Estimating tails of probability distributions (Q1101159): Difference between revisions

Let \(X_ 1,...,X_ n\) be a large random sample from cdf F(x), and let \(Y_ 1,...,Y_ N\) be the excesses over a threshold u, \(Y_ j=X_ i-u\) if \(X_ i\) is the j th exceedance. To estimate the upper tail of F(x), the distribution of the \(Y_ j\) is approximated by the generalized Pareto distribution \(G(y;\sigma,k)=1-(1-ky/\sigma)^{1/k}\), \(k\neq 0\), \(\sigma >0\). Estimating \(\sigma\) and k by ML-estimates \({\hat \sigma}_ N\), \(\hat k_ N\) based on \(Y_ 1,...,Y_ N\) and using N/n as estimator of 1-F(u), the proposed tail estimator of \(1-F(u+y)\) is \[ 1-\hat F(u+y) = n^{-1}N(1-\hat k_ Ny/{\hat \sigma}_ N)^{1/\hat k_ N}\text{ for }\begin{cases} 0<y<\infty &\text{ (if \(\hat k_ N<0)\)} \\ 0<y<{\hat \sigma}_ N/\hat k_ N &\text{ (if \(\hat k_ N>0)\).} \end{cases} \] A practical advantage of this approach is the avoidance of any need to decide a priori among the three limiting types of extreme order statistics. Nevertheless, the working out of the proposed method requires separate consideration of them. This is done in the main part of the paper. But beforehand, the basic idea underlying all the results is presented nicely. The choice of the threshold u is dealt with, too. Also, the proposed method is compared with other estimators.