Why models can't predict climate accurately

 By Christopher Monckton of Brenchley

Dr Gavin Cawley, a computer modeler the University of East Anglia, who posts as “dikranmarsupial”, is uncomfortable with my regular feature articles here at WUWT demonstrating the growing discrepancy between the rapid global warming predicted by the models and the far less exciting changes that actually happen in the real world.

He brings forward the following indictments, which I shall summarize and answer as I go:

1. The RSS satellite global temperature trend since 1996 is cherry-picked to show no statistically-discernible warming [+0.04 K]. One could also have picked some other period [say, 1979-1994: +0.05 K]. The trend on the full RSS dataset since 1979 is a lot higher if one takes the entire dataset [+0.44 K]. He says: “Cherry picking the interval to maximise the strength of the evidence in favour of your argument is bad statistics.”

The question I ask when compiling the monthly graph is this: “What is the earliest month from which the least-squares linear-regression temperature trend to the present does not exceed zero?” The answer, therefore, is not cherry-picked but calculated. It is currently September 1996 – a period of 17 years 6 months. Dr Pachauri, the IPCC’s climate-science chairman, admitted the 17-year Pause in Melbourne in February 2013 (though he has more recently got with the Party Line and has become a Pause Denier).

2. “In the case of the ‘Pause’, the statistical test is straightforward. You just need to show that the observed trend is statistically inconsistent with a continuation of the trend in the preceding decades.”

No, I don’t. The significance of the long Pauses from 1979-1994 and again from 1996-date is that they tend to depress the long-run trend, which, on the entire dataset from 1979-date, is equivalent to a little over 1.2 K/century. In 1990 the IPCC predicted warming at 3 K/century. That was two and a half times the real-world rate observed since 1979. The IPCC has itself explicitly accepted the statistical implications of the Pause by cutting its mid-range near-term warming projection from 2.3 to 1.7 K/century between the pre-final and final drafts of AR5.

3. “No skeptic has made a GCM that can explain the observed climate using only natural forcings.”

One does not need anything as complex as a general-circulation model to explain observed temperature change. Dr Cawley may like to experiment with the time-integral of total solar irradiance across all relevant timescales. He will get a surprise. Besides, observed temperature change since 1950, when we might have begun to influence the warming trend, is well within natural variability. No explanation beyond natural variability is needed.

4. The evidence for an inconsistency between models and data is stronger than that for the existence of a pause, but neither is yet statistically significant.

Dr Hansen used to say one would need five years without warming to falsify his model. Five years without warming came and went. He said one would really need ten years. Ten years without warming came and went. The NOAA, in its State of the Climate report for 2008, said one would need 15 years. Fifteen years came and went. Ben Santer said, “Make that 17 years.” Seventeen years came and went. Now we’re told that even though the Pause has pushed the trend below the 95% significance threshold for very nearly all the models’ near-term projections, it is “not statistically significant”. Sorry – not buying.

5. If the models underestimate the magnitude of the ‘weather’ (e.g. by not predicting the Pause), the significance of the difference between the model mean and the observations is falsely inflated.

In Mark Twain’s words, “Climate is what you expect. Weather is what you get.” Strictly speaking one needs 60 years’ data to cancel the naturally-occurring influence of the cycles of the Pacific Decadal Oscillation. Let us take East Anglia’s own dataset: HadCRUT4. In the 60 years March 1953-February 2014 the warming trend was 0.7 K, equivalent to just 1.1 K/century. CO2 has been rising at the business-as-usual rate.

The IPCC’s mid-range business-as-usual projection, on its RCP 8.5 scenario, is for warming at 3.7 K/century from 2000-2100. The Pause means we won’t get 3.7 K warming this century unless the warming rate is 4.3 K/century from now to 2100. That is almost four times the observed trend of the past 60 years. One might well expect some growth in the so-far lacklustre warming rate as CO2 emissions continue to increase. But one needs a fanciful imagination (or a GCM) to pretend that we’re likely to see a near-quadrupling of the past 60 years’ warming rate over the next 88 years.

