Predictions

A year ago, I predicted that the Covid19 pandemic would kill 35 million or even more people. Although mortality statistics suggest that the current official death count (about two and a half million at the time of writing) significantly underestimates the real number of deaths,¹ it is also becoming increasingly unlikely that we’ll reach numbers anywhere near that pessimistic prediction any time soon.² The reason why my prediction was so far off is the same reason why many other predictions fail: they insufficiently take into account that human behavior changes in response to changes in circumstances.³

Typically, models used to predict the future or some aspect thereof assume that human behavior in all its facets is more or less constant and a given and thus outside the scope of the model. In this sense most models are ahistorical. Without taking human behavioral responses to changing circumstances into account, however, no model can produce a reasonably accurate prediction of the future. For example, Covid19 doesn’t spread as fast as some models (including my own) predicted, because those models did not take into account that when more people are sick, more people change their behavior (even without official policies to force them to change their behavior). And similarly, models of the effects of climate change are unlikely to produce accurate or even useful predictions if they don’t (sufficiently) take changing human behavior in response to climate change into account either. By implication, a prediction of the effects of climate change cannot just be a climatological model – it must also be a sociological, economic, political, ecological, and so forth model.⁴ There are no such broad and integrated models, however, and it is doubtful that such a model is even possible (in a quantitative sense of “model” at least).

About half the articles in this blog thus far deal in some way or other with prediction, and most of those are heavily focused on predicting the effects of climate change. I think it is fair to say that there are few topics that captivate me as much as attempts to predict the future,⁵ both on shorter and longer time scales. It shouldn’t come as a surprise, then, that I spend much time thinking about ways to integrate various predictions, and about revisiting and improving my own. The latter is certainly necessary, because I don’t believe anymore that my predictions for the more distant future in Stages of the Anthropocene are plausible, for example. Like the Covid19 example above, it takes human responses to changing circumstances insufficiently into account, but in addition to that, I think I also made mistakes in my summaries of some of the climate science involved, and there are various other factors that I neglected there. What I keep wondering about is whether and how I could do better. Perhaps I can, but there are some serious obstacles in the way, and the little bit better I might be able to do is probably still not good enough, even by my own standards. Here, I want to highlight a few of those obstacles. The extent to which they can be overcome (at least to some extent), as well as attempts to update my predictions will probably be topics for future articles at this blog. Perhaps, this will develop into a new series,⁶ but at present, there is still way too much that is too unclear to me to even guess where this investigation is heading.

resolution

The most important limitations of climate models themselves (i.e. of models that focus just on the climate and that ignore everything else) is resolution or grid size. A global climate model divides the planet (as well as time) into cells with a size of usually around 100 by 100 km. For each of these cells a number of variables is modeled in each time step. However, for detailed predictions such large cells may not be good enough, and perhaps, they are not even good enough for rough predictions. The relation between resolution and computing power needed to run the model isn’t linear, however, but exponential. For a resolution that is twice as detailed, you’d need about ten times as much computing power, which means that at high resolutions the model will even be impossible to run on a supercomputer. With current constraints on available computing power, we can’t increase the resolution much beyond the aforementioned 100 by 100 km, which means that anything that is smaller than that needs to be “parametrized”, which means that it is approximated by simplifications. The problem is that that may not be good enough. Two of the things you’d ideally be able to incorporate in your model are the landscape – mountains particularly – and clouds. Parametrization doesn’t work well for either. For mountains etcetera, you’d probably need a resolution of a least four times what is currently common. For clouds, you’d need even higher resolutions. Why – you may ask – is this such a problem?

