UPDATE 2024: I’ve ported this over from my old Medium account to my current blog for posterity.

UPDATE 2021: I’ve scored and reviewed these predictions here, on my blog.

Taking the lead of the ever-fascinating David Manheim, I’ll be trying my hand at out-forecasting Vox in 2020. This is the first time I’m making public forecasts and I’m really excited to put my cognitive toolkit to the test.

It’s been a couple of years since I read Superforecasting, but I’m hoping that my ongoing efforts to be a good Bayesian and my curated collection of mental models will help me out in this endeavour.

Being a newbie to public forecasts of world events, I want to try something a little different to what I see others doing. I’m going to go through Vox’s list of 19 statements and make an intuitive estimate for each. The idea is to spend 20–30 seconds thinking about the nature of the question without doing any research and then just making an estimate. Once that’s done, I’ll think about each question a little more, look into the measurement criteria, and employ some adversarial techniques to refine my intuitions (priors) to more informed estimates (posteriors). This may involve various iterations, so I’m going to make my initial predictions now, but allow myself to update this post until Friday 17 January 2020 at 23:59 PM local time. From there, my estimates are locked in for the year. I’ll be using the definitions specified by Vox and only update any constraints if they do.

When Vox publish the follow-up article at the end of the year, I’ll calculate Brier scores for both my “priors” and “posteriors” and see how I did. I’m excited!

UPDATE: I’ve locked in my predictions as of 2020–01–17 6PM local time. I’ve also recorded them on Metaculus for the 17 questions that are reflected there. Let the wait begin.

1. Donald Trump will win reelection

Prior: 70%, Posterior: 55%

Interestingly, my intuitive prior of 70% tracks perfectly with the historical incumbent reelection rate (when accounting for Obama) of 69%. That drops to around 67% if we only look at reelections since 1960. Trump seems to have low approval ratings, but the US economy is fairly strong. These may balance out. Incorporating evidence from the Democratic contenders is difficult as they’re numerous and the polls are quite divided. The PredictWise distillation of US presidential election odds across betting/prediction platforms have Trump at only 55%. However, at this point, he’s effectively running against the sum of every Democrat’s favourite candidate, not the one he’ll actually run against. If a polarising candidate wins the DRC nomination, Trump might pick up votes from defecting Democrats.

Okay, but to win reelection he has to run for reelection. According to the definition, “if Trump dies or is removed from office by any means prior to the election, question resolves negative.” According to some reports, he’s in pretty decent health. That said, the impeachment question complicates things—typically, it would encourage me to update downwards on his chances, but with Republican control of the Senate, I’ve heard that it may even strengthen his 2020 run? I’m not sure what to make of that, but it’s certainly another free variable and that makes me think that tending closer to 50% might be wise.

2. The Democratic nominee will be Joe Biden

Prior: 65%, Posterior: 55%

The PredictWise distillation of US presidential election odds across betting/prediction platforms suggests that Sanders leads Biden by a few percentage points in the DRC nomination race. But they do seem to fluctuate greatly from week to week—presumably due to the news cycle—so I’m not updating too much from the “Biden is the best known” heuristic that informed my initial estimate.

3. The GOP holds the Senate

Prior: 75%, Posterior: 65%

Having done a bit more research, it seems that Senate seats track a lot more to the presidential race than I was originally aware. At 53–47 it’s also much closer than I initially thought. Definitely updating down.

But I also noticed there’s an asymmetry to the number of seats up for reelection across parties. The way I’m currently framing this, the Republicans have 2x the “risk” the Democrats do. But then a cross-reference on PredictWise puts current estimates at 71% Republican majority. So I’ll bump my estimate back up a little.

4. Trump will not get a new Supreme Court appointment

Prior: 55%, Posterior: 75%

My prior was quite uninformed here. Trump has already made 2 nominations. Bush and Obama each had 2 nominations over their double terms. So recent trends seem to suggest Trump won’t get to make another nomination. However, Ginsberg is pretty old and there has been talk of Sotomayor retiring for health reasons.

In other words, the majority of the confidence on this forecast would come from estimating the union of the probabilities of Ginsburg or Sotomayor retiring/dying. Given their ages and what I could find about their health, their all-cause mortality risk for the next 10 years are about 15% and 40%. Assuming linearity in how that maps to annual risk (which is probably a bit dodgy), that’s a union of under 7%. That said, just because they have a low risk of actually dying this year, doesn’t mean they (or someone else) won’t retire.

5. The Supreme Court will allow more abortion restrictions

Prior: 55%, Posterior: 65%

Reading into the detailed definition, it seems this is a lot more likely than my prior (which was informed by tendency to status quo). Given reports that GOP reps. are urging SCOTUS to overrule Roe v. Wade, it seems more likely that at least some additional restrictions could be passed. I’m not well-informed on this topic at all, so I’m maintaining a lower estimate.

