crps {isodistrreg}

Continuous ranked probability score (CRPS)

Description

Computes the CRPS of IDR or raw forecasts.

Usage

crps(predictions, y)

## S3 method for class 'idr'
crps(predictions, y)

## S3 method for class 'data.frame'
crps(predictions, y)

Arguments

predictions

either an object of class idr (output of predict.idrfit), or a data.frame of numeric variables. In the latter case, the CRPS is computed using the empirical distribution of the variables in predictions.

y

a numeric vector of obervations of the same length as the number of predictions, or of length 1. In the latter case, y will be used for all predictions.

Details

This function uses adapted code taken from the function crps_edf of the scoringRules package.

Value

A vector of CRPS values.

References

Jordan A., Krueger F., Lerch S. (2018). "Evaluating Probabilistic Forecasts with scoringRules." Journal of Statistical Software. Forthcoming.

Gneiting, T. and Raftery, A. E. (2007), 'Strictly proper scoring rules, prediction, and estimation', Journal of the American Statistical Association 102(477), 359-378

See Also

predict.idrfit

Examples

data("rain")

## Postprocess HRES forecast using data of 3 years

X <- rain[1:(3 * 365), "HRES", drop = FALSE]
y <- rain[1:(3 * 365), "obs"]

fit <- idr(y = y, X = X)

## Compute CRPS of postprocessed HRES forecast using data of the next 2 years
## (out-of-sample predictions)

data <- rain[(3 * 365 + 1):(5 * 365), "HRES", drop = FALSE]
obs <- rain[(3 * 365 + 1):(5 * 365), "obs"]
predictions <- predict(fit, data = data)
idrCrps <- crps(predictions, y = obs)

## Compare this to CRPS of the raw ensemble of all forecasts (high resolution,
## control and 50 perturbed ensemble forecasts)

rawData <- rain[(3 * 365 + 1):(5 * 365), c("HRES", "CTR", paste0("P", 1:50))]
rawCrps <- crps(rawData, y = obs)

c("idr_HRES" = mean(idrCrps), "raw_all" = mean(rawCrps))