Constraining projections with observations

Can past observations be used to help constrain future temperature projections? This question is particularly relevant given the last decade which has shown relatively less warming than expected.

One of the main approaches used to address this problem is called ‘Detection & Attribution’ (or D&A). Simply, this methodology involves simulating the climate of the past 150 years with various combinations of radiative forcings – usually greenhouse gases only, volcanic & solar only etc – and then recombining these simulations with appropriate scaling factors to best match the observations. The same scaling factors can then be applied to future simulations to make a constrained forecast.

Previously, in the year 2000, Allen et al. applied this technique to the simulations available at the time and made a forecast for the years after 2000 which turned out to be rather accurate.

Although several assumptions are made in this approach – including that the patterns of response to radiative forcings are correct, and that the internal variability in the models is of the right magnitude – the latest CMIP5 simulations offer a chance to revisit these forecasts with a wider range of models.

The result is an open access paper in ERL by Stott et al. The key figure (below) shows how the constrained projection (black lines) is slightly offset and lower than the raw projections from the CMIP5 simulations, suggesting that the very highest sensitivity models are less likely to be consistent with past observations.

The raw CMIP5 range (grey) and constrained projections (black) of decadal mean global temperatures using RCP 4.5, both showing 5-95% ranges.

Stott, P., Good, P., Jones, G., Gillett, N., & Hawkins, E. (2013). The upper end of climate model temperature projections is inconsistent with past warming Environmental Research Letters, 8 (1) DOI: 10.1088/1748-9326/8/1/014024

About Ed Hawkins

Climate scientist in the National Centre for Atmospheric Science (NCAS) at the University of Reading. IPCC AR5 Contributing Author. Can be found on twitter too: @ed_hawkins

17 thoughts on “Constraining projections with observations

  1. Still too much co2, too little Sun. Try thinking about the hysteresis of the ocean’s enormous heat capacity and consider the 70 years of above average solar activity from 1935-2003.

    Co2 change lags behind temperature change at all timescales.

  2. Projections are always difficult, and making them makes scientists hostages to fortune.
    I’m not a scientist, so please be gentle with me as I pose a few questions.
    First, the period between about 1945 and 1975 shows a slight cooling. I remember scientists in the late sixties and early seventies hypothesizing about a new Ice Age. However, as we started cleaning up the particulates from heavy post-war industrialisation – mainly coal-burning, I believe – in the late seventies, the effects of putting CO2 into the atmosphere, hitherto masked, became clear. My point is that we might be doing something similar right now without realising it (for example, is the coal-fired advanced of China and India having the same sort of effect), or have we already factored that possibility into the projections? Second, are the projections merely atmospheric warming, or do they include oceanic warming. Warm seas appear to have exacerbated the storm that recently hit the North-Eastern seaboard of the USA, giving huge amounts of snow; and may also be responsible for the highly unusual plume of seabed methane rising in the Barendts Sea between Norway and Svalbard in January. This warming may then feed back into more atmospheric warming, as well as the somewhat scary prospect of the methane from that and permafrost thaw providing a possible tipping point.
    I’m sure scientists must have covered these points, so could we have an explanation as to how they affect the projections?

    1. Hi Peter,

      The last decade is interesting – and there may be several reasons, including increased aerosol emissions from China.

      The projections include ocean warming, and sub-surface methane and permafrost melt are included to some degree. However, these projections assume certain scenarios for methane (and all other important factors), so any differences from what actually happens will mean a difference in the outcome.


  3. The HADCRUT4 data on that plot goes to 2010. If you extended it to 2012 (such as in the graph on your pervious post) it looks to me like the observational data may have already dropped out of the bottom of even your observationally constrained range. Any thoughts?

    1. Hi HR,

      Couple of thoughts – firstly I think the observations would still be inside the dashed lines, just, but I didn’t make the plot so I can’t add them to check.

      Also, I forgot to add to the figure caption that the two future plumes are for uncertainty in decadal mean temperatures, as discussed in the linked paper.


  4. Hi Ed,

    I have read your repsonse on the previous thread (thanks) and look forward to future additions.

    Back to this thread:

    I am not surprised that the trajectory is lowered but I am a bit surprised that the uncertainty remains well constrained, but there is not enough detail in the Letter to make a better judgement. It has to do with the location of “TRUTH” in the combination of ensemble plus observations. I get puzzled as to why the OBS should necessarily be closer to TRUTH than any of the models, particularly once they have been rescaled for signal (are they also rescaled for variance?).

    I should like to know more about the contents of the D&A references (19-22) but none of them are availabe to me. I expect they contain an important information on how the result can be so well constrained.

    Many Thanks


    1. Hi Alex,
      If you want some extra reading from the papers you mention then please email me!

      Not quite sure I follow your first paragraph above, but I would like to see some ‘perfect sibling’ tests with these types of simulations, i.e. using data from a model as past and future “observations” and see whether another different model can predict the future “observations” using these D&A techniques. No-one has yet done this, but would help judge whether the constraints are broadly realistic or not. There is some variance rescaling I think, and the uncertainty bounds do include a contribution from internal variability.


  5. Hi Ed
    There seems to be a wide range of emission scenarios in CMIP5; are these results constrained to those in which the scenario is similar to the observed rise in forcing, at least to date?

    1. Hi John,
      Yes, there are 4 CMIP5 scenarios – RCP4.5 is the popular one, but the paper includes RCP8.5 as well, which has a faster increasing forcing. However, the observational forcing is only used until 2005, when the RCPs start, so technically the forcings after 2005 are not the observed ones, but are taken from the RCPs.

  6. Ed,

    A couple of questions:

    1. If causality cannot be strictly defined for past variance, how can constraints be established going forward? For example: increasing rate change of warming over 3 warming periods from 1880 to present, how will this affect future warming rates?

    2. Doesn’t this only work in steady state conditions? In other words, with no major change in the system? (i.e. the complete loss of industrial activity leading to zero anthropogenic Sulfate emissions, what would the warming curve look like under this scenario?)

    1. Hi John,
      Not quite sure I follow point (1)… but this approach works only in a system NOT in a steady state as it scales the response to the various forcing agents and assumes the emissions follow the future RCP trajectories, so doesn’t include the possibility of zero sulphate emissions as you describe.

  7. Dear Ed,
    Nice posting. Granted, the mean temperatures seem to be increasing a lot less that the models would predict, but what about the medians? In other words, are the distributions of temperatures over each year similar? My guess would be that the distributions are getting more and more skewed, but the data are seldom showed.

    Another issue: it seems that the Sun will miss its appointment with the next sunspot maximum; rather, it appears that at the appointed date it will actually be at minimum. How does this (if at all) affect the model predictions?

Leave a Reply

Your email address will not be published. Required fields are marked *