weighted_interval_score: Compute weighted interval score

In yardstick: Tidy Characterizations of Model Performance

Compute weighted interval score

Description

Weighted interval score (WIS), a well-known quantile-based approximation of the commonly-used continuous ranked probability score (CRPS). WIS is a proper score, and can be thought of as a distributional generalization of absolute error. For example, see Bracher et al. (2020) for discussion in the context of COVID-19 forecasting.

Usage

weighted_interval_score(data, ...)

## S3 method for class 'data.frame'
weighted_interval_score(
  data,
  truth,
  estimate,
  quantile_levels = NULL,
  na_rm = TRUE,
  quantile_estimate_nas = c("impute", "drop", "propagate"),
  case_weights = NULL,
  ...
)

weighted_interval_score_vec(
  truth,
  estimate,
  quantile_levels = NULL,
  na_rm = FALSE,
  quantile_estimate_nas = c("impute", "drop", "propagate"),
  case_weights = NULL,
  ...
)

Arguments

data	A data.frame containing the columns specified by the truth and estimate arguments.
...	Not Currently used.
truth	The column identifier for the true class results (that is a numeric). This should be an unquoted column name although this argument is passed by expression and supports quasiquotation (you can unquote column names). For ⁠_vec()⁠ functions, a numeric vector.
estimate	The column identifier for the predicted class results (that is also quantile_pred). As with truth this can be specified different ways but the primary method is to use an unquoted variable name. For ⁠_vec()⁠ functions, a quantile_pred vector.
quantile_levels	probabilities. If specified, the score will be computed at this set of levels. Otherwise, those present in x will be used. If quantile_levels do not exactly match those available in x, then some quantiles will have implicit missingness. Handling of these is determined by quantile_estimate_nas.
na_rm	logical. If TRUE, missing values in actual or both implicit and explicit (values of NA present in x), will be ignored (dropped) in the calculation of the summary score. If FALSE (the default), any NAs will result in the summary being NA.
quantile_estimate_nas	character. This argument applies only to x. It handles imputation of individual quantile_levels that are necessary to compute a score. Because each element of x is a hardhat::quantile_pred, it is possible for these to be missing for particular quantile_levels. There are a number of different possibilities for such missingness. The options are as follows: For "impute", both explicit and implicit missing values will be imputed using hardhat::impute_quantiles() prior to the calculation of the score. So the score will be NA only if imputation fails. For "drop", any explicit missing values will be removed before calculating the score for a particular prediction. This may be reasonable due to the weighting. For example, if the estimate has quantile_levels = c(.25, .5, .75) but the median is NA for a particular prediction, it may be reasonable to average the accuracy of c(.25, .75) for that prediction with others that don't have missingness. This option is only works if quantile_levels = NULL or is a subset of the quantile_levels in x. For "propagate", any missing value predictions will result in that element of x having a score of NA. If na_rm = TRUE, then these will be removed before averaging.
case_weights	The optional column identifier for case weights. This should be an unquoted column name that evaluates to a numeric column in data. For ⁠_vec()⁠ functions, a numeric vector, hardhat::importance_weights(), or hardhat::frequency_weights().

Details

Weighted interval score is a metric that should be minimized. The output ranges from 0 to Inf, with 0 indicating perfect predictions.

Value

a vector of nonnegative scores.

See Also

All quantile metrics

Examples

library(hardhat)

quantile_levels <- c(.2, .4, .6, .8)
pred1 <- 1:4
pred2 <- 8:11
preds <- quantile_pred(rbind(pred1, pred2), quantile_levels)
truth <- c(3.3, 7.1)
weighted_interval_score_vec(truth, preds)
weighted_interval_score_vec(truth, preds, quantile_levels = c(.25, .5, .75))

# Missing value behaviours

preds_na <- quantile_pred(rbind(pred1, c(1, 2, NA, 4)), 1:4 / 5)
truth <- c(2.5, 2.5)
weighted_interval_score_vec(truth, preds_na)
weighted_interval_score_vec(truth, preds_na, quantile_levels = 1:9 / 10)
try(weighted_interval_score_vec(
  truth,
  preds_na,
  quantile_levels = 1:9 / 10,
  quantile_estimate_nas = "drop"
))
weighted_interval_score_vec(
  truth,
  preds_na,
  quantile_levels = c(2, 3) / 5,
  quantile_estimate_nas = "drop"
)
weighted_interval_score_vec(
  truth, preds_na, na_rm = TRUE, quantile_estimate_nas = "propagate"
)
weighted_interval_score_vec(
  truth, preds_na, quantile_estimate_nas = "propagate"
)

yardstick documentation built on April 8, 2026, 1:06 a.m.