Estimating the stable index \(\alpha\) in order to measure tail thickness: a critique (Q799047)
From MaRDI portal
scientific article
Language | Label | Description | Also known as |
---|---|---|---|
English | Estimating the stable index \(\alpha\) in order to measure tail thickness: a critique |
scientific article |
Statements
Estimating the stable index \(\alpha\) in order to measure tail thickness: a critique (English)
0 references
1983
0 references
Stable laws are often used as models for long tailed distributions and/or outlier prone models. The paper discusses two data sets D-6 and D-24 with estimates of the index \(\alpha\) as \(1.37\pm .11\) and \(1.23\pm .08\), respectively, which indicates non-normality. The \(\beta\) values are 3.1 and 23.2, respectively. The value 3.1 is not significant at 95\% level but 23.2 is very highly significant. The simple rejection rule to reject normality \((\alpha =2)\) if \({\hat\alpha }<2\) will have both error probabilities going to zero as \(n\to\infty \). It is conjectured that \(P({\hat\alpha }<2|\alpha =2)\simeq K/\log n.\) The paper emphasises that rather than just screening for outliers one should attempt to describe the tail behaviour of the data. In this respect it is recommended that the data outside (10,90) percentiles should be fitted by a generalized Pareto distribution. Although the test based on MLE \({\hat\alpha }\) is very appealing at the first glance the author warns that the MLE would give a very biased measure of the tail behaviour if the true distribution is not a stable one. The paper raises a lot more interesting problems both of theoretical as well as applied nature and perhaps would attract many researchers to work in this area.
0 references
infinite variance
0 references
stable index
0 references
Stable laws
0 references
long tailed distributions
0 references
outlier prone models
0 references
non-normality
0 references
error probabilities
0 references
tail behaviour
0 references
generalized Pareto distribution
0 references