical (1750–2014) anthropogenic emissions of reactive gases and aerosols from the Community Emissions Data System (CEDS)", Geoscientific Model Development, vol. 11, pp. 369-408, 2018.
  9. C.E. Forest, "Inferred Net Aerosol Forcing Based on Historical Climate Changes: a Review", Current Climate Change Reports, vol. 4, pp. 11-22, 2018.
  10. K. Cowtan, R. Rohde, and Z. Hausfather, "Evaluating biases in sea surface temperature records using coastal weather stations", Quarterly Journal of the Royal Meteorological Society, vol. 144, pp. 670-681, 2018.
  11. P.J. Gleckler, K. AchutaRao, J.M. Gregory, B.D. Santer, K.E. Taylor, and T.M.L. Wigley, "Krakatoa lives: The effect of volcanic eruptions on ocean heat content and thermal expansion", Geophysical Research Letters, vol. 33, 2006.

61 Responses to “Unforced Variations vs Forced Responses?”

  1. 51
    dhogaza says:


    “The important difference from coin flipping is the claim that the “hiatus fluke” has been explained.”

    Are you claiming that science is incapable of explaining why a short run of fair coin-flipping might end up all heads?

  2. 52

    jgn, #47–

    Thanks for the thoughts! (And link.)

    “It does bother others as I think we are seeing here (including possibly you?).”

    FWIW, no, I wouldn’t say I’m ‘bothered.’ What I bother about is mostly the progress of emissions mitigation and measures which I think are likely to contribute to it ASAP. But I was/am ‘interested.’

    I’ll add that in high school a friend and I once diverted ourselves with a coin-toss trial, held in my friend’s room. The environment was sufficiently chaotic that we had to add a third column for the time the coin ended up on edge! (I wasn’t too bothered then, either.) :-)

  3. 53
    Ray Ladbury says:

    Matthew Marler,
    So, shall we come up with specific labels for every insignificant fluctuation that happens? For example, shall we call the period from 2015-2017 “the acceleration”.

    The objection to the term “hiatus” is that it implies that the period had a significance that it did not and that it was the invention of denialist charlatans who knew better, but were content to dupe their large, receptive audience of non-experts.

  4. 54
    JCH says:

    The coin flip has been skewed. The theory indicates there should be more tails.

    There are 9 heads in a row. It is not noise. That would have to be explained, so people start writing papers that could explain it.

    Suddenly it comes to light that some of the heads were actually tails, and presto, 9 heads in a row never happened.

  5. 55
    Matthew R Marler says:

    Haustert et al and Kim et al are since the following: Article | OPEN | Published: 20 November 2018

    Global-scale multidecadal variability missing in state-of-the-art climate models
    S. Kravtsov, C. Grimm & S. Gu

    npj Climate and Atmospheric Sciencevolume 1, Article number: 34 (2018)

    I think it may be still an open question whether the Haustert et al and Kim et al explanations are accurate and reliable. I think they probably are, one maybe more so than the other, but the test will be how well they model the 15 + years after the final resolution of the 2015-2016 El Nino.

  6. 56
    t marvell says:

    New research that takes into account trends in supercooled clouds concludes that global warming will be much greater than is commonly believed. Over 5 degrees C for a doubling of CO2.

    I have long argued, in many posts here, that climate scientists have been understating the extent of warming.

    See also:

  7. 57
    Jai Mitchell says:

    I really appreciate your time on this forum, thanks! The magnitude of the late period warming compared with the amount of NH aerosol emission reduction is the key here.

    I do also appreciate the concept of tuning, in my profession we call it a ‘true up’ Trying to use historical data to confirm or disprove a theorized parameter.

    I would use a simple current warming and global energy imbalance model. I would assume a current 1C of warming with effective climate response operating on a 20-year lag. Then I would consider aerosol forcing and known climate feedbacks/tipping points that are on the near horizon.

    This would be the method used for a risk-assessment analysis. It produces a value of warming at current GHG levels slightly greater than 2.5C. If carbon cycle feedbacks are also included then we are already beyond 2X CO2, with every GtC being emitted today equal to 1.7GtC total atmospheric burden (including near term carbon sinks but rapid reduction of sinks turing to net source around 2050).

  8. 58
    jgnfld says:


    Edge landings have also been studied. See Murray and Teare (1993) . (I know of no open source version but I’m pretty sure you have access to a research library.)

    For US nickels of the day on a proper surface they estimated edge landings at 1 per 6000 on both theoretical and observational grounds (though I pity the poor graduate students!).

  9. 59
    Matthew R Marler says:

    53, Ray Ladbury: For example, shall we call the period from 2015-2017 “the acceleration”.

    If it continues, sure. Right now it is called an “El Nin~o”.

  10. 60
    jgnfld says:


    El Ninos, la Ninas, and neutral periods have precisely zero effect on the trend. In point of fact, any proper time series analysis partials out significant cyclic component(s) as independent variate(s) uncorrelated with trend. I forget the ref, but this can be seen directly if one regresses and graphs the 3 periods separately. The same trend emerges regardless. Someone can point to it if you are doubtful about this.

    This makes labeling insignificant subperiods in the trend with causal language implying pauses and accelerations (1) fairly silly as any time one adds a line and a cycle together you get these by definition and (2) much more speculation and magic than anything empirical as no mechanism actually affecting the trend is implied or tested.

    Retailers, for example, do no get overly excited about the rise in activity before Christmas and the drop afterwards. The actual trend with cyclic components partialed out over many Christmases is what is important to the long term viability of the business. As, for example, Toy-R-Us found out.

  11. 61
    Peter K Miles says:

    Causes of Variation and the impact of the ONI:

    Looking at the HadCrut 4.0 annual data (or GISS for that matter) from 1970 to date shows a remarkably linear relationship with time. For the HadCrut 4.0 data the correlation has an R^2 value of 87.9% and the model is given as ‘Temp Increase = -34.05 + 0.01737 year’. Calendar year as input, result relative to the mean of a 1850-1899 base line.

    More sophisticated models do not seem to do any better than the simple linear regression equation shown here.

    However, the ‘residuals'(The actual-predicted values) of the regression model are worth studying. They appear to be bi-modal rather than normally distributed around the best fit line. This suggests that there is some kind of binary or on/off influence at work. One obvious candidate is the ONI, so I decided to try and correlate the annual residual values generated by the model with the prior ONI data. Transitions from El Nino and La Nina or back seem to be relatively quick compared with the length of a El Nino or La Nina phase. I tested this with offsets of 0 to 12 months between the averaged annual ONI data, and the residual value for the calendar year. The best correlation was for a 6 month delay between the ONI and its impact on the residual. The R^2 number of 48.9% suggests that about half of the variation we see in the annual temperatures over and beyond the relentless linear increase is due to the ONI. Is this consistent with current thinking?