Japan is a nice example of the relevance of mountains. (There probably are better examples, but I live in Japan, and it is the example I am most familiar with.) In winter there is a dominant westerly wind in Japan. This wind picks up moisture from the sea of Japan (between Japan and Russia/Korea). When the wind hits the mountain range that runs all along Japan’s main island, Honshū, it moves up and cools down, which reduces its capacity to hold moisture resulting in clouds, snow, and rain. On the other side of the mountains the air descends again, increasing its capacity to hold moisture resulting in disappearing clouds and dry air. Because of this, the northwest side of Honshū is nearly permanently cloudy in winter and has lots and lots of snow (and rain), while the southeast side has nearly permanent blue skies and is very dry. The distance between those areas is small, and when you travel by bullet train you’ll find a very sharp boundary between the two areas. At some point, you’ll enter a tunnel in snowy terrain below a cloudy sky, to re-emerge in a dry and sunny landscape on the other side (or the other way around, of course).

The point of the example is that mountains, and other aspects of the landscape, have very significant regional effects, and consequently, if you want your model to make predictions on the regional level (and a “region” in this sense can be bigger than most countries), then you need to take mountains etcetera into account and thus need a higher resolution. And this matters because we need somewhat reliable predictions on the regional level for at least two reasons. Firstly, the global climate is a complex system, meaning that small differences in initial conditions may lead to large differences further down the line. Or in other words, regional effects often don’t stay regional – the may have global impacts, and those global impacts are often unpredictable. By implication, without taking the landscape into account (and thus increasing the resolution) the model probably can’t even make reasonably accurate predictions on the global scale. Thus far, the predictions of the main climate models haven’t been far off, but we haven’t pushed the Earth system much yet, and there is no guarantee that it will stay like this. In the contrary, there is little doubt that some regional changes that can’t really be predicted (yet) will have significant global effects.

Secondly, and perhaps even more importantly, it makes a huge difference were exactly droughts and other natural disasters hit and how bad exactly they get. Small differences in precipitation and temperature matter, and so do small differences in location. The social, political, and economic effects of severe drought and deadly heatwaves are very different if they affect a densely populated urban area than when they affect a nearby rural and sparsely populated region. About 150 million people live in the southern half of Nigeria, for example.⁷ If that region becomes too dry and/or too hot to sustain much more than a single digit percentage of its current population, then that would have severe socioeconomic and political implications for much of Africa and even Europe. Hence, to be able to forecast numbers and flows of refugees, the impact of disasters and their various effects, and so on, and so forth, we need reasonably accurate predictions on the regional level. However, if you compare current maps of drought prediction, then you’ll see that they all make fairly similar predictions on very large scales, but when you zoom in, they completely disagree and there are very high levels of uncertainty. With the current state of climate modeling and forecasting, when it comes to the regional level, we know almost nothing.

Clouds matter for reasons similar to the first reason why regional predictions matter: complexity. As mentioned, clouds are much too small, but also much too complex, to fit in current climate models (and it is actually rather unlikely that this will change) and are, therefore, approached by means of parametrization. What isn’t clear, however, is whether that parametrization significantly affects the reliability of the models. Two years ago, a paper was published that focused on clouds in a small area of tropical ocean.⁸ Its most surprising finding was that stratocumulus clouds disappear if atmospheric CO₂ exceeds 1200ppm, resulting in a 8°C temperature rise in addition to what current models expect. This is just a single study, and it is very unclear whether and how its results can be extrapolated, but it illustrates that clouds (or the lack there off) may have effects that far exceed their parametrization. As mentioned above, the models have been reasonably accurate thus far, but if we further push the climate outside its “normal” range, there is no guarantee that it will stay like that – suddenly, minor details may turn out to have major implications.⁹

gaps and lags

A second problem with climate models is uncertainty with regards to gaps, lags, and time scales. The main problem in this respect is not the time lag between a hypothetical end of CO₂ emissions and a cessation of further heating (as a direct consequence thereof). There is some uncertainty in this respect, but research suggests that it is unlikely that this time lag is more than a decade – most likely it is much shorter.¹⁰ My main problem is of a different nature (and may be as much due to my ignorance as to limitations and uncertainty in models): often it is unclear whether supposed effects are immediate or gradual.

