We propose a new model with three parameters called bimodal Gumbel (BG) as a generalization of the Gumbel distribution. The advantage of our model in comparison to other generalizations of the Gumbel distribution is the number of parameters and the fact that it can be used to model extreme data with one or two modes.
You can install the released version of bgumbel from CRAN with:
or using devtools
dbgumbel(x = 0, mu = -2, sigma = 1, delta = -1)
curve(dbgumbel(x, mu = -2, sigma = 1, delta = -1), xlim = c(-5, 10))
integrate(dbgumbel, mu = -2, sigma = 1, delta = -1, lower = -5, upper = 0)
pbgumbel(0, mu = -2, sigma = 1, delta = -1)
integrate(dbgumbel, mu = -2, sigma = 1, delta = -1, lower = -Inf, upper = 0)
pbgumbel(0, mu = -2, sigma = 1, delta = -1, lower.tail = FALSE)
curve(pbgumbel(x, mu = -2, sigma = 1, delta = -1), xlim = c(-5, 10))
It is recommended to set up a pbgumbel graph to see the starting and ending range of the desired quantile.
curve(pbgumbel(x, mu = -2, sigma = 1, delta = -1), xlim = c(-5, 5))
(value <- qbgumbel(.25, mu = -2, sigma = 1, delta = -1, initial = -4, final = -2))
pbgumbel(value, mu = -2, sigma = 1, delta = -1)
x <- rbgumbel(100000, mu = -2, sigma = 1, delta = -1)
hist(x, probability = TRUE)
curve(dbgumbel(x, mu = -2, sigma = 1, delta = -1), add = TRUE, col = 'blue')
lines(density(x), col = 'red')
(EX <- m1bgumbel(mu = -2, sigma = 1, delta = -1))
x <- rbgumbel(100000, mu = -2, sigma = 1, delta = -1)
mean(x)
abs(EX - mean(x))/abs(EX) # relative error
# grid 1
mu <- seq(-5, 5, length.out = 100)
delta <- seq(-5, 5, length.out = 100)
z <- outer(X <- mu, Y <- delta, FUN = function(x, y) m1bgumbel(mu = x, sigma = 1, delta = y))
persp(x = mu, y = delta, z = z, theta = -60, ticktype = 'detailed')
# grid 2
mu <- seq(-5, 5, length.out = 100)
delta <- seq(-5, 5, length.out = 100)
sigmas <- seq(.1, 10, length.out = 20)
for (sigma in sigmas) {
z <- outer(X <- mu, Y <- delta, FUN = function(x, y) m1bgumbel(mu = x, sigma = sigma, delta = y))
persp(x = mu, y = delta, z = z, theta = -60, zlab = 'E(X)')
Sys.sleep(.5)
}
(EX2 <- m2bgumbel(mu = -2, sigma = 1, delta = -1))
x <- rbgumbel(100000, mu = -2, sigma = 1, delta = -1)
mean(x^2)
abs(EX2 - mean(x))/abs(EX2) # relative error
# Variance
EX <- m1bgumbel(mu = -2, sigma = 1, delta = -1)
EX2 - EX^2
var(x)
abs(EX2 - EX^2 - var(x))/abs(EX2 - EX^2) # relative error
# grid 1
mu <- seq(-5, 5, length.out = 100)
delta <- seq(-5, 5, length.out = 100)
z <- outer(X <- mu, Y <- delta, FUN = function(x, y) m2bgumbel(mu = x, sigma = 1, delta = y))
persp(x = mu, y = delta, z = z, theta = -30, ticktype = 'detailed')
# grid 2
mu <- seq(-5, 5, length.out = 100)
delta <- seq(-5, 5, length.out = 100)
sigmas <- seq(.1, 10, length.out = 20)
for (sigma in sigmas) {
z <- outer(X <- mu, Y <- delta, FUN = function(x, y) m2bgumbel(mu = x, sigma = sigma, delta = y))
persp(x = mu, y = delta, z = z, theta = -45, zlab = 'E(X^2)')
Sys.sleep(.5)
}
# Let's generate some values
set.seed(123)
x <- rbgumbel(1000, mu = -2, sigma = 1, delta = -1)
# Look for these references in the figure:
hist(x, probability = TRUE)
lines(density(x), col = 'blue')
abline(v = c(-2.5, -.5), col = 'red')
text(x = c(c(-2.5, -.5)), y = c(.05, .05), c('mu\nnear here', 'delta\nnear here'))
# If argument auto = FALSE
fit <- mlebgumbel(
data = x,
# try some values near the region. Format: theta = c(mu, sigma, delta)
theta = c(-3, 2, -2),
auto = FALSE
)
print(fit)
# If argument auto = TRUE
fit <- mlebgumbel(
data = x,
auto = TRUE
)
print(fit)
# Kolmogorov-Smirnov Tests
mu.sigma.delta <- fit$estimate$estimate
ks.test(
x,
y = 'pbgumbel',
mu = mu.sigma.delta[[1]],
sigma = mu.sigma.delta[[2]],
delta = mu.sigma.delta[[3]]
)
Please, send to: https://github.com/pcbrom/bgumbel/issues