As I’ve noted previously, a group of researchers is tracking the performance of the much-cited University of Washington COVID-19 model. Their initial paper earlier this month — focusing on the model’s short-term, state-level predictions of how many deaths would be recorded — showed that the “95 percent” intervals the model provided contained the correct number of deaths only 30 percent of the time, though the errors were not consistently in one direction or the other.
They’ve now updated the paper to account for new data and changes in the model. The good news is that the 95 percent intervals overwhelmingly include the right numbers nowadays. The bad news? This apparently happened only because the folks at UW made the intervals wider. The main estimates are no more accurate than they used to be, and now they’re consistently overpredicting deaths.
Meanwhile, as Jim Geraghty notes below, we’ve passed 50,000 deaths and keep adding more than 2,000 a day nationally, while the model still predicts a total of just 68,000. In the final analysis, I don’t think the model will prove to be consistently alarmist or optimistic, but I do think it will prove to be highly imprecise.