Choosing Return Distributions for Bulk Fit and Tail Risk

Summary

The document raises a model-selection question for financial log returns: should analysts identify a distribution that fits the bulk of observations before separately modeling extreme tails? It notes a claim that empirical cumulative distributions, compared with candidate cumulative distribution functions, can favor Laplace or logistic fits over stable, normal, or Student’s t distributions, including after ARMA/GARCH filtering. The text presents this as an observation, without supplying data, estimation details, or formal comparisons.

It challenges the emphasis on Student’s t distributions in tail-risk work and mentions the Cauchy and stable distributions as alternatives for heavy tails. It also flags infinite variance as a concern for some such choices. The document does not resolve when any distribution is preferable, explain the constraints behind common practice, or offer a fitting procedure. Its contribution is to frame the distinction between fitting typical returns and representing tail behavior, while leaving empirical validation and risk-model suitability open.

Key ideas

• Distribution choice for returns may differ between fitting the central observations and modeling extremes.
• The document reports a claim favoring Laplace and logistic fits, but gives no supporting dataset or fitting details.
• Student’s t, Cauchy, and stable distributions are raised as heavy-tail modeling choices.
• Some heavy-tailed distributions can have infinite variance, which complicates their use.
• The text poses the model-selection question without establishing a preferred distribution.

Tags

# Fitting Student t-distributions to log-returns


# Fitting Student t-distributions to log-returns












It seems that some tail-risk centric groups are bent on using Paretian and t-distributions to account for tail risk when fitting log-returns. It has been observed, however, that with and without filtering log-returns with ARMA/GARCH, fitting ecdf's to cdf's still results in better Laplace and logistic distribution fits when compared with stable, normal, or Student's t. Given this, is there a reason why the bulk of the data and its best fitting distribution is not identified first, followed by consideration of tails? If the focus is mostly tails, then the Cauchy is not bad, and a Cauchy with heavy tails could be realized by a stable distribution. Independent of the infinite variance problem, why the constraint for mostly t-distributions?

Shown in full with attribution under the source's licence. Licence: CC BY-SA 4.0 (Stack Exchange)

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.