Whether you're looking at COVID-19 projections or baseball projections from ZiPS, remember: all models are wrong, some are useful. We're not giving you the right answer, but how wrong an answer is likely to be under certain parameters.
That's not to say these aren't useful, because they are. But models aren't time machines and are at the mercy of the data. Much of the error is unexplainable and the more open the system and the more unknowns, the larger the errors.
We can model a team's run scored very accuratedly without knowing any run data. But it's also an *extremely* closed system with a very small number of specific, well-described, events.
And even for runs, there's an error in there that there's *no* answer out there for. It's not a matter of finding the right answer, because that answer isn't out there. No amount of overparameterization will find it.
Now, with something like a novel coronavirus, you *expect* very large errors. It's not a problem for models - imagine trying to develop WAR for players in a game that has never been played before, has uncertain rules, and just has a few vague similarities to baseball.
Modeling a virus like this one is a bit like generating performance statistics for a game of Calvinball. It's not shocking that models have been this wrong but that models have been this *right*. All we can do (well, they, I'm not an epidemiologist), is light a candle.
And as the saying goes, it's better to light a candle than curse the darkness. We don't have a sun.
Developing and running a popular baseball model for >15 years hasn't made me more confident in predicting the future, it's made me *less* so. I am keenly aware now of how wrong I should be and how right I could never possibly be.
You can follow @DSzymborski.
Tip: mention @twtextapp on a Twitter thread with the keyword “unroll” to get a link to it.

Latest Threads Unrolled: