There are a lot of epidemiological models flying around right now. They make predictions for what will happen if we enact this or that intervention, or none at all. These models are extremely useful. But they don’t answer the basic question: what is going to happen? They only tell us what to expect conditional on some other action being taken.
The Good Judgment Project is likely the best group in the world for forecasting the outcomes of global events. Today I won’t bother to talk about how they do it or why I think they’re so great, besides recommending the book Superforecasting. The good people at Open Philanthropy, who concern themselves, among other things, with fighting global pandemics, have commissioned a public forecast of coronavirus by the GJP. Specifically, they are forecasting counts of infections and fatalities in the US and globally by March 31, 2021. These are unconditional forecasts. It’s not, “what will happen if blah blah blah”, it’s, “what’s going to happen?”
Like all good forecasters, they don’t make a single point estimate. They give a probability distribution. They say they’ll update daily. They’ve been doing it for a week now, though I only learned about it today. They haven’t kept a public record of their last week’s worth of forecasts. I intend to record their forecast every day and to convert their probability distribution to a single expected value. That throws out a lot of nuance. Pandemics follow a power law: the more people who get infected, the more more people will get infected. Unlike a normal distribution, the most likely single outcome is not the same as the average outcome. ‘Average’, ‘mean’, and ‘expected value’ all mean the same thing. I understand that for people who haven’t taken probability classes it’s counterintuitive that ‘expected value’ isn’t the same as ‘most likely outcome’. I don’t expect to solve that confusion today except to point it out for anyone who reads this. With those caveats, tracking the average by day seems like the best way to answer the simple question: did things get better or worse today?
Technical note: the GJP forecasts the probability distribution in exponentially sized bins. To convert to an expected value, I need to assign a point value to each bin. I’ll take the geometric mean of the lower and upper bound of each bin. That’s equivalent to taking the arithmetic mean in log space. For the lowest bin, with a lower bound of 0, I’ll set the lower bound to be roughly the value as of today, March 19th. For the highest bin, I’ll just set the whole bin equal to its lower bound. This cuts off even worse outcomes, but their highest bin is so unlikely the effect should be negligible.
These forecasts are the “give it to me straight” take. It’s not trying to make us feel better. It’s not trying to scare us into taking action. The distribution is broad, so if the expected value seems high that doesn’t mean we can’t take action to bring it down. And if the expected value seems low that doesn’t mean we should relax; it’s already baking in assumptions about actions we’ll take to suppress the virus.
With that, here’s today’s forecast. [Updated 3/22/2020]
My spreadsheet is here, along with the full distribution for each of the four outcomes.
How to 115 