Closed form expressions for the continuous ranked probability score

Alexander Jordan

2017-11-03

1 Introduction

The continuous ranked probability score can be given in multiple equivalent forms, \[\begin{align*} \label{eq:kernel}\mathrm{CRPS}(F,y) &= \mathbb{E}_F|Y-y| - \frac{1}{2}\mathbb{E}_F|Y-Y'| \\ &= \int_{-\infty}^y F(x)^2 \, \mathrm{d} x + \int_y^{\infty} \left(1 - F(x)\right)^2 \, \mathrm{d} x \\ &= 2\int_0^{F(y)} \alpha\left(y - F^{-1}(\alpha)\right) \, \mathrm{d} \alpha + 2\int_{F(y)}^1 (1 - \alpha)\left(F^{-1}(\alpha) - y\right) \, \mathrm{d} \alpha, \end{align*}\]

where the first is the kernel representation, followed by the threshold decomposition, and lastly the quantile decomposition. The threshold decomposition corresponds to the integral of the Brier score over all event thresholds, while the quantile decomposition is the integral of the quantile score over all probabilities.

2 Distributions for variables on the real line

2.1 Laplace distribution ("lapl")

CRPS formula: \[\begin{align*} \mathrm{CRPS}(F, y) &= |y| + \exp(-|y|) - \frac{3}{4}, \\ \mathrm{CRPS}(F_{\mu, \sigma}, y) &= \sigma\, \mathrm{CRPS}\left(F, \tfrac{y - \mu}{\sigma}\right). \end{align*}\] Laplace distribution: \[\begin{align*} F(x) &= \begin{cases} \frac{1}{2} \exp(x), & x < 0,\\ 1 - \frac{1}{2} \exp(-x), & x \geq 0, \end{cases} \\ F_{\mu, \sigma}(x) &= F\left(\tfrac{x - \mu}{\sigma}\right). \end{align*}\]

Parameters:

Name Domain
"location" \(\mu \in \mathbb{R}\)
"scale" \(\sigma > 0\)

See also: Two-piece exponential distribution

2.2 Logistic distribution ("logis")

CRPS formula: \[\begin{align*} \mathrm{CRPS}(F, y) &= y - 2\log(F(y)) - 1, \\ \mathrm{CRPS}(F_{\mu, \sigma}, y) &= \sigma\, \mathrm{CRPS}\left(F, \tfrac{y - \mu}{\sigma} \right). \end{align*}\] Logistic distribution: \[\begin{align*} F(x) &= \frac{1}{1 + \exp(-x)}, \\ F_{\mu, \sigma}(x) &= F\left(\tfrac{x - \mu}{\sigma}\right). \end{align*}\]

Parameters:

Name Domain
"location" \(\mu \in \mathbb{R}\)
"scale" \(\sigma > 0\)

See also: Generalized truncated/censored logistic distribution

2.3 Normal distribution ("norm")

CRPS formula: \[\begin{align*} \mathrm{CRPS}(F, y) &= y\left(2F(y)-1\right) + 2f(y) - \frac{1}{\sqrt{\pi}}, \\ \mathrm{CRPS}(F_{\mu, \sigma}, y) &= \sigma\, \mathrm{CRPS}\left(F, \tfrac{y - \mu}{\sigma} \right). \end{align*}\] Normal distribution: \[\begin{align*} f(x) &= \tfrac{1}{\sqrt{2\pi}}\exp\left(-\tfrac{x^2}{2}\right), \\ F(x) &= \int_{-\infty}^x f(t)\, \mathrm{d} x, \\ F_{\mu, \sigma}(x) &= F\left(\tfrac{x - \mu}{\sigma}\right). \end{align*}\]

Parameters:

Name Domain
"mean" \(\mu \in \mathbb{R}\)
"sd" \(\sigma > 0\)

Notes:

See also: Mixture of normal distributions, Two-piece normal distribution, Generalized truncated/censored normal distribution

2.4 Mixture of normal distributions ("mixnorm")

CRPS formula: \[\begin{align*} \mathrm{CRPS}(F,y) &= \sum_{i=1}^M \omega_i A\left(y-\mu_i,\sigma_i^2\right) - \frac{1}{2} \sum_{i=1}^M\sum_{j=1}^{M}\omega_i \omega_j A\left(\mu_i-\mu_j,\sigma_i^2+\sigma_j^2\right). \end{align*}\] Finite mixture of normal distributions: \[\begin{align*} F(x) &= \sum_{i=1}^M \omega_i \Phi\left(\tfrac{x-\mu_i}{\sigma_i}\right), \\ A\left(\mu,\sigma^2\right) &= \mu\left(2\Phi\left(\tfrac{\mu}{\sigma}\right) -1 \right) + 2\sigma \varphi\left(\tfrac{\mu}{\sigma}\right), \\ \varphi(x) &= \tfrac{1}{\sqrt{2\pi}}\exp\left(-\tfrac{x^2}{2}\right), \\ \Phi(x) &= \int_{-\infty}^x \varphi(t)\, \mathrm{d} x. \end{align*}\]

Parameters:

Name Domain
"m" \(\mu_1, \ldots, \mu_M \in \mathbb{R}\)
"s" \(\sigma_1, \ldots, \sigma_M > 0\)
"w" \(\omega_1, \ldots, \omega_M > 0\), \(\omega_1 + \ldots + \omega_M = 1\)

Notes:

See also: Normal distribution

2.5 Student’s \(t\)-distribution ("t")

CRPS formula: \[\begin{align*} \mathrm{CRPS}(F_\nu, y) &= y\Big(2F_\nu(y) - 1\Big) + 2f_\nu(y) \left(\frac{\nu + y^2}{\nu - 1}\right) - \frac{2\sqrt{\nu}}{\nu - 1}\frac{B(\tfrac{1}{2}, \nu - \tfrac{1}{2})}{B(\tfrac{1}{2}, \tfrac{\nu}{2})^2}, \\ \mathrm{CRPS}(F_{\nu, \mu, \sigma}, y) &= \sigma\, \mathrm{CRPS}\left(F_\nu, \tfrac{y - \mu}{\sigma} \right). \end{align*}\] Student’s \(t\)-distribution: \[\begin{align*} f_\nu(x) &= \frac{1}{\sqrt{\nu}B(\tfrac{1}{2},\tfrac{\nu}{2})}\left(1+\frac{x^2}{\nu}\right)^{-\tfrac{\nu+1}{2}}, \\ F_\nu(x) &= \frac{1}{2} + \frac{x\ {}_2F_1(\tfrac{1}{2},\tfrac{\nu+1}{2};\tfrac{3}{2};-\tfrac{x^2}{\nu})}{\sqrt{\nu} B(\tfrac{1}{2},\tfrac{\nu}{2})}, \\ F_{\nu, \mu, \sigma}(x) &= F_\nu\left(\tfrac{x - \mu}{\sigma}\right). \end{align*}\]

Parameters:

Name Domain
"df" \(\nu > 1\) (for \(\nu \in (0, 1]\) the CRPS is undefined)
"location" \(\mu \in \mathbb{R}\)
"scale" \(\sigma > 0\)

Mathematical functions:

Symbol Name
\({}_2F_1\) hypergeometric function
\(B\) beta function

See also: Generalized truncated/censored Student’s \(t\)-distribution

2.6 Two-piece exponential distribution ("2pexp")

CRPS formula: \[\begin{align*} \mathrm{CRPS}(F_{\sigma_1, \sigma_2}, y) &= \begin{cases} \left\lvert y \right\rvert + \frac{2\sigma_1^2}{\sigma_1 + \sigma_2}\exp\left(-\left\lvert \frac{y}{\sigma_1}\right\rvert \right) - \frac{ 2\sigma_1^2}{\sigma_1 + \sigma_2} + \frac{\sigma_1^3 + \sigma_2^3}{2(\sigma_1 + \sigma_2)^2}, & y < 0, \\ \left\lvert y \right\rvert + \frac{2\sigma_2^2}{\sigma_1 + \sigma_2}\exp\left(-\left\lvert \frac{y}{\sigma_2}\right\rvert \right) - \frac{ 2\sigma_2^2}{\sigma_1 + \sigma_2} + \frac{\sigma_1^3 + \sigma_2^3}{2(\sigma_1 + \sigma_2)^2}, & y \ge 0, \end{cases} \\ \mathrm{CRPS}(F_{\mu, \sigma_1, \sigma_2}, y) &= \mathrm{CRPS}(F_{\sigma_1, \sigma_2}, y - \mu). \end{align*}\] Two-piece exponential distribution: \[\begin{align*} F_{\sigma_1, \sigma_2}(x) &= \begin{cases} \frac{\sigma_1}{\sigma_1 + \sigma_2}\exp\left(\frac{x}{\sigma_1}\right), & x < 0, \\ 1 - \frac{\sigma_2}{\sigma_1 + \sigma_2}\exp\left(-\frac{x}{\sigma_2}\right), & x \ge 0,\end{cases} \\ F_{\mu, \sigma_1, \sigma_2}(x) &= F_{\sigma_1, \sigma_2}(x - \mu). \end{align*}\]

Parameters:

Name Domain
"location" \(\mu \in \mathbb{R}\)
"scale1" \(\sigma_1 > 0\)
"scale2" \(\sigma_2 > 0\)

See also: Laplace

2.7 Two-piece normal distribution ("2pnorm")

CRPS formula: \[\begin{align*} \mathrm{CRPS}(F_{\sigma_1, \sigma_2}, y) &= \sigma_1\, \mathrm{CRPS}\left(F_{-\infty, 0}^{0, \sigma_2/(\sigma_1 + \sigma_2)}, \tfrac{\min(0, y)}{\sigma_1}\right) \\ &\quad + \sigma_2\, \mathrm{CRPS}\left(F_{0, \sigma_1/(\sigma_1 + \sigma_2)}^{\infty, 0}, \tfrac{\max(0, y)}{\sigma_2}\right), \\ \mathrm{CRPS}(F_{\mu, \sigma_1, \sigma_2}, y) &= \mathrm{CRPS}(F_{\sigma_1, \sigma_2}, y - \mu), \end{align*}\]

with \(F_{l, L}^{u, U}\) as in Generalized truncated/censored normal distribution.

Two-piece normal distribution: \[\begin{align*} F_{\sigma_1,\sigma_2}(x) &= \begin{cases} \frac{2\sigma_1}{\sigma_1+\sigma_2}\Phi\left(\frac{x}{\sigma_1}\right), & x < 0,\\ \frac{\sigma_1-\sigma_2}{\sigma_1+\sigma_2} + \frac{2\sigma_2}{\sigma_1+\sigma_2} \Phi\left(\frac{x}{\sigma_2}\right), & x \ge 0, \end{cases} \\ F_{\mu, \sigma_1, \sigma_2}(x) &= F_{\sigma_1, \sigma_2}(x - \mu), \\ \Phi(x) &= \int_{-\infty}^x \varphi(t)\, \mathrm{d} x, \\ \varphi(x) &= \tfrac{1}{\sqrt{2\pi}}\exp\left(-\tfrac{x^2}{2}\right). \end{align*}\]

Parameters:

Name Domain
"location" \(\mu \in \mathbb{R}\)
"scale1" \(\sigma_1 > 0\)
"scale2" \(\sigma_2 > 0\)

Notes:

See also: Normal distribution

3 Distributions for non-negative variables

3.1 Exponential distribution ("exp")

CRPS formula: \[\begin{align*} \mathrm{CRPS}(F_\lambda, y) &= |y| - \frac{2F_\lambda(y)}{\lambda} + \frac{1}{2\lambda}. \end{align*}\] Exponential distribution: \[\begin{align*} F_\lambda(x) = \begin{cases} 1 - \exp(-\lambda x), & x \ge 0, \\ 0, & x < 0.\end{cases} \end{align*}\]