Take a map of expected aridity at an average global temperature increase of 2°C as an example. Does that map show expected aridity at the point in time when we reach +2°C? Or does it show what will eventually be the new situation if warming stabilizes at that level? And in the latter case, how long does it take until that new situation is reached? And what happens in between? Answers to these questions matter. When and where droughts and other changes and disasters hit matters greatly if you want to predict refugee flows, for example. But I almost never find clear answers to these questions in the climate science papers I read. Because warming is slow, giving systems some time to adapt, I usually assume that the answer to my first question is “yes”, which makes the other questions moot, but it remains unclear to me whether that is the right answer.

after ecosystem collapse

Another major uncertainty due to (my) ignorance is that I have no clue what happens after an ecosystem collapses. I was hoping to find clues by reading some (parts of) books about ecology and biogeography, but thus far that hope turned out to be in vain.

Ecosystems are (usually) complex networks of interdependencies between (usually) many species in an area. Even if one of those species is driven to extinction or emigration due to changes in circumstances, that can have ramifications throughout the network resulting in its collapse. Climate change results in rather extreme changes to circumstances, which usually affect very many species in an ecosystem, and because of that, even at 2°C of warming many ecosystems will already collapse, and for every further increase there will be more. I’m not sure whether an average increase of 3°C (which means and average of roughly 4.5°C on land, and even 6°C in cold climates) would leave any ecosystem unaffected. If we heat up the planet that much (and it is very likely that we will, unfortunately), ecosystem collapse might be a global phenomenon. It is for this reason that it is necessary to understand what happens after ecosystem collapse to be able to make predictions on slightly longer timescales (i.e. more than a few decades).

However, as mentioned, I haven’t been able to find much useful information about this. Perhaps, I shouldn’t have expected otherwise, as the circumstances of collapse are very different from those that ecologists typically study. The gradual arising of some (new) ecosystem in an area were there was none is called “succession”. In case of primary succession, the process starts from scratch; in case of secondary succession remnants of a previous ecosystem and/or its conditions are still present. Primary succession describes the slow arising of an ecosystem (or actually, a succession of ecosystems) on more or less new (-ly exposed) land created by, for example, a volcano or a retreating ice cap. In such cases, the process often has to start with soil formation (which takes a lot of time) and the first organisms entering the new habitat are fungi, lichens, algae, and so forth. Secondary succession takes place after the previous ecosystem was destroyed, but that destruction has left the soil mostly intact and didn’t destroy all roots and seeds (etc.) either. This happens, for example, after a forest fire or flood.

The kind of ecological succession that would (or might?) take place after ecological collapse due to climate change is neither primary nor secondary succession, however. In most cases there will be soil (as in secondary succession) and that soil will also contain roots and seeds of plants, but those plants can no longer thrive (or even survive) in those conditions. Furthermore, the succession doesn’t start on a more or less blank slate – not everything that lived in the area will be wiped out at once and completely. Because of that, some typical pioneer species may not be able to perform their usual pioneering role, resulting in a further deviation from “standard” succession patterns. The most important difference, however, is that in the kind of ecological succession we’re used to, seeds etcetera of plants that can survive in the new habitat are available at relatively close distances and thus spread into the new habitat quite easily. In case of ecosystem collapse due to climate change, that is not the case, however. The new circumstances in some area may not have had an equivalent prior to collapse anywhere nearby. Plants that can no longer thrive or survive in their current habitats will have to travel far greater distances to find new suitable locations than they are capable of. Birds, of course, can transport some seeds over great distances, but that only works for some seeds, and only in areas that aren’t too hot for birds to survive.

