Very recent country-days are less probable to have full covariate sets. One way to model this explicitly would be a temporal recency variable, which took the value observation date - Sys.Date().
Relatedly, next model fit should possible use a covariate set at least as old as the most recent target data, to account for the same pattern. In concert, these could help stabilize very recent predictions.
sondreus
commented
2 years ago
Very recent country-days are less probable to have full covariate sets. One way to model this explicitly would be a temporal recency variable, which took the value observation date - Sys.Date().
Relatedly, next model fit should possible use a covariate set at least as old as the most recent target data, to account for the same pattern. In concert, these could help stabilize very recent predictions.