In jobstdavid/eppverification: Verification Tools for the Statistical Postprocessing of Ensemble Forecasts

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  fig.path = "man/figures/",
  out.width = "100%"
)

eppverification: Verification Tools for the Statistical Postprocessing of Ensemble Forecasts

An R package providing user-friendly univariate and multivariate verification tools for the statistical ensemble post-processing. It allows to score and assess the calibration (reliability) and sharpness of ensemble forecasts and predictive distributions. In addition this package can be used to create useful contemporary visualizations for verification.

Installation

You can install the latest development version from GitHub with:

#install.packages("remotes")
remotes::install_github("jobstdavid/eppverification")

Package overview

The goal of probabilistic forecasting is to maximize the sharpness of the probabilistic forecast F (CDF) subject to calibration. Therefore this package contains tools for assessing:

• calibration. It refers to the statistical consistency between the predictive probabilistic forecast F and the associated observation y. Consequently it is a joint property of predictions and verifications. The predictive probabilistic forecast F is calibrated, if the observation y can not be distinguished from a random draw from the predictive probabilistic forecast F.
• sharpness. It refers to the concentration of the predictive probabilistic forecast F. Additionally it is a property of the predictive probabilistic forecast F, only. The more concentrated/narrower the predictive probabilistic forecast F is, the sharper the forecast is.
• calibration and sharpness simultaneously. Proper scoring rules assess calibration and sharpness properties of the predictive probabilistic forecast F simultaneously. They are functions of the predictive probabilistic forecast F and the associated observation y. A smaller score of a proper scoring rule indicates a "better" forecast.

In the following you find univariate and multivariate verification tools for calibration, sharpness and proper scoring rules, where the corresponding function provided within this package is written in brackets.

Univariate Verification Tools

• Calibration: Verification Rank Histogram (vr.hist), Reliability Index (ri), Entropy (ent), PIT Histogram (pit.hist), Central Prediction Interval Coverage (cpi).
• Sharpness: Root Mean Variance (rmv), Central Prediction Interval Width (cpi).
• Dispersion: Variance of PIT-Values (var.pit).
• Proper Scoring Rules: Continuous Ranked Probability Score (crps), Logarithmic Score (logs), Interval Score (is), Quantile Score (qs), Brier Score (bs), Dawid-Sebastiani Score (dss), Absolute Error (ae), Squared Error (se).

Multivariate Verification Tools

• Calibration: Multivariate Verification Rank Histogram (mvr.hist), Multivariate Reliability Index (mri), Multivariate Entropy (ment).
• Sharpness: Determinant Sharpness (ds).
• Proper Scoring Rules: Logarithmic Score (logs), Energy Score (es), Euclidean Error (ee), Variogram Score (vs).

Further functions for model comparison and visualizations in this package are:

• dm.test: This function performs a Diebold-Mariano-Test for two forecasts.
• bh.test: This function performs a Benjamini-Hochberg-Correction for different p-values.
• ss: This function calculates the skill score.
• cpi.plot: This function plots the central prediction intervals for a certain interval range.
• cov.plot: This function plots the central prediction interval coverage for certain interval ranges of different models.
• score.plot: This function plots scores of different models rated by selected measures as heatmap.

Examples

# load R package
library(eppverification)

# set.seed for reproducibility 
set.seed(2023)

Univariate Verification Tools

Calibration

# simulated data
n <- 30
m <- 50
y <- rnorm(n)
x <- matrix(rnorm(n*m), ncol = m)

# Verification Rank Histogram
vr.hist(y = y, x = x, bins = 3, reliability = TRUE, entropy = TRUE)

# simulated data
n <- 10000
u <- runif(n)

# PIT Histogram
pit.hist(u = u, bins = 5, dispersion = TRUE, bias = TRUE)

Sharpness

# Root Mean Variance
rmv(x = x)

Proper Scoring Rules

# simulated data
n <- 30
m <- 50
y <- rnorm(n)
x <- matrix(rnorm(n*m), ncol = m)

# Continuous Ranked Probability Score
crps(y = y, x = x, method = "ens", aggregate = mean)

# simulated data
n <- 30
y <- rnorm(n, mean = 1:n)
nominal.coverage <- 90
alpha <- (100-nominal.coverage)/100
lower <- qnorm(alpha/2, rnorm(n, mean = 1:n))
upper <- qnorm((1-alpha/2), rnorm(n, mean = 1:n))

