← Back to RiskAlgo Research

Zhao's q-distribution

A probability distribution with asymmetric and heavy tail property

In probability theory, there are many symmetric real valued continuous distributions. Gaussian distribution is the most popular example of such distributions. For modeling asymmetric data, the model choice of the distributions are very limited. In this article, we introduce Zhao's q-distribution [1], which is an extension of Gaussian distribution with heavy tails and asymmetry.

This distribution is part of the research behind RiskAlgo's approach to modeling volatility skew and smile — see the math behind RiskAlgo for how it connects to the app's own pricing.

Zhao's q-distribution is an asymmetric and heavy-tailed probability distribution of a real-valued random variable [1]. It was published in 2024 in finance literature for modeling financial asset returns, which usually have heavier tails than the Gaussian distribution and also exhibit asymmetry.

The q-distribution can be represented by four parameters explicitly with a closed form representation. These parameters represent the location, scaling parameters of positive and negative deviations and heaviness of tails. When one or two parameters reach their boundary, e.g. zero or infinity, the limiting distributions can be identical to one of the classical distributions. The Gaussian, Laplace, logistic, and exponential are among the known special cases.[2]

Definition

Open figure (PDF)
Density function of a symmetric q-distribution versus Gaussian distribution
Open figure (PDF)
Density of an asymmetric q-distribution versus Gaussian distribution

The q-distribution has a probability density function[1]

\[ f(x) = \frac{1}{C(\alpha,\beta,s)} \left[ \left( 1 + e^{\frac{x-\mu}{\alpha s}} \right) \left( 1 + e^{-\frac{x-\mu}{\beta s}} \right) \right]^{-s} \]

for \(x\in \mathbb{R}\), where \(C(\alpha,\beta,s)\) is a normalizing constant that often needs to be computed numerically except for the symmetric cases when \(s\) is an integer.

The family of q-distributions has four parameters:

\(\alpha\), \(\beta\), and \(s\) must be positive.

Symmetric case

The q-distribution is symmetric when the scaling parameters \(\alpha\) and \(\beta\) are identical. Let \(\sigma=\alpha s\). The probability density function is simplified to

\[ f(x) = \frac{1}{\sigma B(s,s)}\left[\left(1 + e^{\frac{x-\mu}{\sigma}}\right) \left(1 + e^{-\frac{x-\mu}{\sigma}}\right)\right]^{-s}, \]

where the normalizing constant can be written as the Beta function. When the shape parameter \(s\) is an integer, the normalizing constant can be easily calculated. For instance, when \(s = 1\) and \(B(1,1) = 1\), one gets the logistic distribution. The symmetric q-distribution can be directly derived from Beta distribution[2] through variable exchange with a logistic function.

Other limiting cases

A few classical distributions can be derived from the q-distribution. These are the boundary cases when one or more of the parameters reaches zero or infinity.

Properties

The q-distribution has the following known properties:

Computational method

With a data sample \(\{x_1,x_2,\dots, x_n\}\), the parameters of q-distribution can be estimated using the maximum likelihood method.

In general, for asymmetric q-distribution, there is no closed-form expression for the cumulative distribution function or the compensator \(C(\alpha,\beta,s)\). However, expectation related to q-distribution can be computed effectively using a Gaussian quadrature approach since the distribution has exponential tails.

For any smooth function, the following integral can be effectively computed using Gauss–Laguerre quadrature[5] method:

\[ \int_0^\infty e^{-x} G(x) \, \mathrm{d}x = \sum_{i=1}^k w_i G(x_i) \]

where \(\{x_i, w_i\}\) are the quadrature abscissas and weights.

Any expectation with respect to q-density can be written into the above format via variable exchanges.[1] This makes it convenient to compute the compensator \(C(\alpha,\beta,s)\) as well as the cumulative distribution function and any other types of expected values.

Applications

For applications where a Gaussian distribution is used, but is not perfect because the data set presents either asymmetry or long tails, the q-distribution can jump in as a helper.

The four parameters of the q-distribution have the flexibility to match moments up to the fourth order instead of the second order as with the Gaussian. The distribution has potential applications in option pricing for modeling volatility skew and smile [1] , risk management in projecting asymmetric tail risk, and other scientific fields where data under study exhibits the behavior of asymmetry and heavy tails.

See also

The following classical distributions are related to the q-distribution.

References

  1. Zhao, Q. (2024). "A Leptokurtic Distribution Explains Volatility Skew and Smile". Journal of Derivatives. 32: 81–108. doi:10.3905/jod.2024.1.211.
  2. Johnson, N. L. and Kotz, S. (1970). Continuous Univariate Distributions I–II. John Wiley and Sons.
  3. Stukel, T. (1988). "Generalized logistic models". Journal of the American Statistical Association. 83: 426–431.
  4. Kotz, S.; Kozubowski, T. J.; Podgorski, K. (2002). "Maximum likelihood estimation of asymmetric Laplace parameters". Annals of the Institute of Statistical Mathematics. 54: 816–826.
  5. Press, W. H.; Teukolsky, S. A.; Vetterling, W. T.; Flannery, B. P. (1992). Numerical Recipes in C. Cambridge University Press.