Parameters:

Name Domain
"rate" \(\lambda > 0\)

See also: Exponential distribution with point mass, Generalized Pareto distribution with point mass

3.2 Gamma distribution ("gamma")

CRPS formula: \[ \mathrm{CRPS}(F_{\alpha,\beta},y) = y\left(2F_{\alpha,\beta}(y)-1\right) - \frac{\alpha}{\beta}\left(2F_{\alpha + 1, \beta}(y) -1\right) - \frac{1}{\beta B\left(\tfrac{1}{2},\alpha\right)} \] Gamma distribution: \[\begin{align*} F_{\alpha,\beta}(x) &= \begin{cases}\frac{\Gamma_l(\alpha,\beta x)}{\Gamma(\alpha)}, & x \geq 0, \\ 0, & x < 0. \end{cases} \end{align*}\]

Parameters:

Name Domain
"shape" \(\alpha > 0\)
"rate" \(\beta > 0\)

Mathematical functions:

Symbol Name
\(B\) beta function
\(\Gamma\) gamma function
\(\Gamma_l\) lower incomplete gamma function

Notes:

3.3 Log-Laplace distribution ("llapl")

CRPS formula: \[\begin{align*} \mathrm{CRPS}(F_{\mu, \sigma}, y) &= y\Big(2F_{\mu, \sigma}(y) - 1\Big) + \exp(\mu) \left(\tfrac{\sigma}{4 - \sigma^2} + A(y) \right) \end{align*}\] Log-Laplace distribution: \[\begin{align*} F_{\mu, \sigma}(x) &= \begin{cases} 0, & x \leq 0,\\ \frac{1}{2}\exp\left(\frac{\log x - \mu}{\sigma}\right), & 0 < x < \exp(\mu), \\ 1 - \frac{1}{2}\exp\left(-\frac{\log x - \mu}{\sigma}\right), & x \geq \exp(\mu), \end{cases} \\ A(x) &= \begin{cases} \frac{1}{1 + \sigma}\left(1-\left(2F_{\mu, \sigma}(x)\right)^{1+\sigma}\right), & x < \exp(\mu), \\ -\frac{1}{1-\sigma}\left(1-\left(2(1-F_{\mu, \sigma}(x))\right)^{1-\sigma}\right), & y \geq \exp(\mu). \end{cases} \end{align*}\]

Parameters:

Name Domain
"locationlog" \(\mu \in \mathbb{R}\)
"scalelog" \(\sigma \in (0, 1)\) (for \(\sigma \ge 1\) the CRPS is undefined)

3.4 Log-logistic distribution ("llogis")

CRPS formula: \[\begin{align*} \mathrm{CRPS}(F_{\mu, \sigma}, y) &= y\left(2F_{\mu, \sigma}(y) - 1\right) - 2\exp(\mu) B(F_{\mu, \sigma}(y); 1 + \sigma, 1 - \sigma) \\ &\quad + \exp(\mu)(1 - \sigma)B(1 + \sigma, 1 - \sigma) \end{align*}\] Log-logistic distribution: \[\begin{align*} F_{\mu, \sigma}(x) &= \begin{cases} 0, & x \leq 0, \\ \left(1 + \exp\left(-\tfrac{\log x - \mu}{\sigma}\right)\right)^{-1}, & x > 0, \end{cases} \end{align*}\]

Parameters:

Name Domain
"locationlog" \(\mu \in \mathbb{R}\)
"scalelog" \(\sigma \in (0, 1)\) (for \(\sigma \ge 1\) the CRPS is undefined)

Mathematical functions:

Symbol Name
\(B(\cdot, \cdot)\) beta function
\(B(x; \cdot, \cdot)\) incomplete beta function

Notes:

3.5 Log-normal distribution ("lnorm")

\[\begin{align*} \mathrm{CRPS}(F_{\mu,\sigma},y) &= y\left(2F_{\mu, \sigma}(y) - 1\right) - 2 \exp(\mu+\sigma^2/2)\left(\Phi\left(\tfrac{\log y -\mu - \sigma^2}{\sigma}\right) + \Phi\left(\tfrac{\sigma}{\sqrt{2}}\right) - 1\right) \end{align*}\] Log-normal distribution: \[\begin{align*} F_{\mu,\sigma}(x) &= \begin{cases} 0, & x\leq 0, \\ \Phi\left(\tfrac{\log x - \mu}{\sigma}\right), & x > 0, \end{cases} \\ \varphi(x) &= \tfrac{1}{\sqrt{2\pi}}\exp\left(-\tfrac{x^2}{2}\right), \\ \Phi(x) &= \int_{-\infty}^x \varphi(t)\, \mathrm{d} x. \end{align*}\]

Parameters:

Name Domain
"locationlog" \(\mu \in \mathbb{R}\)
"scalelog" \(\sigma > 0\)

Notes:

4 Distributions with flexible support and/or point masses

4.1 Beta distribution ("beta")

CRPS formula: \[\begin{align*} \mathrm{CRPS}(F_{\alpha, \beta}, y) &= y(2F_{\alpha, \beta}(y) - 1) + \frac{\alpha}{\alpha + \beta} \left(1 - 2F_{\alpha + 1, \beta}(y) - \frac{2B(2\alpha, 2\beta)}{\alpha B(\alpha, \beta)^2} \right), \\ \mathrm{CRPS}(F_{l, \alpha, \beta}^{u}, y) &= (u - l)\, \mathrm{CRPS}\left(F_{\alpha, \beta}, \tfrac{y - l}{u - l} \right). \end{align*}\] Beta distribution: \[\begin{align*} F_{\alpha, \beta}(x) &= \begin{cases} 0 & x < 0\\ I(x; \alpha, \beta) & 0 \leq x < 1\\ 1 & x \geq 1 \end{cases}, \\ F_{l, \alpha, \beta}^{u}(x) &= F_{\alpha, \beta}\left(\tfrac{x - l}{u - l}\right). \end{align*}\]

