Climate Modeling Revisited

Jan 23, 2016



As a core part of my professional career, I have done extensive time series modeling and predictive analysis. Tens of millions of dollars are at stake in my getting it right.

I have understood for a long time the limitations of predictive modeling. First, no single model fits every time series dataset under analysis. Second, any model, no matter how good the fit (coefficient of determination), has little to no predictive skill. Any prediction made must necessarily be short term and based on an intuitive understanding of the data itself. Thus, my predictions are not based solely on model output. The model output helps me to visualize the data, which helps me to intuitively make predictions up to several months out based on trajectories and an underlying assumption that all things will remain constant.

It is because of my experience with time series models that I have been critical of the Global Circulation Models (GCM) used in climatology to make predictions 100 years out. I marvel at the notion that anyone takes them seriously. But, most of the public does not understand models and believe they can actually represent reality. They don’t and they can’t, at least not presently.




I came across an excellent dialog between Christopher Monckton and Dr. Robert Brown of Duke University that I think puts climate and models into a proper perspective. I emboldened the key points. And now for your reading enjoyment and enlightenment:

Monckton: Good questions to ask you are do you understand that volcanic eruptions cause cooling? And do you think that this cooling should be considered when trying to determine the effects of CO2?

Dr. Brown: They are excellent questions indeed. Now add the other ten confounding elements of the climate. Do you think that ENSO causes warming and cooling respectively? Do you think that the effect of ENSO should be considered when trying to determine the effects of CO_2 AND volcanoes? The evidence for this is precisely the same as, only more dramatic and of more permanent action, than the evidence for volcanic effects. Do you think that the phase of the PDO has a causal effect on the climate that can augment or reduce any multivariate warming or cooling trend produced by the other factors, including CO_2? The evidence that it does is actually very strong, stronger than the combined evidence that CO_2 plays an important role, to the extent that one can extract a CO_2 “signal” from what is now three other natural factors that appear to have a significant impact on the time evolution of the climate. What about the NAO? What about the state of the global thermohaline circulation and complex feedbacks at the parts of the world where haline density overcomes thermal stratification and the surface current sinks to return at great depth or the parts of the world where the current at depth is displaced by high density haline sinking fluid to rise, carrying an image of system state laid down some centuries earlier to the surface? What about solar state, both the direct variability of the solar output power (which is small) and the solar magnetic interaction with the Earth that very definitely has observable effects on radiation screening and things like ozone production in the upper atmosphere (which is largely unknown in its effects on the climate, but is at least partially correlated with major climate eras of the past in the complex multivariate system)? What about the eccentricity of the Earth’s orbit, which produces a 90 watts/m^2 variation of the total solar insolation at the top of the atmosphere over the course of a year, a variation that dwarfs all of the rest of the variability in forcing put together)?

What about clouds? Most of this is highly, highly, nonlinear. All of it is coupled. The effect (if any) of variable solar state could be dependent on the state of the entire Earth climate system and its past state as the time evolution of the climate requires either the completely detailed solution of every major heat source, sink and capacity from the mantel of the Earth on up to the TOA or one has to solve a non-Markovian problem where the climate today depends in part on what the climate was ten years ago or a hundred years ago when (for example) the water that is welling to the surface near Antarctica now was actually last on the surface where its state was directly coupled to the climate of that time.

The complexity of the problem is partially revealed in the Perturbed Parameter Ensemble runs of the GCMs. Tiny parameter changes relative to a given initial condition don’t produce bundle of solutions tightly bound to a nice, deterministic trajectory. It produces a diverging bundle of solutions. If one makes even major changes — completely rebalances the effects of CO_2 compared to other stuff — one simply gets a differently diverging bundle, one that very likely overlaps the bundle originally produced. What that means — technically — is that the inverse problem is not solvable. One quite literally cannot look at the climate and infer the effect of CO_2 from the temperature series, or predict the temperature series even from a perfect knowledge of the physics of CO_2. One cannot even do a good job of producing probabilities that any given model’s assignment of a “total climate sensitivity” to additional CO_2 are correct, partly because the overlap and lack of an inverse allow inference to be used in precisely the wrong direction, as is the rule and not the exception in climate science.

The right direction is this. Nature is probably “right” — that is, what happens in nature is likely to be the most probably outcome of the physics, not the least probable outcome of the physics. Any other assumption is madness and an open invitation to confirmation bias, cherrypicking, storytelling, and all of the manifold abuses of science attendant upon a claim that a model is more likely to be correct than the nature the model is modelling. Models — as is the rule throughout all physical science! — must indeed be tested against nature (and not the other way around) in order to be validated as plausibly being correct models and sufficiently accurate to be of predictive use. When an untested model fails to agree with nature, we do not assume that nature is off on a comparatively improbable track, we assume that the model has failed unless and until it produces good agreement with nature.

Finally, as anyone who does modelling professionally well knows, one cannot validate any model with its training data, with a reference set used to tune the model parameters. Most complex models are effectively overcomplete bases and can easily fit almost any behavior over a finite interval while being completely wrong outside of that interval. That’s the fundamental problem with all of heuristic curve fitting of the temperature record to sine functions, linear trends, correlations across some finite segment with some proposed external causal agency (one at a time or all together). One can get an absolute perfect curve fit to a small segment of the data (as HenryP insists on doing) with some set of basis functions, but there is quite literally no mathematical reason to expect that the fit will extrapolate outside of the training set being fit. It might. It might not. It might for a while and then suddenly decide to change. This is absurdly true for chaotic trajectories, characterized by the property of never being able to be extrapolated forward with simple curves for arbitrary time intervals. (I could say much more about Taylor series, polynomial or non-polynomial representations, uniform convergence on intervals, and so on, but either you’ve taken real math and I don’t have to or else I’d have to give you a whole course in functional analysis with trivial examples of the substantial risk of building extrapolatory models without a sound physical foundation.)

So yes, please, think about volcanic aerosols, human aerosols, volcanic and human particulates, the interaction of the above with patterns of humidity, cloud formation, rainfall, vertical heat transport in the form of latent heat, and the substantial variation of effective albedo brought about both by the direct effect of the aerosols themselves and their effect on cloud nucleation in semisaturated air. Think about how the decadal oscillations, the global atmospheric oscillations and variations in trade winds and the jet stream vary the pattern of delivery of humid air to concentrations of aerosols (which are often not particularly well mixed because they are being produced by sources localized in space and time). Think about how the particular month of the year might matter since the Earth might be getting 90 watts/m^2 more TOA TSI at one time of the year compared to another, so a volcano that goes off in the northern hemisphere in the winter might have a completely different effect than one that goes off in the southern hemisphere in the summer, and both might have completely different climate impact compared to a tropical volcano at any time of year. Consider how that effect might be further modulated by what the sea surface and thermohaline circulation are doing, by modulation of stratospheric water content or ozone, by solar magnetic effects. Then tell me that this is settled science, that we know what the net impact of increased CO_2 in this non-Markovian whirl of natural nonlinear multivariate dynamics is.

I think not. I don’t think we even have a good idea.


People are welcome to believe in climate models if they like. Because they believe, it doesn’t make them the least bit representative of reality today or in the future.

The GCM models diverge dramatically from themselves and Mother Nature as depicted in the comparison of the model predictions to actual empirical temperature measurements shown in the image below.

Click on the above image for a higher resolution view


To coin an old adage, there are two rules to remember about Mother Nature and Models:

  1. Mother Nature is always right.
  2. In the event the models say Mother Nature is wrong, refer back to rule #1.






blog comments powered by "logo-disqus">Disqus************************************************************ -->