Further review on cdc_baseline_forecaster

[brookslogan]

cdc_baseline_forcecaster can successfully reproduce historical Flusight-baseline forecasts from last season; see this script. However, some extra features in cdc_baseline_forecaster that aren't needed to produce Flusight-baseline forecasts may need some extra review.

[also review claims re. covid if they still exist]

Rough points to revisit; some probably don't actually apply

• check if runs when doing geo pooling

• .$ in epi_slide… prefer .x$? or fine?

• locale-independent Saturday check?

• styler

• .data$ / .env$ to calm checks? or fn Nat used in epiprocess?

• why !!outcome?

• why predictor on keys? could this be follow "id variable" example in `?add<sub>role</sub>`?

• (1group_by(across(……))1 vs. 1group_by(pick(…….))1? deprecation planned in future https://www.tidyverse.org/blog/2023/02/dplyr-1-1-0-pick-reframe-arrange/. but for compatibility better to keep around??)

• missing a step_epinaomit for the training window? (but what about test data selection?)

• side

  • `get<sub>testdata</sub> `arg validation & fixup a little weird looking (might be able to combine some, `allow<sub>null</sub>=TRUE` inside non-NULL branch, -Inf thing weird, class != class)
  • ~ min(.x$lag %||% Inf) — why min? also, mapping across all steps…. need to remember this if doing archive-based recipes b/c we may not want this
  • check… max lags & max horizon also might not be archive-backcast-compatible unless already doing transform to epi`<sub>df</sub>
  • what??`?

  if (is.null(n_recent)) n_recent <- min_required + 1 # one extra for filling
  if (n_recent <= min_required) n_recent <- min_required + n_recent

• appears to be flatline + iterated (as if independent) symmetrized 1-week
  differences, separately for each geo (w/ no time window, no transformation)

• no need for `if (args<sub>list</sub>$nonneg) f <- layer<sub>threshold</sub>(f, ".pred")`? or does it need to be before `cdc<sub>flatlinequantiles</sub>`? or not?

• what type of warning are we trying to suppress with suppressWarnings? be more selective?

• incomplete propagate test

• major

  • `data<sub>frequenc</sub>`y not considered in layer?
  • check on hhs… we don't want filling through forecast_date
  • nsims much smaller?
  • do we really want warning + something different rather than error when
    `by<sub>key</sub>` cols aren't available? also, the warnings don't trigger?? probably
    just casualty of suppressWarnings
  • no clue about the reasoning here

`nafill<sub>buffer</sub>`: At predict time, recent values of the training data are
used to create a forecast. However, these can be 'NA' due to,
e.g., data latency issues. By default, any missing values
will get filled with less recent data. Setting this value to
'NULL' will result in 1 extra recent row (beyond those
required for lag creation) to be used. Note that we require
at least 'min(lags)' rows of recent data per '`geo<sub>value</sub>`' to
create a prediction. For this reason, setting 'nafill_buffer
< min(lags)' will be treated as additional allowed recent
data rather than the total amount of recent data to examine.

• semimajor

  • assume this doesn't actually work with `time<sub>type</sub>` = week?
  • data frequency was not 1 week for covid-19 forecasts
  • need to do a `step<sub>epilag</sub>` 1? data_frequency? to get the right training window selection?
  • if there are gaps, are deltas appropriately NA?
  • with hhs latency, `forecast<sub>date</sub>` & `target<sub>date</sub>` setting is awkward
  • for non-flatline, had issue with contrasts on 1 geo and with residuals not matching size on mult geos. maybe missing `step<sub>epinaomit</sub>`? but adding `step<sub>epinaomit</sub>` gives `Warning: Values from `q` are not uniquely identified; output will contain list-cols.`

• awkward…:

  if (max_ahead > 1L) {
  for (iter in 2:max_ahead) {

• filter to Saturdays & no time_type update…

• `epiprocess::guess<sub>period</sub>` useful?

• `state<sub>census</sub>$fips` should be chr

• maybe avoid `sample` due to length-1-numeric case? unlikely to encounter but bad…