Consider implementing suggested model (Krymova et al.)

Migrating from e-mail thread. Consider implementing the model described in "Trend estimation and short-term forecasting of COVID-19 cases and deaths worldwide" by Krymova et al. ([preprint](https://arxiv.org/abs/2106.10203)) as a baseline-ish model. From a quick skim, it looks like it breaks the past into 6wk windows, LOESS-STL-smooths each, recombines, and then, to forecast, linearly extrapolates on either the original or log scale, depending on the original-scale trend slope (trying to _avoid_ overshooting or undershooting). It appears to be a purely AR-flavor model without extra covariates, done for each geo individually. Somehow they obtain distributional forecasts, but I didn't immediately see where this was described.