Parameters:

Name Domain
"shape1" \(\alpha > 0\)
"shape2" \(\beta > 0\)
"lower" \(l \in \mathbb{R}\), \(l < u\)
"upper" \(u \in \mathbb{R}\), \(l < u\)

Mathematical functions:

Symbol Name
\(B\) beta function
\(I\) regularized incomplete beta function

Notes:

See also: Continuous uniform distribution

4.2 Continuous uniform ("unif")

CRPS formula: \[\begin{align*} \mathrm{CRPS}(F, y) &= |y - z| + z^2 - z + \frac{1}{3}, \\ \text{where}\quad z &= \begin{cases} 0, & y < 0, \\ y, & 0 \le y < 1, \\ 1, & y \ge 1. \end{cases} \end{align*}\] Continuous uniform: \[\begin{align*} F(x) &= \begin{cases} 0, & x < 0,\\ x, & 0 \leq x < 1,\\ 1, & x > 1. \end{cases} \end{align*}\]

See also: Beta distribution

4.2.1 Generalized continuous uniform ("unif")

CRPS formula: \[\begin{align*} \mathrm{CRPS}(F_{L}^{U}, y) &= |y - z| + z^2(1 - L - U) - z(1 - 2L) \\ &\quad + \frac{(1 - L - U)^2}{3} + (1 - L)U, \\ \text{where}\quad z &= \begin{cases} 0, & y < 0, \\ y, & 0 \le y < 1, \\ 1, & y \ge 1, \end{cases} \\ \mathrm{CRPS}(F_{l, L}^{u, U}, y) &= (u - l)\, \mathrm{CRPS}\left(F_{L}^{U}, \tfrac{y - l}{u - l} \right). \end{align*}\] Continuous uniform with point masses and support transformation: \[\begin{align*} F_{L}^{U}(x) &= \begin{cases} 0, & x < 0, \\ L + (1 - L - U) x, & 0 \leq x < 1,\\ 1, & x \geq 1, \end{cases} \\ F_{l, L}^{u, U}(x) &= F_{L}^{U}\left(\tfrac{x - l}{u - l}\right). \end{align*}\]

Parameters:

Name Domain
"min" \(l \in \mathbb{R}\), \(l < u\)
"max" \(u \in \mathbb{R}\), \(l < u\)
"lmass" \(L \ge 0\), \(L + U < 1\)
"umass" \(U \ge 0\), \(L + U < 1\)

4.3 Exponential distribution with point mass ("expM")

\[\begin{align*} \mathrm{CRPS}(F_{M}, y) &= |y| - 2 (1 - M)F(y) + \frac{(1 - M)^2}{2}, \\ \mathrm{CRPS}(F_{M, \mu, \sigma}, y) &= \sigma\, \mathrm{CRPS}\left(F_M, \tfrac{y - \mu}{\sigma} \right). \end{align*}\] Exponential distribution with point mass: \[\begin{align*} F_M(x) &= \begin{cases} M + (1 - M)F(x), & x \ge 0, \\ 0, & x < 0, \end{cases} \\ F(x) &= \begin{cases} 1 - \exp(-x), & x \ge 0, \\ 0, & x < 0, \end{cases} \\ F_{M, \mu, \sigma}(x) &= F_M\left(\tfrac{x - \mu}{\sigma}\right). \end{align*}\]

Parameters:

Name Domain
"location" \(\mu \in \mathbb{R}\)
"scale" \(\sigma > 0\)
"mass" \(M \in [0, 1]\)

See also: Exponential distribution, Generalized Pareto distribution with point mass

4.4 Generalized extreme value distribution ("gev")

CRPS formula: \[\begin{align*} \mathrm{CRPS}(F_{\xi, \mu, \sigma}, y) &= \sigma\, \mathrm{CRPS}\left(F_{\xi}, \tfrac{y - \mu}{\sigma} \right), \\ \text{For $\xi = 0$:}\quad \mathrm{CRPS}(F_{\xi},y) &= - y - 2 \mathrm{Ei}(\log F_{\xi}(y)) + C - \log 2, \\ \text{For $\xi \neq 0$:} \quad \mathrm{CRPS}(F_\xi, y) &= y\left(2F_\xi(y) - 1\right) - 2G_\xi(y) - \frac{1 - \left(2 - 2^\xi\right)\Gamma(1 - \xi)}{\xi}. \end{align*}\] Generalized extreme value distribution: \[\begin{align*} F_{\xi, \mu, \sigma}(x) &= F_{\xi}\left(\tfrac{x - \mu}{\sigma}\right), \\ \text{for $\xi = 0$:}\quad F_{\xi}(x) &= \exp\left(-\exp(-x)\right) \\ \text{for $\xi > 0$:}\quad F_{\xi}(x) &= \begin{cases} 0, & x \le -\frac{1}{\xi}, \\ \exp\left(-(1+\xi x)^{-1/\xi}\right), & x > -\frac{1}{\xi}, \end{cases}\\ G_{\xi}(x) &= \begin{cases} 0, & x \le -\frac{1}{\xi}, \\ -\frac{F_\xi(x)}{\xi} + \frac{\Gamma_u(1 - \xi, -\log F_\xi(x))}{\xi}, & x > -\frac{1}{\xi}, \end{cases} \\ \text{for $\xi < 0$:}\quad F_{\xi}(x) &= \begin{cases} \exp\left(-(1+\xi x)^{-1/\xi}\right), & x < -\frac{1}{\xi}, \\ 1, & x \ge -\frac{1}{\xi}, \end{cases} \\ G_\xi(x) &= \begin{cases} -\frac{F_\xi(x)}{\xi} + \frac{\Gamma_u(1 - \xi, -\log F_\xi(x))}{\xi}, & x < -\frac{1}{\xi}, \\ -\frac{1}{\xi} + \frac{\Gamma(1 - \xi)}{\xi}, & x \ge -\frac{1}{\xi}. \end{cases} \end{align*}\]

