Make `adjust_latency` work or give nicer debug messages (on weekly data?)

[brookslogan]

I don't know what I'm doing wrong. There might be something funky going on with this data source right now (maybe just a tad out of date), but inspecting the edf, I don't see anything too egregious. Perhaps the latency adjustments are requesting shifts that don't actually exist in this weekly data set? But how is "extend_lags" getting lag 73, 80, and 87?

suppressPackageStartupMessages({
library(epidatr)
library(epiprocess)
library(epipredict)
library(dplyr)
})
cce <- covidcast_epidata()
as_of <- as.Date("2026-01-21")
nhsn_flu_tbl <- cce$signals$"nhsn:confirmed_admissions_flu_ew"$call("state", "*", "*", as_of = as_of)
nhsn_flu_edf <- nhsn_flu_tbl %>%
  select(geo_value, time_value, value) %>%
  as_epi_df(as_of = as_of)
max(nhsn_flu_tbl$issue)
#> [1] "2026-01-25"
range(nhsn_flu_edf$time_value)
#> [1] "2020-08-02" "2026-01-18"
nhsn_flu_edf %>%
  arx_forecaster("value", args_list = arx_args_list(adjust_latency = "none"))
#> ══ A basic forecaster of type ARX Forecaster ═══════════════════════════════════
#> 
#> This forecaster was fit on 2026-02-10 12:38:41.
#> 
#> Training data was an <epi_df> with:
#> • Geography: state,
#> • Time type: week,
#> • Using data up-to-date as of: 2026-01-21.
#> • With the last data available on 2026-01-18
#> 
#> ── Predictions ─────────────────────────────────────────────────────────────────
#> 
#> A total of 54 predictions are available for
#> • 54 unique geographic regions,
#> • At forecast date: 2026-01-18,
#> • For target date: 2026-01-25,
#> 
nhsn_flu_edf %>%
  arx_forecaster("value", args_list = arx_args_list(adjust_latency = "extend_ahead"))
#> Error in `check_enough_data_core()` at epipredict/R/check_enough_data.R:95:3:
#> ! The following columns don't have enough data to train: lag_0_value,
#>   lag_7_value, and lag_14_value.
nhsn_flu_edf %>%
  arx_forecaster("value", args_list = arx_args_list(adjust_latency = "extend_lags"))
#> Error in `check_enough_data_core()` at epipredict/R/check_enough_data.R:95:3:
#> ! The following columns don't have enough data to train: lag_73_value,
#>   lag_80_value, and lag_87_value.

^{Created on 2026-02-10 with reprex v2.1.1}

[brookslogan]

One attempt to work around by lying about the as_of [though it'd be better to do something similar by shifting the time_value and keeping the as_of accurate] generates error or output+warning:

suppressPackageStartupMessages({
library(epidatr)
library(epiprocess)
library(epipredict)
library(dplyr)
})
cce <- covidcast_epidata()
as_of <- as.Date("2026-01-21")
nhsn_flu_tbl <- cce$signals$"nhsn:confirmed_admissions_flu_ew"$call("state", "*", "*", as_of = as_of)
nhsn_flu_edf <- nhsn_flu_tbl %>%
  select(geo_value, time_value, value) %>%
  as_epi_df(as_of = as_of)
max(nhsn_flu_tbl$issue)
#> [1] "2026-01-25"
range(nhsn_flu_edf$time_value)
#> [1] "2020-08-02" "2026-01-18"
nhsn_flu_edf %>%
  arx_forecaster("value", args_list = arx_args_list(adjust_latency = "none"))
#> ══ A basic forecaster of type ARX Forecaster ═══════════════════════════════════
#> 
#> This forecaster was fit on 2026-02-10 12:54:30.
#> 
#> Training data was an <epi_df> with:
#> • Geography: state,
#> • Time type: week,
#> • Using data up-to-date as of: 2026-01-21.
#> • With the last data available on 2026-01-18
#> 
#> ── Predictions ─────────────────────────────────────────────────────────────────
#> 
#> A total of 54 predictions are available for
#> • 54 unique geographic regions,
#> • At forecast date: 2026-01-18,
#> • For target date: 2026-01-25,
#> 
nhsn_flu_edf %>%
  {attr(., "metadata")$as_of <- as.Date("2026-01-18"); .} %>%
  arx_forecaster("value", args_list = arx_args_list(adjust_latency = "extend_ahead"))
#> Error in `bake()` at recipes/R/recipe.R:794:5:
#> ! There is more latency at bake time than there was at prep time. You
#>   will need to fit a model with more latency to predict on this dataset.
nhsn_flu_edf %>%
  {attr(., "metadata")$as_of <- as.Date("2026-01-18"); .} %>%
  arx_forecaster("value", args_list = arx_args_list(adjust_latency = "extend_lags"))
#> Warning in bake(step, new_data = new_data): There is less latency at bake time than there was at prep time. This will still
#> fit, but will discard the most recent data.
#> ══ A basic forecaster of type ARX Forecaster ═══════════════════════════════════
#> This forecaster was fit on 2026-02-10 12:54:31.
#> 
#> Training data was an <epi_df> with:
#> • Geography: state,
#> • Time type: week,
#> • Using data up-to-date as of: 2026-01-18.
#> • With the last data available on 2026-01-18
#> 
#> ── Predictions ─────────────────────────────────────────────────────────────────
#> 
#> A total of 54 predictions are available for
#> • 54 unique geographic regions,
#> • At forecast date: 2026-01-18,
#> • For target date: 2026-01-25,
#> 

^{Created on 2026-02-10 with reprex v2.1.1}

[brookslogan]

Aside: I've also put count data into a model that shouldn't be run on count data here. But I hope that's unrelated.