Autism and Forecasting
Note: If you haven’t heard of forecasting, I would recommend at least skimming the Superforecaster Wikipedia page to get a brief overview of what I’m talking about before diving in.
Being an excellent forecaster is correlated with certain traits (and certain demographic variables). The main paper that digs into this topic seems to just talk about stuff that’s fairly obvious to people who have some experience with forecasting or have already read Superforecasting. One famous example is that forecasters do better on the Cognitive Reflection Test, which asks questions like ‘A bat and a ball cost $1.10 in total. The bat costs $1.00 more than the ball. How much does the ball cost?’, with the correct answer being 5 cents rather than 10 cents, which is the answer that most people give. Unsurprisingly, political knowledge and active open-mindedness (essentially, the willingness to listen to evidence that contradicts your views) also predict forecasting ability, as do traits like deliberation time and frequency of updating beliefs.
I also have the impression, although I can’t substantiate it with survey evidence, that excellent forecasters are disproportionately autistic. I’ve done a very brief review of the literature, and I don’t think there’s any research showing that forecasters are more likely to be (high-functioning) autistics. I also emailed Phil Tetlock just to double check if I had missed anything, and he replied that he isn’t aware of any work on the correlation between autism and forecasting ability. I think it’s fairly safe to say that there isn’t much rigorous evidence for the correlation that I believe exists. But having spoken to quite a few excellent forecasters, I’ve definitely come away with the impression that many of them have traits that resemble those of people who are autistic, and others have explicitly mentioned that they have been diagnosed with autism.
If my impression is right, why might this be the case? To some extent, it might not exactly be that there’s anything that links forecasting specifically to being autistic, but rather that forecasting is just the type of hobby that attracts autistic people. A fair number of forecasters that I’ve met are competent programmers, and programmers are also disproportionately autistic. Forecasters are almost certainly disproportionately male, and men are more likely to be autistic than women. So, we would sort of expect forecasters to be more autistic than the wider population just as a function of forecasters being more male, and because being a competent programmer lends itself quite well to being good at forecasting.
But I think there might also be reasons to think that forecasters being disproportionately autistic isn’t only a result of demographic variables, it might be that autistic people have traits that actually make them better forecasters. Even though there isn’t (as far as I’m aware) any research showing a direct correlation between forecasting ability and autism, we can see that some traits that are correlated with forecasting ability are also correlated with being autistic.
The dual-process account of reasoning, which was made famous by Daniel Kahneman’s book Thinking Fast and Slow, is the idea that humans have two systems for thinking about the world. System 1 makes use of shortcuts and is fairly automatic, whereas System 2 is much more intentional (and probably accurate), but requires more resources. The Dual-process theory of Autism proposes that people with autism rely more on System 2 than on System 1, and deliberate more when trying to answer a question than do neurotypical people. Lewton et al. (2018) carried out two studies looking at the relationship between being autstic and deliberative processing, and found that autistic people do better at the Cognitive Reflection Test discussed earlier. As already mentioned, doing well at the CRT is positively correlated with forecasting ability, so if both findings are legit, it makes sense that autistic people might be better at forecasting.
The paper also found that autistic people seem to be better at ‘Syllogistic reasoning’. What does that mean? In the study, people were presented with eight logical arguments, and had to figure out whether the arguments were valid or invalid. Four of the arguments were consistent with reality, and four of the arguments were not consistent with reality. For example, the argument ‘All birds have feathers. Robins are birds. Robins have feathers’ is both logically valid and consistent with reality (what they call ‘valid-believable’), whereas the argument ‘All mammals walk. Whales are mammals. Whales walk’ is logically valid but not consistent with reality (they call it ‘valid-unbelievable’). So, we have four different types of argument: valid-believable, valid-unbelievable, invalid-believable, and invalid-unbelievable. A ‘belief bias’ occurs when people score worse on ‘incongruent problems’, the ones which are either valid-unbelievable or invalid-believable, than they do on congruent problems. In other words, someone with low belief bias is more likely to notice that although whales cannot walk, an argument that concludes that whales can walk may not be logically invalid. The result of the study, as you might have guessed, is that autistic people show much lower levels of belief bias than do neurotypical people.
There are other papers that lend credence to the assertion that autistic people are likely to make better forecasters. Morsanyi et al. (2010) show that autistic adolescents are less susceptible to the conjunction fallacy, a common flaw in human probabilistic reasoning. Testing for the conjunction fallacy works like this - you give people a description of a woman called Linda, who is described as ‘a philosophy major, concerned with discrimination and social justice, and a participant in antinuclear demonstrations’, and then you ask people if it is more likely that ‘Linda is a bank teller’, or that ‘Linda is a bank teller who is active in the feminist movement’. If someone takes the view that the second is more likely, they have committed the conjunction fallacy, as if it is the case that Linda is a bank teller and a feminist, it must be the case that Linda is a bank teller, so it can’t be more likely that she is both a bank teller and a feminist than being just a bank teller. If autistic people are better at this sort of probabilistic reasoning, we should also probably assume that they’ll be better at the sort of probabilistic reasoning that is so important in forecasting.
I feel pretty confident that my impression that skilled forecasters are disproportionately autistic is correct. We would expect forecasters to be disproportionately autistic just from knowing that they’re mostly men and that programmers are overrepresented among forecasters, but I’m also inclined to believe that high-functioning autistic people may be less susceptible to the types of biases that screw up most peoples’ forecasts. It would be pretty cool if Good Judgment Inc or Metaculus did a survey of their top forecasters that asked people whether they have been diagnosed with autism, and I would be happy to arrange a bet about how overrepresented autistic people are likely to be if anyone is interested.