6. The Democratic primary will be settled on Super Tuesday*

Prior: 50%, Posterior: 55%

*According to the definition, “This question resolves positively if electionbettingodds.com assigns one candidate at least 90% of winning the Democratic primary at some point within one week after Super Tuesday.”

The idea behind the forecast is that having a greater number of candidates increases the likelihood that an early lead will consolidate power behind one candidate and end the race early. I’m not sure where I can find good historical data for this one, so I’m going to mainly go on instinct and speculation, tempered with a conservative upper bound.

7. The number of people in global poverty will fall

Prior: 75%, Posterior: 70%

Definition: “Will we see fewer than 734.5M people in extreme poverty, worldwide in the year 2020, according to World Bank estimates?”

Playing with some data, it seems like the trend toward increased income (and therefore decreased poverty) is pretty steady, despite some mention that it might decline. Given my “the world’s getting better on average” model, I’d feel pretty confident with this one. But others are more conservative, so I probably need to find more data.

8. Brexit (finally) happens

Prior: 80%, Posterior: 90%

According to the definition, this is accepted if there’s either (1) a “hard” Brexit, or (2) a withdrawal agreement with a transition period. The fact that Brexit has been repeatedly deferred for years made be quite bearish in my initial estimates. But it seems like things are different this time, and a quick look at the betting markets makes me pretty confident it will go ahead sometime in 2020.

9. The US invades and attempts a regime change in Iran in 2020

Prior: 30%, Posterior: 20%

The definition of this one actually differs quite a lot to what Vox’s forecast is. As with all these forecasts, I’ll follow the rigorous definitions of Metaculus. But I hadn’t looked that up when I set my (now inverted from 70% for the negation) prior. I thought it likely that the US wouldn’t invade Iran, but to avoid the conjunction fallacy I should now be even more sure that they won’t invade and enforce a regime change. Adding the regime change conditions reduces the affirmative set to a subset of the original “US invades Iran” set.

I’m already quite bearish on a US invasion occurring, as Iran seems to be backing down since they shot down the Ukranian airliner. Moreover, if the US invades but doesn’t officially call for Khamenei’s resignation before the end of the year, the statement evaluates in the negative.

10. China will fail to curtail its internment camp programs for Uyghurs and Muslims in 2020

Prior: 70%, Posterior: 80%

This one also has a very different definition from the specification in the Vox article. I based my initial prior on the original statement. By definition, China would need to lower the inmate population by 200k or reduce the number of camps by 200. From what I could find, the latest estimates are between 1 and 3 million interred. Requiring a 6–20% reduction in the interred seems viable in a single year if it were an explicit goal. But even as a reluctant (signalling) effort, I’d expect lower. China is powerful and influential in the world in recent years, and I don’t see any major players forcing them to make changes. From what I know, their own populace is also not in a position to oppose national agendas, so pressure from within is not strong either. I doubt international opinion on social media platforms (that are banned in China) will have much influence on national interests. But perhaps they would trade this interest for another. It’s low-hanging fruit for international relations.

11. Netanyahu will not be unseated as Israeli prime minister

Prior: 70%, Posterior: 55%

Netanyahu’s party seems to be trailing the opposition in opinion polls and, with his recent indictments, it seems like public opinion could sway against him. That said, maintenance of the status quo is typically more likely as a general rule. I’m not very well-informed on this one anyway, so I’m tending down toward 50%.

12. No gene drives to fight malaria-carrying mosquitoes will be launched in any part of the world

Prior: 70%, Posterior: 80%

I actually have some background here, but whilst that makes me familiar with the concept and caveats of gene drives, it doesn’t help me predict the major causal factor here—whether humans in organisations try to implement one in the wild. I think there is wonderful potential for good here, so I’m bullish on the concept, especially in the context of malaria prevention. I probably need to counter for that preference when forecasting. Predicting the ecological effects that an IRL gene drive would have is difficult. I think approaches that render the offending species extinct (by making them infertile) are more dangerous. We’d be better served by conferring malaria resistance genes without removing species from their ecological niche. From what I understand of the problem, that’s more difficult to do, which decreases the chance that a gene drive will be implemented this year.

13. No new CRISPR-edited babies will be born

Prior: 45%, Posterior: 75%

Once again, I have at least some domain knowledge here. I’ve long maintained that editing embryos to remove disease risk would be the justifiable beginning of human genetic engineering and would eventually trigger a genetic arms race. All that was needed was one rogue group to defy the status quo and trigger the cascade. In 2019, I thought He Jiankui had done just that. But it seems that my prediction was wrong (or at least that there is some lag time).

One of my few controversial theories is that a lot of human gene editing probably occurs out of the public eye. I think that may very well involve CRISPR-edited babies in 2020. I think China publicly punishing He may have been a signalling strategy—the scapegoat who triggered the events many were waiting for. However, this forecast only applies to publicly-announced births. Because of that (and the present unpopularity of this concept in popular discourse), I don’t really expect any announcements this year. But some shady labs might be discovered, which I expect would result in something that qualifies as a public announcement. So I’m leaving some probability density on the table for that eventuality.