Hence, neither primary succession, nor secondary succession describes what is likely to happen after ecosystem collapse due to climate change, which makes it very difficult – for me at least – to say anything about future habitability and carrying capacity (particularly with regards to how many people an area can potentially feed). On the other hand, it isn’t even clear how much ecosystem collapse due to climate change matters in this respect. According to a recent study by Rui Yin and colleagues, the effects of intensive land use outweigh those of climate change.¹¹ The areas that we rely on most for food production are the product of intensive “gardening” and can only persist through gardening.¹² We don’t really rely on “natural” ecosystems anyway (in as far as those still exist), but on gardening the planet. Climate change will just increase the amount of gardening we need to do, and destroy the environments that we weren’t gardening intensively yet. Consequently, for future food production it won’t matter much what ecosystem could potentially arise in some area through ecological succession, but what kind of gardening we can do there. (How much gardening we can do in circumstances of widespread societal collapse due to natural disasters and refugee flows is another question.)

Nevertheless, ecosystem collapse matters and will have severe and unpredictable consequences. One consequence is that the animals living in a collapsing ecosystem will be driven out of their habitats before they migrate to another suitable habitat or go extinct, which brings them into closer contact with humans. That increase in contact of humans with various “wild” animals will increase the spread of zoonotic diseases. Covid19 is one example hereof, but the virus that causes Covid19 is a rather mild one – there are many others that are far worse (Ebola and Marburg, for example), and ecosystem collapse will expose humans to some of those viruses. Because of this, there will be pandemics in the future that make the current pandemic look like a minor variety of the usual flu season. This, of course, ads another layer of unpredictability. It is impossible to say when the next pandemic will start, how bad it will be, and/or how fast and far it will spread. All we do know with near perfect certainty is that there will be other pandemics.¹³

the human factor

The main obstacle to prediction isn’t climatological or ecological, however – it’s us. While there are models of various aspects of human behavior, none of those reach acceptable levels of accuracy and some are even farcically inaccurate (with economic models taking the crown, given that according to mainstream economic models economic crises like the 2007~2009 “Great Recession” are impossible). Human behavior is far too complex to model, and consequently, guessing how groups of people respond to climate change and the direct and indirect effects thereof can really never be more than that: guesswork. It may be theoretically informed and well-argued guesswork in some cases, but it’s still guesswork. A wrong guess or oversight can result in a prediction that is wide off its mark, however.

There is reason to believe that rising temperatures will lead to an increase in violent conflict,¹⁴, for example, but it is nearly impossible to predict where and when exactly conflicts take place and how violent they get. Take India and Pakistan, as an example. Both countries are already affected by climate change (floods, heatwaves, droughts), and this will only get worse. Both countries claim Kashmir, which is an important source of increasingly scarce (and sometimes overabundant) water for both. The conflict between the two countries already becomes violent occasionally, and it is not at all unthinkable that hostilities at some point develop into war. Both countries have nuclear weapons, and if the conflict escalates, they might resort to using those. The results would – obviously – be devastating. If they’d use between 15 and 20 average sized bombs that would kill millions of people directly, but the amount of soot that those bombs together would put in the atmosphere would reduce rainfall and shrink growing seasons globally for anywhere between 5 and 25 years. In some areas precipitation could decrease by up to 80%, and the resulting famines and other indirect effects could kill a billion people or even more.¹⁵

Now, I’m not saying that this will happen, but it is a possibility, and it is a possibility that would change everything. The problem is that there are hundreds (if not thousands) such possibilities, things that might happen and that would change everything. That is what makes predicting the future – at least, a future involving humans – fundamentally impossible.¹⁶ It could be argued, of course, that in the long run this kind of unpredictable events average out, and perhaps that’s true. However, such events may have a non-negligible influence on total CO₂ emissions. If in the above scenario Pakistan’s ally China and India’s ally the USA get involved in the conflict, it might even evolve into a world war, which could result in a nuclear winter wiping out most (but not all) of mankind and thereby pretty much terminating CO₂ emissions. I don’t think that this is a likely scenario, but it would result in far lower total emissions than a scenario without a world war (resulting in a far more habitable world for the survivors). Less extreme scenarios will have less extreme effects, of course, but the point is that unpredictable events in the short term may have (very) big effects in the long term.