# Central Prediction Interval Values
cpi(y = y, lower = lower, upper = upper, nominal.coverage = nominal.coverage,
separate = c("is", "overprediction", "underprediction", "width", "coverage"), aggregate = mean)

Multivariate Verification Tools

Calibration

# simulated data
n <- 30
m <- 50
y <- cbind(rnorm(n), rgamma(n, shape = 1))
x <- array(NA, dim = c(m, 2, n))
x[, 1, ] <- rnorm(n*m)
x[, 2, ] <- rgamma(n*m, shape = 1)

# Multivariate Verification Rank Histogram
mvr.hist(y = y, x = x, method = "mv", type = "absolute", bins = 17)

Sharpness

# simulated data
n <- 30
m <- 50
x <- array(NA, dim = c(2, 2, n))
for (i in 1:n) {
x[, , i] <- cov(cbind(rnorm(m), rgamma(m, shape = 1)))
}

#Determinant Sharpness
ds(x = x, covmat = TRUE, aggregate = mean)

Proper Scoring Rules

# simulated data
n <- 30
m <- 50
y <- cbind(rnorm(n), rgamma(n, shape = 1))
x <- array(NA, dim = c(m, 2, n))
x[, 1, ] <- rnorm(n*m)
x[, 2, ] <- rgamma(n*m, shape = 1)

# Energy Score
es(y = y, x = x, method = "ens", aggregate = mean)

# Euclidean Error
ee(y = y, x = x, method = "median", aggregate = mean)

Model Comparison and Visualizations

# simulated data
n <- 365
s1 <- arima.sim(list(ar = 0.7), sd = 0.5, 100)
s2 <- arima.sim(list(ar = 0.7), sd = 0.5, 100) - 0.2
p <- runif(100, min = 0, max = 0.05)

# Diebold-Mariano-Test
dm.test(s1, s2, alternative = "two.sided", h = 1)

# Benjamini-Hochberg-Procedure
bh.test(p, alpha = 0.05)

# simulated data
n <- 30
x <- seq(Sys.Date(), by = "day", length.out = n)
y <- rnorm(n, mean = 1:n)
nominal.coverage <- 90
alpha <- (100-nominal.coverage)/100
lower <- qnorm(alpha/2, rnorm(n, mean = 1:n))
upper <- qnorm((1-alpha/2), rnorm(n, mean = 1:n))

# Central Prediction Intervals Plot
cpi.plot(x = x, y = y, lower = lower, upper = upper, nominal.coverage = nominal.coverage, x.lab = "Date", y.lab = "Value", info = TRUE)

# simulated data
n <- 30
x <- matrix(runif(n)*100, ncol = 3)
x <- apply(x, 2, sort)
nominal.coverage <- seq(5, 95, length.out = 10)
models <- c("A", "B", "C")

# Central Prediction Interval Coverage Model Comparison
cov.plot(x = x, models = models, nominal.coverage = nominal.coverage)

# simulated data
x <- matrix(c(0.5, 0.3, 0.8, 0.21, 1.5, 0.7, 2, 1), byrow = TRUE, ncol = 4)
models <- c("A", "B", "C", "D")
measures <- c("CRPS", "LogS")

# Score Plot
score.plot(x = x, models = models, measures = measures)

Contact

Feel free to contact [e-mail] if you have any questions or suggestions.

References

Gneiting, T. and Raftery, A. (2007). Strictly Proper Scoring Rules, Prediction, and Estimation. Journal of the American Statistical Association. 102(477). 359-378.

jobstdavid/eppverification documentation built on May 13, 2024, 5:20 p.m.