14. The number of drug-resistant infections will increase

Prior: 75%, Posterior: 85%

There isn’t a formal definition for this question on Metaculus, so I’m going off what Vox say in the article. To me, this seems obvious. Species tend to diversify. Drug-resistance is highly selected for (especially within hospitals). Short of some breakthrough in antibiotics that applies to a wide spectrum of bacteria, it seems like a given that there will be more kinds of drug-resistant bacteria. But what it seems like Vox are forecasting here is the number of reported cases of a patient with a drug-resistant infection. This is only slightly-correlated with the number of drug-resistant bacteria subspecies. So it’s not a given, but it’s still quite likely.

15. Facial recognition will be banned in at least three more US cities

Prior: 65%, Posterior: 65%

There appear to be two competing trends here. Firstly, the majority of Americans trust city law enforcement to use facial recognition responsibly. And the maturation of the technology in recent years makes it a highly valuable asset in the toolkit of any city department. Secondly, multiple US cities have already banned facial recognition use (by city departments).

The question is which trend will dominate in 2020. The more I look into it, the happier I am to stick with my prior here. Only two cities, Portland and Springfield, are apparently considering bans, but they might not pass them before the end of the year.

16. Beyond Meat will outperform the general stock market

Prior: 40%, Posterior: 55%

Once again, there’s no Metaculus definition for this one, so I’ll rely on the (very little) information in Vox’s original post. I don’t know what they’re defining the “general stock market” as, but I know that beating the market over time is really really hard and that informed my prior. But one year is not “over time,” and the “general stock market” might mean something other than S&P 500. But plant-based diets are increasingly in vogue and Beyond Meat might just ride that wave in 2020, so I’ll give them a slight forecast advantage over the negative case.

17. Global carbon emissions will increase

Prior: 70%, Posterior: 85%

This also seems like a given to me, but the fact that it’s a forecast question made me set a more conservative prior. In almost all recent years, the estimated emissions have been higher than the previous year. In general, emissions increase year-on-year. But, fortunately, it seems like the second derivative of emissions is negative. That is to say, the rate at which emissions are increasing is slowing down. That does cast some doubt on this forecast, but I’m gonna bump up my estimate.

18. Average world temperatures will increase relative to 2019

Prior: 60%, Posterior: 55%

This one is a little less obvious than global emissions. The weather is definitely correlated with the climate, which is correlated with global emissions. But there are two layers of chaotic systems separating them. As you can see in the figures below, the general trend in average temperature is clear, but from year to year there’s a lot of “wobble.” Given that, it makes sense to have only a slight expectation that 2020 will be hotter than 2019. It’s like a slightly-loaded coin.

19. California has a wildfire among the 10 most destructive in state history

Prior: 45%, Posterior: 58%

“The overall trend in California is troubling too. Six of California’s 10 most destructive wildfires on record have hit in just the past three years” (source).

But regression to the mean is also a thing. Just because you had 10 heads in a row, doesn’t mean you’re more likely to get heads again. But coin flips are considered truly independent events. Wildfire is definitely linked to changes in climate and human activity.

Moreover, the trend of “there’s a record fire every other year” becomes harder and harder the more extreme the recent fires are. But it’s hard to grasp without looking at the data.

The thing about this forecast is that it’s evaluated on the number of structures destroyed, which is a pretty weird way to quantify the severity of the fire. But regardless, here is a plot of the 10 worst fires in terms of destroyed structures.

2018 was outrageously extreme. But then there’s this long tail of sub-3000 fires. So breaking into this top-10 list isn’t actually that difficult for now. Moreover, the wildfires are treated as independent, so some of these are from the same year. There may be more than one wildfire outbreak in 2020, so it’s an additive probability. Even if we assume a uniform distribution of wildfires, there’s over a 10% chance that any year’s worst fire would be in the top 10 (given records dating back to 1923). But the current top-10 list is 70% composed of fires from 2015–2018. So I’m kinda unsure which way to update my prior here, but upwards seems more reasonable.

However, since 2017 and 2018 were awful years for fires, I’d presume way more resources are being thrown at the problem now. Preventative measures and early-response systems are in place, so it’s less likely that 2020 fires would cause as much damage. The fires were often caused by electrical and powerline issues, which PG&E are probably working fairly hard to address. It’s typically cheaper to prevent issues than pay for the resulting lawsuits.

Moreover, the most vulnerable structures and regions were likely destroyed in 2017 and 2018 and either no longer exist or are now built to be more fire-resistant. It’s kinda like that story of where to place armour on WWII era planes. Once again, the severity of fires is measured in terms of structures (barns, sheds, homes, etc.) destroyed. A fire bigger than any in recent history might destroy fewer structures simply because the structures are different/absent. Because of those effects, I’ll temper my update.