Score model output predictions — score_model

Scores model outputs with a single output_type against observed data.

Usage

score_model_out(
  model_out_tbl,
  oracle_output,
  metrics = NULL,
  relative_metrics = NULL,
  baseline = NULL,
  summarize = TRUE,
  by = "model_id",
  output_type_id_order = NULL
)

Arguments

model_out_tbl: Model output tibble with predictions
oracle_output: Predictions that would have been generated by an oracle model that knew the observed target data values in advance
metrics: Character vector of scoring metrics to compute. If NULL (the default), appropriate metrics are chosen automatically. See details for more.
relative_metrics: Character vector of scoring metrics for which to compute relative skill scores. The relative_metrics should be a subset of metrics and should only include proper scores (e.g., it should not contain interval coverage metrics). If NULL (the default), no relative metrics will be computed. Relative metrics are only computed if summarize = TRUE, and require that "model_id" is included in by.
baseline: String with the name of a model to use as a baseline for relative skill scores. If a baseline is given, then a scaled relative skill with respect to the baseline will be returned. By default (NULL), relative skill will not be scaled with respect to a baseline model.
summarize: Boolean indicator of whether summaries of forecast scores should be computed. Defaults to TRUE.
by: Character vector naming columns to summarize by. For example, specifying by = "model_id" (the default) will compute average scores for each model.
output_type_id_order: For ordinal variables in pmf format, this is a vector of levels for pmf forecasts, in increasing order of the levels. For all other output types, this is ignored.

Value

A data.table with scores

Details

See the hubverse documentation for the expected format of the oracle output data.

Default metrics are provided by the scoringutils package. You can select metrics by passing in a character vector of metric names to the metrics argument.

The following metrics can be selected (all are used by default) for the different output_types:

Quantile forecasts: (output_type == "quantile")

wis
overprediction
underprediction
dispersion
bias
ae_median
"interval_coverage_XX": interval coverage at the "XX" level. For example, "interval_coverage_95" is the 95% interval coverage rate, which would be calculated based on quantiles at the probability levels 0.025 and 0.975.

See scoringutils::get_metrics.forecast_quantile for details.

Nominal forecasts: (output_type == "pmf" and output_type_id_order is NULL)

log_score

(scoring for ordinal forecasts will be added in the future).

See scoringutils::get_metrics.forecast_nominal for details.

Median forecasts: (output_type == "median")

ae_point: absolute error of the point forecast (recommended for the median, see Gneiting (2011))

See scoringutils::get_metrics.forecast_point for details.

Mean forecasts: (output_type == "mean")

se_point: squared error of the point forecast (recommended for the mean, see Gneiting (2011))

See scoringutils::add_relative_skill for details on relative skill scores.

References

Gneiting, Tilmann. 2011. "Making and Evaluating Point Forecasts." Journal of the American Statistical Association 106 (494): 746–62. <doi: 10.1198/jasa.2011.r10138>.

Examples

# compute WIS and interval coverage rates at 80% and 90% levels based on
# quantile forecasts, summarized by the mean score for each model
quantile_scores <- score_model_out(
  model_out_tbl = hubExamples::forecast_outputs |>
    dplyr::filter(.data[["output_type"]] == "quantile"),
  oracle_output = hubExamples::forecast_oracle_output,
  metrics = c("wis", "interval_coverage_80", "interval_coverage_90"),
  relative_metrics = "wis",
  by = "model_id"
)
quantile_scores
#> Key: <model_id>
#>             model_id      wis interval_coverage_80 interval_coverage_90
#>               <char>    <num>                <num>                <num>
#> 1: Flusight-baseline 329.4545                  0.0               0.1250
#> 2:   MOBS-GLEAM_FLUH 315.2393                  0.5               0.5625
#> 3:          PSI-DICE 227.9527                  0.5               0.5000
#>    wis_relative_skill
#>                 <num>
#> 1:          1.1473659
#> 2:          1.0978597
#> 3:          0.7938733

# compute log scores based on pmf predictions for categorical targets,
# summarized by the mean score for each combination of model and location.
# Note: if the model_out_tbl had forecasts for multiple targets using a
# pmf output_type with different bins, it would be necessary to score the
# predictions for those targets separately.
pmf_scores <- score_model_out(
  model_out_tbl = hubExamples::forecast_outputs |>
    dplyr::filter(.data[["output_type"]] == "pmf"),
  oracle_output = hubExamples::forecast_oracle_output,
  metrics = "log_score",
  by = c("model_id", "location", "horizon")
)
head(pmf_scores)
#>             model_id location horizon   log_score
#>               <char>   <char>   <int>       <num>
#> 1: Flusight-baseline       25       0  0.02107606
#> 2: Flusight-baseline       25       1  6.69652380
#> 3: Flusight-baseline       25       2 17.73313203
#> 4: Flusight-baseline       25       3         Inf
#> 5: Flusight-baseline       48       0  2.18418007
#> 6: Flusight-baseline       48       1  7.49960792