On statistical inference based on record values (Q1297900)

The shape parameter \(\alpha\) of the distribution function \(F(x)=1-x^{- \alpha}K(x)\) is estimated using the record values \(Y_1\), \(Y_2\),\dots, \(Y_N\) and record times \(L(1)\), \(L(2)\),\dots, \(L(N)\) of a sequence of i.i.d. random values. Here \(K(x)\) is a function with \[ \log K(x)=d[1+d_1(\log x)^{-\alpha_1}+ O((\log x)^{-\alpha_1})] \text{as} x\to\infty\;\text{with} \alpha,\alpha_1,d>0. \] Two cases are considered: (a) the records are observed for a data set of given length \(n\), so \(N\) is a random variable; (b) \(N\) is a predetermined non-random value. The proposed estimator \((\hat\alpha,\hat d)\) is a solution of the equations system: \[ \sum_{i=1}^N\{L(i+1)-L(i)-1\}(1-dY_i^{-\alpha})^{-1}=L(N+1)-2, \] \[ (N-1)\alpha^{-1}+\sum_{i=1}^N\{L(i+1)-L(i)-1\}(1-dY_i^{-\alpha})^{-1}\log Y_i= \] \[ =\sum_{i=1}^N\{L(i+1)-L(i)\}\log Y_i-\log Y_1. \] (Such choice of \((\hat\alpha,\hat d)\) is provided by the MLE for the parametric model \(F(x)=1-dx^{-\alpha}\) for \(x>x_0\)). The asymptotic behavior of the estimators is investigated as \(N\to\infty\). In the case of non-random \(N\): \[ \hat\alpha-\alpha=b_\alpha N^{-(\alpha_1+1)}+12^{1/2}\alpha N^{-3/2}\zeta_1+ o_p(N^{-(\alpha_1+1)}+N^{-3/2}), \] \[ \hat d-d=b_d N^{-\alpha_1}+2dN^{-1/2}\zeta_2+o_p(N^{-\alpha}+N^{-1/2}), \] where \((\zeta_1,\zeta_2)\) is a bivariate normal random variable with zero mean, unit variance and correlation coefficient \(3^{1/2}/2\), \[ b_\alpha=-6\alpha^{1- \alpha_1}d_1(1-\alpha_1)^{-1},\;b_d=-4\alpha^{1-\alpha_1}dd_1(1-\alpha_1)^{-1} \text{if} 0<\alpha_1<1/2, \] and \(b_\alpha=b_d=0\) otherwise. Results of a simulation are presented.