Approaches for Modeling Fat Tails in Asset Returns
Summary
The document surveys approaches to representing heavy tails in financial return data and asks whether Brownian motion is adequate. One answer recommends discrete-time GARCH with standardized innovations drawn from a Student-t or another heavy-tailed distribution; the resulting returns can have tails heavier than those of the innovations alone. Another answer points to extreme value theory, including the generalized Pareto distribution for tail analysis, and mentions likelihood, moment-based, and probability-weighted estimation methods, along with tail-index estimators.
It also lists several ways to quantify or compare tail heaviness, including power-law tail indices and kurtosis. The material is a collection of suggestions rather than a comparative evaluation: it gives no fitted models or performance evidence establishing which method is best. Tail estimation can be sensitive to finite samples and modeling choices. The appropriate distribution and fitting approach depend on whether the data are discrete or continuous and on the purpose of the analysis.
Key ideas
- GARCH models can use standardized Student-t or other heavy-tailed innovations to represent return tails.
- Extreme value theory and generalized Pareto models focus on behavior in the extremes.
- Tail estimation can use likelihood, moments, probability-weighted moments, or tail-index methods.
- Different measures describe tail heaviness, and results depend on distributional assumptions and sample size.
Tags
Full text
# What are the most common methods to model fat tails in the changes of asset prices? # What are the most common methods to model fat tails in the changes of asset prices? I was wondering what the most common, or most popular, ways - in both academia, and industry - there were to model the fat tails of volatility in asset prices changes. I am presuming a basic Brownian motion random walk, is not what is used, because it will not replicate fat tails. Is that correct? Or am I wrong, and in most cases, a basic Brownian motion is "good enough"? What are more advanced methods that are used, whether it be in terms of stochastic calculus, statistical methods, etc.? ## Answer by Richard Hardy (score 2, accepted) https://quant.stackexchange.com/a/69085 For the case of discrete time, consider a GARCH model with standardized innovations that follow a Student-$t$ or another (somewhat) heavy-tailed distribution. The dependent variable will have a tail heavier than that due to the GARCH model. (The model generates heavier tails than present in the distribution assumed for the standardized innovations.) ## Answer by JeeyCi (score 1) https://quant.stackexchange.com/a/79831 see Extreme Value Theory for Risk Estimation coded to find VaR & C.I. for extreme value risk estimates of financial time series - example for stocks is given by link... and look through Regression Estimator for the Tail Index as non-parametric approach... and others at 19.Strategies for Modeling and Predicting Heavy-Tailed Data Generalized Pareto distribution is a very important distribution in the extreme value investigation > there is a set of Estimation procedures, such as the maximum likelihood (ML), the method of moments (MOM) and the probability weighted moments (PWM) method can see here implementation in R or scikit-extremes-package in Python or phat-tails-package or example of Hill estimator use... Though, data-driven tail index finding through "minimizing the asymptotic mse do not perform well in finite samples"... So, algorithmization can vary for fitting Pareto distribution with real data, and model-free approches perhaps can benefit compared with pareto modeling or change distribution estimation with some sorts of indexes to make analytical aims easier to achieve algorithmically. In general, there are 4 Ways to Quantify Fat Tails - Power Law Tail Index. ... - Kurtosis (i.e. non-Gaussianity) ... - Log-normal's σ ... - Taleb's κ Tail Comparison & survival_probability_plots can see here for BTC as part of FergM's work - seems to be interesting approach for analysing purposes of tails comparison! p.s. Pareto optimality (or multi-objective optimization) p.p.s. In general there is vast number of fat-tailed distributions & you should always make assumption about your distribution to consider either for Outliers or for Fat-tails in your analysis, as e.g. here- Identifying multiple outliers in heavy-tailed distributions with an application to market crashes -- remember the nature (either discrete or continuous) of your distribution & choose the appropriate one for fitting depending on aims of your analysis... p.p.p.s. extrapolation over the unit of the return period
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.