Parameters:

Name Domain
"shape" \(\xi < 1\) (for \(\xi \ge 1\) the CRPS is undefined)
"location" \(\mu \in \mathbb{R}\)
"scale" \(\sigma > 0\)

Mathematical constants and functions:

Symbol Name
\(C\) Euler-Mascheroni constant
\(\mathrm{Ei}\) exponential integral
\(\Gamma\) gamma function
\(\Gamma_u\) upper incomplete gamma function

Notes:

4.5 Generalized Pareto distribution with point mass ("gpd")

CRPS formula: \[\begin{align*} \mathrm{CRPS}(F_{M, \xi}, y) &= |y| - \frac{2(1 - M)}{1 - \xi}\left(1 - \left(1 - F_\xi(y)\right)^{1 - \xi}\right) + \frac{(1 - M)^2}{2 - \xi}, \\ \mathrm{CRPS}(F_{M, \xi, \mu, \sigma}, y) &= \sigma\, \mathrm{CRPS}\left(F_{M, \xi}, \tfrac{y - \mu}{\sigma} \right). \end{align*}\] Generalized Pareto distribution with point mass: \[\begin{align*} F_{M, \xi}(x) &= \begin{cases} M + (1 - M)F_\xi(x), & x \ge 0, \\ 0, & x < 0, \end{cases} \\ F_{M, \xi, \mu, \sigma}(x) &= F_{M, \xi}\left(\tfrac{x - \mu}{\sigma}\right) \\ \text{for $\xi > 0$:}\quad F_\xi(x) &= \begin{cases} 0, & x < 0, \\ 1 - (1 + \xi x)^{-1/\xi}, & x \ge 0, \end{cases} \\ \text{for $\xi < 0$:}\quad F_\xi(x) &= \begin{cases} 0, & x < 0, \\ 1 - (1 + \xi x)^{-1/\xi}, & 0 \le x < |\xi|^{-1}, \\ 1, & x \ge |\xi|^{-1}, \end{cases} \\ \text{for $\xi = 0$:}\quad F_\xi(x) &= \begin{cases} 0, & x < 0, \\ 1 - \exp(-x), & x \ge 0. \end{cases} \end{align*}\]

Parameters:

Name Domain
"shape" \(\xi < 1\) (for \(\xi \ge 1\) the CRPS is undefined)
"location" \(\mu \in \mathbb{R}\)
"scale" \(\sigma > 0\)
"mass" \(M \in [0, 1]\)

Notes:

See also: Exponential distribution, Exponential distribution with point mass

4.6 Generalized truncated/censored logistic distribution ("gtclogis")

CRPS formula: \[\begin{align*} \mathrm{CRPS}\left(F_{l, L}^{u, U}, y\right) &= |y - z| + uU^2 - lL^2 \\ &\quad - \left(\frac{1 - L - U}{F(u) - F(l)}\right)z \left(\frac{(1 - 2L)F(u) + (1 - 2U)F(l)}{1 - L - U}\right) \\ &\quad - \left(\frac{1 - L - U}{F(u) - F(l)}\right)\left(2\log F(-z) - 2G(u)U - 2G(l)L\right) \\ &\quad - \left(\frac{1 - L - U}{F(u) - F(l)}\right)^2 (H(u) - H(l)), \\ \text{where}\quad z&= \begin{cases} l, & y < l, \\ y, & l \le y < u, \\ u, & y \ge u, \end{cases} \\ \mathrm{CRPS}(F_{l, L, \mu, \sigma}^{u, U}, y) &= \sigma\, \mathrm{CRPS}\left(F_{(l - \mu)/\sigma, L}^{(u - \mu)/\sigma, U}, \tfrac{y - \mu}{\sigma} \right). \end{align*}\] Generalized truncated/censored logistic distribution: \[\begin{align*} F_{l, L}^{u, U}(x) &= \begin{cases} 0, & x < l, \\ \frac{1 - L - U}{F(u) - F(l)} F(z) - \frac{1 - L - U}{F(u) - F(l)} F(l) + L, & l \leq x < u, \\ 1, & x \geq u, \end{cases} \\ F_{l, L, \mu, \sigma}^{u, U}(x) &= F_{(l - \mu)/\sigma, L}^{(u - \mu)/\sigma, U}\left(\tfrac{x - \mu}{\sigma}\right), \\ F(x) &= \frac{1}{1 + \exp(-x)}, \\ G(x) &= xF(x) + \log F(-x), \\ H(x) &= F(x) - xF(x)^2 + (1 - 2F(x))\log F(-x). \end{align*}\]

Parameters:

Name Domain
"location" \(\mu \in \mathbb{R}\)
"scale" \(\sigma > 0\)
"lower" \(l \in \mathbb{R}\), \(l < u\)
"upper" \(u \in \mathbb{R}\), \(l < u\)
"lmass" \(L \ge 0\), \(L + U < 1\)
"umass" \(U \ge 0\), \(L + U < 1\)

See also: Logistic distribution

4.6.1 Censored logistic distribution ("clogis")

CRPS formula: \[\begin{align*} \mathrm{CRPS}\left(F_{l}^{u}, y\right) &= |y - z| + z + \log \left(\frac{F(-l)F(u)}{F(z)^2}\right) - F(u) + F(l), \\ \text{where}\quad z&= \begin{cases} l, & y < l, \\ y, & l \le y < u, \\ u, & y \ge u. \end{cases} \end{align*}\] Censored logistic distribution: \[\begin{align*} F_{l}^{u}(x) = \begin{cases} 0, & x < l, \\ F(x), & l \leq x < u, \\ 1, & x \geq u, \end{cases} \end{align*}\]

where all other symbols are as given in Generalized truncated/censored logistic distribution.

