< List of probability distributions
Stable distributions are a general family of probability distributions that share certain properties. They were first described by Paul Lévy (1925) and so are also sometimes informally called Lévy distributions. However, this can cause confusion as a “Lévy Distribution” is a specific member of the Stable Distribution family.
Most of these distributions do not have a distinct probability density function, with the exception of the Cauchy Distribution, Lévy Distribution and Normal distribution) but they do share certain properties, like skewness and heavy tails.
Properties of Stable Distributions
There are three main types of stable distributions: the Cauchy, Lévy and Normal distributions. Each type has its own unique probability density function (PDF), which gives us insight into how likely certain events are to occur given certain sets of data.
The Cauchy distribution is especially interesting due to its heavy-tailed nature and its ability to model fat-tailed phenomena such as stock market crashes or extreme weather events. The Lévy distribution is also known for its ability to model large-scale phenomena such as global sea surface temperatures or large financial transactions. Finally, the Normal distribution is probably the most well-known type of stable distribution due to its widespread use in statistical analysis and forecasting models.
The general stable distribution has four parameters (Barndorff-Nielsen et. al):
- Index of stability: α ∈ (0,2). This parameter determines the probability in the extreme tails (i.e., it tells you something about the height of the tails). A normal distribution has α = 2. Distributions below that number (i.e. with 0 < α ≤ 2) will be more tail heavy. The Cauchy distribution has α = 1.
- Skewness parameter: β. If β = 0, then the distribution is a symmetrical distribution. A normal distribution has β = 0.
- Scale parameter:γ. A measure of dispersion. For the normal distribution, γ = half of the population variance. For other symmetric distributions in the family, one suggestion is to exclude the top and bottom 28% of observations (Fama and Roll).
- Location parameter: δ. This parameter equals the median. When α > 1, it also equals the mean. For a normal distribution, the sample mean can be used as an estimate for δ. For other distributions, it may be necessary to discard extreme values in order to get a good estimate for δ. Depending on how heavy the tails are, you may need to exclude the first and last quartiles — only using the central half of observations.
One of the most important mathematical characteristics of Stable Distributions is that the distributions retain the same α and β under convolution of random variables (a calculus term which describes what happens when two functions f and g are combined to form a third function).
Advantages and Disadvantages for Practical Use
Stable distributions play an important role in probability calculations because they can help us understand the underlying phenomenon behind various events or outcomes. For example, if we know the PDFs associated with each type of stable distribution, then we can better predict how likely certain outcomes are given different sets of data – something that would otherwise be impossible without this knowledge! Furthermore, understanding how these PDFs interact with one another can help us make more accurate predictions about future events or outcomes based on past data sets. This could be invaluable information for businesses looking to plan ahead for future scenarios or investors trying to make sound financial decisions based on past trends.
Another very useful property of stable distributions is that they are scalable (up to a certain factor) — a small part of the distribution looks just like the whole.
A major drawback to the use of stable distributions is that any moment greater than α isn’t defined. That means in general, any theory based on variance (the second moment) isn’t useful. However, it’s sometimes possible to modify the distributions (for example, truncate the tails). This does require you to have a good grasp of the particular distribution you’re dealing with as well as the discipline from which your data was drawn in the first place. For example, Paul & Baschnagel describe how a Lévy distribution can be used to model a human heart beat if the tails are truncated — but you would only know to truncate the tails if you were aware that deviations in heartbeats can’t be arbitrarily large.
Fama, E.F. and Roll, R. (1968). Some Properties of Symmetric Stable Distributions.
Paul Lévy (1925). Calcul des probabilities.
Ole E. Barndorff-Nielsen, Thomas Mikosch, Sidney I. Resnick (2001). Lévy Processes: Theory and Applications
Jitendar S. Mann, Richard G. Heifner, United States. Dept. of Agriculture. Economic Research Service. The distribution of shortrun commodity price movements, Issues 1535-1538. U.S. Dept. of Agriculture, Economic Research Service, 1976 – Business & Economics – 68 pages. Free ebook.
Wolfgang Paul, Jörg Baschnagel. Stochastic Processes: From Physics to Finance.