Convert a baselinenowcast_df object to a forecast_sample object

This function converts a baselinenowcast_df object as returned by baselinenowcast() to a forecast_sample object which can be used for scoring with the scoringutils package.

Usage

# S3 method for class 'baselinenowcast_df'
as_forecast_sample(
  data,
  latest_obs,
  observed = "count",
  model = "baselinenowcast",
  ...
)

Arguments

data

A baselinenowcast_df object as returned by baselinenowcast() with output_type = "samples".

latest_obs

A data.frame containing the truth to score against, with one row per (reference_date, strata) combination. Must contain a reference_date column and a column with observed counts named according to observed (default "count"). Additional columns shared with data (such as strata columns) are used as merge keys. latest_obs should hold the reported total at each reference date evaluated at the same max_delay horizon used for the nowcast (rolling truth), not the partial total available when the nowcast was run.

observed

Character string giving the name of the column in latest_obs that holds the observed value. Defaults to "count" to match the input format of baselinenowcast.data.frame().

model

Character string used as the value of the model column on the returned forecast object. Lets scoringutils::summarise_scores() run with its default by = "model". Defaults to "baselinenowcast". Pass NULL to omit the column (e.g. when data already carries its own model column).

...

Additional arguments passed to scoringutils::as_forecast_sample().

Value

A forecast_sample object as returned by scoringutils::as_forecast_sample().

Details

The nowcast samples in data are merged with the latest available observations in latest_obs on reference_date and any other shared columns (e.g. strata columns such as age_group or location). The merged data is then passed to scoringutils::as_forecast_sample(), using pred_count as the predicted value, draw as the sample identifier, and the column named by observed as the observed value.

See also

Main nowcasting interface functions as_forecast_point.baselinenowcast_df(), assert_baselinenowcast_df(), baselinenowcast(), baselinenowcast.data.frame(), baselinenowcast.reporting_triangle(), baselinenowcast_df-class, new_baselinenowcast_df()

Examples

if (FALSE) { # interactive() && requireNamespace("scoringutils", quietly = TRUE)
library(scoringutils)

max_delay <- 25
nowcast_date <- as.Date("2026-04-01")

# Build the full reporting triangle once from the long-form data
full_tri <- as_reporting_triangle(syn_nssp_df) |>
  truncate_to_delay(max_delay = max_delay)

# As-of view for the nowcast: drop rows past `nowcast_date`, then apply the
# diagonal reporting structure so cells reported after it become NA
n_drop <- sum(as.Date(rownames(full_tri)) > nowcast_date)
rep_tri <- full_tri |>
  truncate_to_row(t = n_drop) |>
  apply_reporting_structure() |>
  tail(n = 40)

# Run a probabilistic nowcast
nowcast <- baselinenowcast(rep_tri, draws = 100)

# Truth: total reports observed within `max_delay` of each reference date,
# taken from the same `full_tri` (rolling truth). `as_forecast_sample()`
# scores only the right-truncated reference dates that were actually
# nowcast (the `nowcast` column of the nowcast); the inner-join merge then
# restricts the truth to those dates. The scoring vignette explains why we
# use a rolling rather than latest-vintage truth.
truth_df <- as.data.frame(full_tri)
latest_obs <- aggregate(count ~ reference_date, data = truth_df, FUN = sum)

# Convert and score
fs <- as_forecast_sample(nowcast, latest_obs)
fs
scores <- score(fs)
scores
summarise_scores(scores)
}