Notes:

  • Taillardat et al. (2016) give a CRPS formula for a left-censored logistic distribution with boundary 0.

4.6.2 Truncated logistic distribution ("tlogis")

\[\begin{align*} \mathrm{CRPS}\left(F_{l}^{u}, y\right) &= |y - z| - \frac{zF(u) + zF(l) + 2\log F(-z)}{F(u) - F(l)} - \frac{H(u) - H(l)}{(F(u) - F(l))^2}, \\ \text{where}\quad z&= \begin{cases} l, & y < l, \\ y, & l \le y < u, \\ u, & y \ge u. \end{cases} \end{align*}\] Truncated logistic distribution: \[\begin{align*} F_{l}^{u}(x) = \begin{cases} 0, & x < l, \\ \frac{F(x) - F(l)}{F(u) - F(l)}, & l \leq x < u, \\ 1, & x \geq u, \end{cases} \end{align*}\]

where all other symbols are as given in Generalized truncated/censored logistic distribution.

Notes:

  • Möller and Scheuerer (2015) give a CRPS formula for a left-truncated logistic distribution with boundary 0.

4.7 Generalized truncated/censored normal distribution ("gtcnorm")

CRPS formula: \[\begin{align*} \mathrm{CRPS}\left(F_{l, L}^{u, U}, y\right) &= |y - z| + uU^2 - lL^2 \\ &\quad + \left(\frac{1 - L - U}{F(u) - F(l)}\right)z\left(2F(z) - \frac{(1 - 2L)F(u) + (1 - 2U)F(l)}{1 - L - U}\right) \\ &\quad + \left(\frac{1 - L - U}{F(u) - F(l)}\right)\left(2f(z) - 2f(u)U - 2f(l)L\right) \\ &\quad - \left(\frac{1 - L - U}{F(u) - F(l)}\right)^2 \left(\frac{1}{\sqrt{\pi}}\right) \left(F\left(u\sqrt{2}\right) - F\left(l\sqrt{2}\right)\right), \\ \text{where}\quad z&= \begin{cases} l, & y < l, \\ y, & l \le y < u, \\ u, & y \ge u, \end{cases} \\ \mathrm{CRPS}(F_{l, L, \mu, \sigma}^{u, U}, y) &= \sigma\, \mathrm{CRPS}\left(F_{(l - \mu)/\sigma, L}^{(u - \mu)/\sigma, U}, \tfrac{y - \mu}{\sigma} \right). \end{align*}\] Generalized truncated/censored normal distribution: \[\begin{align*} F_{l, L}^{u, U}(x) &= \begin{cases} 0, & x < l, \\ \frac{1 - L - U}{F(u) - F(l)} F(z) - \frac{1 - L - U}{F(u) - F(l)} F(l) + L, & l \leq x < u, \\ 1, & x \geq u, \end{cases} \\ F_{l, L, \mu, \sigma}^{u, U}(x) &= F_{(l - \mu)/\sigma, L}^{(u - \mu)/\sigma, U}\left(\tfrac{x - \mu}{\sigma}\right), \\ F(x) &= \int_{-\infty}^x f(t)\, \mathrm{d}t, \\ f(x) &= \frac{1}{\sqrt{2\pi}}\exp(-x^2/2). \end{align*}\]

Parameters:

Name Domain
"location" \(\mu \in \mathbb{R}\)
"scale" \(\sigma > 0\)
"lower" \(l \in \mathbb{R}\), \(l < u\)
"upper" \(u \in \mathbb{R}\), \(l < u\)
"lmass" \(L \ge 0\), \(L + U < 1\)
"umass" \(U \ge 0\), \(L + U < 1\)

See also: Normal distribution

4.7.1 Censored normal distribution ("cnorm")

CRPS formula: \[\begin{align*} \mathrm{CRPS}\left(F_{l}^{u}, y\right) &= |y - z| + uF(-u)^2 - lF(l)^2 \\ & \quad + z(2F(z) - 1) \\ & \quad + 2f(z) - 2f(u)F(-u) - 2f(l)F(l) \\ & \quad - \left(\frac{1}{\sqrt{\pi}}\right) \left(F\left(u\sqrt{2}\right) - F\left(l\sqrt{2}\right)\right), \\ \text{where}\quad z&= \begin{cases} l, & y < l, \\ y, & l \le y < u, \\ u, & y \ge u. \end{cases} \end{align*}\] Censored normal distribution: \[\begin{align*} F_{l}^{u}(x) = \begin{cases} 0, & x < l, \\ F(x), & l \leq x < u, \\ 1, & x \geq u, \end{cases} \end{align*}\]

where all other symbols are as given in Generalized truncated/censored normal distribution.

Notes:

  • Thorarinsdottir and Gneiting (2010) give a CRPS formula for a left-censored normal distribution with boundary 0.

4.7.2 Truncated normal distribution ("tnorm")

CRPS formula: \[\begin{align*} \mathrm{CRPS}\left(F_{l}^{u}, y\right) &= |y - z| \\ & \quad + \left(\frac{1}{F(u) - F(l)}\right) z\left(2F(z) - F(u) - F(l)\right) \\ & \quad + \left(\frac{1}{F(u) - F(l)}\right) 2f(z) \\ & \quad - \left(\frac{1}{F(u) - F(l)}\right)^2 \left(\frac{1}{\sqrt{\pi}}\right) \left(F\left(u\sqrt{2}\right) - F\left(l\sqrt{2}\right)\right), \\ \text{where}\quad z&= \begin{cases} l, & y < l, \\ y, & l \le y < u, \\ u, & y \ge u. \end{cases} \end{align*}\] Truncated normal distribution: \[\begin{align*} F_{l}^{u}(x) = \begin{cases} 0, & x < l, \\ \frac{F(x) - F(l)}{F(u) - F(l)}, & l \leq x < u, \\ 1, & x \geq u, \end{cases} \end{align*}\]

where all other symbols are as given in Generalized truncated/censored normal distribution.

