Log-t or Log-Student t

Parameters	${\displaystyle {\hat {\mu }}}$ ( real), location parameter ${\displaystyle \displaystyle {\hat {\sigma }}>0\!}$ (real), scale parameter ${\displaystyle \nu }$ (real), degrees of freedom ( shape) parameter
Support	${\displaystyle \displaystyle x\in (0,+\infty )\!}$
PDF	${\displaystyle p(x\mid \nu ,{\hat {\mu }},{\hat {\sigma }})={\frac {\Gamma ({\frac {\nu +1}{2}})}{x\Gamma ({\frac {\nu }{2}}){\sqrt {\pi \nu }}{\hat {\sigma }}\,}}\left(1+{\frac {1}{\nu }}\left({\frac {\ln x-{\hat {\mu }}}{\hat {\sigma }}}\right)^{2}\right)^{-{\frac {\nu +1}{2}}}}$
Mean	infinite
Median	${\displaystyle e^{\hat {\mu }}\,}$
Variance	infinite
Skewness	does not exist
Excess kurtosis	does not exist
MGF	does not exist

In probability theory, a log-t distribution or log-Student t distribution is a probability distribution of a random variable whose logarithm is distributed in accordance with a Student's t-distribution. If X is a random variable with a Student's t-distribution, then Y = exp(X) has a log-t distribution; likewise, if Y has a log-t distribution, then X = log(Y) has a Student's t-distribution. ^[1]

Characterization

The log-t distribution has the probability density function:

${\displaystyle p(x\mid \nu ,{\hat {\mu }},{\hat {\sigma }})={\frac {\Gamma ({\frac {\nu +1}{2}})}{x\Gamma ({\frac {\nu }{2}}){\sqrt {\pi \nu }}{\hat {\sigma }}\,}}\left(1+{\frac {1}{\nu }}\left({\frac {\ln x-{\hat {\mu }}}{\hat {\sigma }}}\right)^{2}\right)^{-{\frac {\nu +1}{2}}}}$,

where ${\displaystyle {\hat {\mu }}}$ is the location parameter of the underlying (non-standardized) Student's t-distribution, ${\displaystyle {\hat {\sigma }}}$ is the scale parameter of the underlying (non-standardized) Student's t-distribution, and ${\displaystyle \nu }$ is the number of degrees of freedom of the underlying Student's t-distribution. ^[1] If ${\displaystyle {\hat {\mu }}=0}$ and ${\displaystyle {\hat {\sigma }}=1}$ then the underlying distribution is the standardized Student's t-distribution.

If ${\displaystyle \nu =1}$ then the distribution is a log-Cauchy distribution. ^[1] As ${\displaystyle \nu }$ approaches infinity, the distribution approaches a log-normal distribution. ^{[1] [2]} Although the log-normal distribution has finite moments, for any finite degrees of freedom, the mean and variance and all higher moments of the log-t distribution are infinite or do not exist. ^[1]

The log-t distribution is a special case of the generalized beta distribution of the second kind. ^{[1] [3] [4]} The log-t distribution is an example of a compound probability distribution between the lognormal distribution and inverse gamma distribution whereby the variance parameter of the lognormal distribution is a random variable distributed according to an inverse gamma distribution. ^{[3] [5]}

Applications

The log-t distribution has applications in finance. ^[3] For example, the distribution of stock market returns often shows fatter tails than a normal distribution, and thus tends to fit a Student's t-distribution better than a normal distribution. While the Black-Scholes model based on the log-normal distribution is often used to price stock options, option pricing formulas based on the log-t distribution can be a preferable alternative if the returns have fat tails. ^[6] The fact that the log-t distribution has infinite mean is a problem when using it to value options, but there are techniques to overcome that limitation, such as by truncating the probability density function at some arbitrary large value. ^{[6] [7] [8]}

The log-t distribution also has applications in hydrology and in analyzing data on cancer remission. ^{[1] [9]}

Multivariate log-t distribution

Analogous to the log-normal distribution, multivariate forms of the log-t distribution exist. In this case, the location parameter is replaced by a vector μ, the scale parameter is replaced by a matrix Σ. ^[1]

References