navigating unpredictability

There are two common ways of dealing with this kind of unpredictability. The first is ignoring it. That’s obviously not recommended if you want to make somewhat accurate predictions. The second is to develop different predictions for different uncertain outcomes – or in other words, different scenarios. Climate change forecasting typically chooses the second of these options,¹⁷ but whether the scenarios adopted by climate scientists are plausible from a social-scientific point of view is quite debatable.¹⁸ More realistic scenarios need to be based on more realistic economics and sociology, for example, and integrate many other ecological, social, political, and so forth phenomena than the Shared Socioeconomic Pathways do. To what extent that is possible is an open question, but it would be interesting to try.

update (June 12)

The first two episodes of a new series on attempting to predict the mid- and long-term future of aspects of the global climate and their impact on humanity have now been published. The series introduction is titled Stages of the Anthropocene, Revisited.

If you found this article and/or other articles in this blog useful or valuable, please consider making a small financial contribution to support this blog, 𝐹=𝑚𝑎, and its author. You can find 𝐹=𝑚𝑎’s Patreon page here.

1. Most estimates based on comparisons of mortality with previous years that I have seen suggest that the real death toll of Covid19 might be about 50% higher.
2. In the long run, we’ll surely reach such high numbers, on the other hand, as Covid19 is not going to disappear, but my prediction wasn’t for the long run.
3. Actually, I did take that into account to some extent in that article, as I also wrote that if “at some point during the virus’s spread we … find better ways to treat patients or even a working vaccine (that can be produced quickly and abundantly)” then that “would lower the numbers”. I’m presenting my own predictions here as being (even) worse than they really were to make a point.
4. And this in turn implies that predicting the effects of climate change is not just climate science, but by necessity involves much science that climate scientists have no expertise in. This is a topic that I addressed before in Michael Mann versus the “Doomists”.
5. Few, but not zero.
6. Like the previous two series at this blog, Crisis and Inertia and No Time for Utopia.
7. And that number is expected to double at least by 2050.
8. Tapio Schneider, Colleen Kaul, & Kyle Pressel (2019). “Possible climate transitions from breakup of stratocumulus decks under greenhouse warming”, Nature Geoscience 12: 163-167.
9. That’s just an implication of complexity (in its technical sense), and everyone agrees that the climate is a complex system (in this sense.)
10. Katharine Ricke & Ken Caldeira (2014), “Maximum Warming Occurs about One Decade after a Carbon Dioxide Emission”, Environmental Research Letters 9.124002. Andrew MacDougall et al. (2020), “Is there Warming in the Pipeline? A Multi-model Analysis of the Zero Emissions Commitment from CO₂”, Biogeosciences 17: 2987-3016.
11. Rui Yin et al. (2020), “Soil functional biodiversity and biological quality under threat: Intensive land use outweighs climate change”, Soil Biology and Biochemistry 147: 107847.
12. “Gardening” here refers to agriculture and all other major changes in the natural environment by humans.
13. Epidemiologist and virologists have been warning for this for well over a decade by the way. For an accessible (and fascinating) introduction into zoonotic viruses, see David Quammen (2012), Spillover: Animal Infections and the Next Human Pandemic (London: Bodley Head).
14. Solomon Hsiang, Marshall Burke, & Edward Miguel (2013). “Quantifying the Influence of Climate on Human Conflict”, Science 341 (13 September 2013).
15. Owen Toon, Alan Robock, & Richard Turco (2008). “Environmental Consequences of Nuclear War”, Physics Today, December 2008: 37-42. Adam Liska, Tyler White, Eric Holley, & Robert Oglesby (2017). “Nuclear Weapons in a Changing Climate: Probability, Increasing Risks, and Perception”, Environment 59.4: 22-33. See also Crisis and Inertia (3) – Technological Threats and Crises.
16. If there are very many risks that each have a very low probability, then there still is a high probability that at least one of them will occur. For example, if there are a thousand risks that each are only 0.1% likely, then there is a 63.2% chance that at least one of them will happen.
17. Mainstream economists go for the first.
18. See Stages of the Anthropocene.