Notes:

  • Gneiting et al. (2006) give a CRPS formula for a left-truncated normal distribution with boundary 0.

4.8 Generalized truncated/censored Student’s \(t\)-distribution ("gtct")

CRPS formula: \[\begin{align*} \mathrm{CRPS}\left(F_{l, L, \nu}^{u, U}, y\right) &= |y - z| + uU^2 - lL^2 \\ &\quad + \left(\frac{1 - L - U}{F_\nu(u) - F_\nu(l)}\right) z\left(2F_\nu(z) - \frac{(1 - 2L)F_\nu(u) + (1 - 2U)F_\nu(l)}{1 - L - U}\right) \\ &\quad - \left(\frac{1 - L - U}{F_\nu(u) - F_\nu(l)}\right)\left(2G_\nu(z) - 2G_\nu(u)U - 2G_\nu(l)L\right) \\ &\quad - \left(\frac{1 - L - U}{F_\nu(u) - F_\nu(l)}\right)^2 \bar{B}_\nu \left(H_\nu(u) - H_\nu(l)\right), \\ \text{where}\quad z&= \begin{cases} l, & y < l, \\ y, & l \le y < u, \\ u, & y \ge u, \end{cases} \\ \mathrm{CRPS}(F_{l, L, \nu, \mu, \sigma}^{u, U}, y) &= \sigma\, \mathrm{CRPS}\left(F_{(l - \mu)/\sigma, L, \nu}^{(u - \mu)/\sigma, U}, \tfrac{y - \mu}{\sigma} \right). \end{align*}\] Generalized truncated/censored Student’s \(t\)-distribution: \[\begin{align*} F_{l, L, \nu}^{u, U}(x) &= \begin{cases} 0, & x < l, \\ \frac{1 - L - U}{F(u) - F(l)} F(z) - \frac{1 - L - U}{F(u) - F(l)} F(l) + L, & l \leq x < u, \\ 1, & x \geq u, \end{cases} \\ F_{l, L, \nu, \mu, \sigma}^{u, U}(x) &= F_{\tfrac{l - \mu}{\sigma}, L, \nu}^{\tfrac{u - \mu}{\sigma}, U}\left(\tfrac{x - \mu}{\sigma}\right), \\ f_\nu(x) &= \frac{1}{\sqrt{\nu}B\left(\tfrac{1}{2}, \tfrac{\nu}{2}\right)}\left(1 + \frac{x^2}{\nu}\right)^{-(\nu + 1)/2}, \\ F_\nu(x) &= \frac{1}{2} + \frac{x\ {}_2F_1\left(\tfrac{1}{2},\tfrac{\nu+1}{2};\tfrac{3}{2};-\tfrac{x^2}{\nu}\right)}{\sqrt{\nu} B\left(\tfrac{1}{2},\tfrac{\nu}{2}\right)}, \\ G_\nu(x) &= -\left(\frac{\nu + x^2}{\nu - 1}\right) f_\nu(x), \\ H_\nu(x) &= \frac{1}{2} + \frac{1}{2}\, \mathrm{sgn}(x)\, I \left(\tfrac{x^2}{\nu + x^2}; \tfrac{1}{2}, \nu - \tfrac{1}{2}\right), \\ \bar{B}_\nu &= \left(\frac{2\sqrt{\nu}}{\nu - 1}\right)\frac{B\left(\tfrac{1}{2}, \nu - \tfrac{1}{2}\right)}{B\left(\tfrac{1}{2}, \tfrac{\nu}{2}\right)^2}. \end{align*}\]

Parameters:

Name Domain
"df" \(\nu > 0\) (for \(\nu \le 1\) the CRPS is undefined)
"location" \(\mu \in \mathbb{R}\)
"scale" \(\sigma > 0\)
"lower" \(l \in \mathbb{R}\), \(l < u\)
"upper" \(u \in \mathbb{R}\), \(l < u\)
"lmass" \(L \ge 0\), \(L + U < 1\)
"umass" \(U \ge 0\), \(L + U < 1\)

Mathematical functions:

Symbol Name
\({}_2F_1\) hypergeometric function
\(B\) beta function
\(I\) regularized incomplete beta function
\(\mathrm{sgn}\) sign function

See also: Student’s \(t\)-distribution

4.8.1 Censored \(t\)-distribution ("ct")

CRPS formula: \[\begin{align*} \mathrm{CRPS}\left(F_{l, \nu}^{u}, y\right) &= |y - z| + uF_\nu(-u)^2 - lF_\nu(l)^2 \\ & \quad + z(2F_\nu(z) - 1) \\ & \quad - 2G_\nu(z) + 2G_\nu(u)F_\nu(-u) + 2G_\nu(l)F_\nu(l) \\ & \quad - \bar{B}_\nu \left(H_\nu(u) - H_\nu(l) \right), \\ \text{where}\quad z&= \begin{cases} l, & y < l, \\ y, & l \le y < u, \\ u, & y \ge u. \end{cases} \end{align*}\] Censored Student’s \(t\)-distribution: \[\begin{align*} F_{l, \nu}^{u}(x) &= \begin{cases} 0, & x < l, \\ F_\nu(x), & l \leq x < u, \\ 1, & x \geq u, \end{cases} \end{align*}\]

where all other symbols are as given in Generalized truncated/censored Student’s \(t\)-distribution.