6. It is better to understand the science than to reject the models, which are “the best method we currently have for reasoning about the effects of our (in)actions on future climate”.

No one is “rejecting” the models. However, they have accorded a substantially greater weighting to our warming influence than seems at all justifiable on the evidence to date. And Dr Cawley’s argument at this point is a common variant of the logical fallacy of arguing from ignorance. The correct question is not whether the models are the best method we have but whether, given their inherent limitations, they are – or can ever be – an adequate method of making predictions (and, so far, extravagantly excessive ones at that) on the basis of which the West is squandering $1 billion a day to no useful effect.

The answer to that question is No. Our knowledge of key processes – notably the behavior of clouds and aerosols – remains entirely insufficient. For example, a naturally-recurring (and unpredicted) reduction in cloud cover in just 18 years from 1983-2001 caused 2.9 Watts per square meter of radiative forcing. That natural forcing exceeded by more than a quarter the entire 2.3 W m^–2 anthropogenic forcing in the 262 years from 1750-2011 as published in the IPCC’s Fifth Assessment report. Yet the models cannot correctly represent cloud forcings.

Then there are temperature feedbacks, which the models use to multiply the direct warming from greenhouse gases by 3. By this artifice, they contrive a problem out of a non-problem: for without strongly net-positive feedbacks the direct warming even from a quadrupling of today’s CO2 concentration would be a harmless 2.3 Cº.

But no feedback’s value can be directly measured, or theoretically inferred, or distinguished from that of any other feedback, or even distinguished from the forcing that triggered it. Yet the models pretend otherwise. They assume, for instance, that because the Clausius-Clapeyron relation establishes that the atmosphere can carry near-exponentially more water vapor as it warms it must do so. Yet some records, such as the ISCCP measurements, show water vapor declining. The models are also underestimating the cooling effect of evaporation threefold. And they are unable to account sufficiently for the heteroskedasticity evident even in the noise that overlies the signal.

But the key reason why the models will never be able to make policy-relevant predictions of future global temperature trends is that, mathematically speaking, the climate behaves as a chaotic object. A chaotic object has the following characteristics:

1. It is not random but deterministic. Every change in the climate happens for a reason.

2. It is aperiodic. Appearances of periodicity will occur in various elements of the climate, but closer inspection reveals that often the periods are not of equal length (Fig. 1).

3. It exhibits self-similarity at different scales. One can see this scalar self-similarity in the global temperature record (Fig. 1).

4. It is extremely sensitive to the most minuscule of perturbations in its initial conditions. This is the “butterfly effect”: a butterfly flaps its wings in the Amazon and rain falls on London (again).

5. Its evolution is inherently unpredictable, even by the most sophisticated of models, unless perfect knowledge of the initial conditions is available. With the climate, it’s not available.

Figure 1. Quasi-periodicity at 100,000,000-year, 100,000-year, 1000-year, and 100-year timescales, all showing cycles of lengths and magnitudes that vary unpredictably. Click each image to enlarge it.

Not every variable in a chaotic object will behave chaotically: nor will the object as a whole behave chaotically under all conditions. I had great difficulty explaining this to the vice-chancellor of East Anglia and his head of research when I visited them a couple of years ago. When I mentioned the aperiodicity that is a characteristic of a chaotic object, the head of research sneered that it was possible to predict reliably that summer would be warmer than winter. So it is: but that fact does not render the climate object predictable.

By the same token, it would not be right to pray in aid the manifest chaoticity with which the climate object behaves as a pretext for denying that we can expect or predict that any warming will occur if we add greenhouse gases to the atmosphere. Some warming is to be expected. However, it is by now self-evident that trying to determine how much warming we can expect on the basis of outputs from general-circulation models is futile. They have gotten it too wrong for too long, and at unacceptable cost.

The simplest way to determine climate sensitivity is to run the experiment. We have been doing that since 1950. The answer, so far, is a warming trend so far below what the models have predicted that the probability of major warming diminishes by the month. The real world exists, and we who live in it will not indefinitely throw money at modelers to model what the models have failed to model: for models cannot predict future warming trends to anything like a sufficient resolution or accuracy to justify shutting down the West.