4.8.2 Truncated \(t\)-distribution ("tt")

CRPS formula: \[\begin{align*} \mathrm{CRPS}\left(F_{l, \nu}^{u}, y\right) &= |y - z| \\ & \quad + \left(\frac{1}{F_\nu(u) - F_\nu(l)}\right) z\left(2F_\nu(z) - F_\nu(u) - F_\nu(l)\right) \\ & \quad - \left(\frac{1}{F_\nu(u) - F_\nu(l)}\right) 2G_\nu(z) \\ & \quad - \left(\frac{1}{F_\nu(u) - F_\nu(l)}\right)^2 \bar{B}_\nu \left(H_\nu(u) - H_\nu(l)\right), \\ \text{where}\quad z&= \begin{cases} l, & y < l, \\ y, & l \le y < u, \\ u, & y \ge u. \end{cases} \end{align*}\] Truncated Student’s \(t\)-distribution: \[\begin{align*} F_{l, \nu}^{u}(x) &= \begin{cases} 0, & x < l, \\ \frac{F_\nu(x) - F_\nu(l)}{F_\nu(u) - F_\nu(l)}, & l \leq x < u, \\ 1, & x \geq u, \end{cases} \end{align*}\]

where all other symbols are as given in Generalized truncated/censored Student’s \(t\)-distribution.

5 Distributions for discrete variables

5.1 Negative binomial distribution ("nbinom")

CRPS formula: \[\begin{align*} \mathrm{CRPS}(F_{n, p}, y) &= y\left(2F_{n, p}(y) - 1\right) \\ &\quad - \frac{n(1-p)}{p^2}\left(p\left(2F_{n+1,p}(y-1) - 1\right) + {\ }_2F_1\left(n+1, \tfrac{1}{2}; 2; -\tfrac{4(1-p)}{p^2}\right)\right) \end{align*}\] Negative binomial distribution: \[\begin{align*} F_{n, p}(x) &= \begin{cases} I\left(p; n, \lfloor x+1\rfloor\right), & x \geq 0, \\ 0, & x < 0, \end{cases} \\ f_{n, p}(x) &= \begin{cases}\frac{\Gamma(x+n)}{\Gamma(n) x!} p^n (1-p)^x, & x = 0, 1, 2, \ldots, \\ 0, & \text{otherwise}. \end{cases} \end{align*}\]

Parameters:

Name Domain
"size" \(n > 0\)
"prob" \(p \in (0, 1]\)

Mathematical functions:

Symbol Name
\(\lfloor\cdot\rfloor\) floor function
\({}_2F_1\) hypergeometric function
\(I\) regularized incomplete beta function

Notes:

5.2 Poisson distribution ("pois")

CRPS formula: \[\begin{align*} \mathrm{CRPS}(F_\lambda, y) &= (y - \lambda) \left(2F_\lambda(y) - 1\right) + 2\lambda f_\lambda\left(\lfloor y\rfloor\right) - \lambda \exp(-2\lambda)\left(I_0(2\lambda) + I_1(2\lambda)\right) \end{align*}\] Poisson distribution: \[\begin{align*} F_\lambda(x) &= \begin{cases} \frac{\Gamma_u(\lfloor x+1\rfloor, \lambda)}{\Gamma(\lfloor x+1 \rfloor)}, & x \geq 0,\\ 0, & x < 0, \end{cases}\\ f_\lambda(x) &= \begin{cases}\frac{\lambda^x}{x!}e^{-\lambda}, & x = 0, 1, 2, \ldots, \\ 0, & \text{otherwise}, \end{cases} \end{align*}\]

Parameter:

Name Domain
"lambda" \(\lambda > 0\)

Mathematical functions:

Symbol Name
\(\lfloor \cdot \rfloor\) floor function
\(\Gamma\) gamma function
\(\Gamma_u\) upper incomplete gamma function
\(I_m\) modified Bessel function of the first kind

Notes:

References

Baran, S, and S Lerch. 2015. “Log-Normal Distribution Based EMOS Models for Probabilistic Wind Speed Forecasting.” Quarterly Journal of the Royal Meteorological Society 141: 2289–99.

Friederichs, P, and T L Thorarinsdottir. 2012. “Forecast Verification for Extreme Value Distributions with an Application to Probabilistic Peak Wind Prediction.” Environmetrics 23: 579–94.

Gneiting, T, and T L Thorarinsdottir. 2010. “Predicting Inflation: Professional Experts Versus No-Change Forecasts.” ArXiv Preprint ArXiv:1010.2318.

Gneiting, T, K Larson, K Westrick, M G Genton, and E Aldrich. 2006. “Calibrated Probabilistic Forecasting at the Stateline Wind Energy Center: The Regime-Switching Space–time Method.” Journal of the American Statistical Association 101: 968–79.

Gneiting, T, A E Raftery, A H Westveld III, and T Goldman. 2005. “Calibrated Probabilistic Forecasting Using Ensemble Model Output Statistics and Minimum CRPS Estimation.” Monthly Weather Review 133: 1098–1118.

Grimit, E P, T Gneiting, V J Berrocal, and N A Johnson. 2006. “The Continuous Ranked Probability Score for Circular Variables and Its Application to Mesoscale Forecast Ensemble Verification.” Quarterly Journal of the Royal Meteorological Society 132: 2925–42.

Möller, D, and M Scheuerer. 2015. “Probabilistic Wind Speed Forecasting on a Grid Based on Ensemble Model Output Statistics.” Annals of Applied Statistics 9: 1328–49.

Taillardat, M, O Mestre, M Zamo, and P Naveau. 2016. “Calibrated Ensemble Forecasts Using Quantile Regression Forests and Ensemble Model Output Statistics.” Monthly Weather Review 144: 2375–93.

Thorarinsdottir, T L, and T Gneiting. 2010. “Probabilistic Forecasts of Wind Speed: Ensemble Model Output Statistics by Using Heteroscedastic Censored Regression.” Journal of the Royal Statistical Society: Series A (Statistics in Society) 173: 371–88.

Wei, W, and L Held. 2014. “Calibration Tests for Count Data.” TEST 23: 787–805.