Extreme value distribution of a recursive-type detector in linear model (Q2271708)

A sequential CUSUM-type algorithm is considered for the detection of a change in the linear regression coefficients \(\beta_i\) by the observations of a sequence \(y_i=x_i^T\beta_i+\varepsilon_i\), \(i=1,2,\dots, m, m+1,\dots\) The algorithm detects the change at the moment \(\tau_m=\inf_{k\geq 1}\{| \tilde Q(m,k)| >\sqrt{m}h(k/m)\}\), where \(\tilde Q(m,k)=\tilde\varepsilon_{m+1}+\dots+\tilde\varepsilon_{m+k}\), \(\tilde\varepsilon_{m+k}\) are the least squares residuals of the model, and \(h\) is some fixed function, e.g., \(h_\gamma(t)=t^\gamma(1+t)^{1-\gamma}\), \(0\leq \gamma<1/2\). An extreme value asymptotics is derived for \(| \tilde Q(m,k)| \). The case is considered when \(\varepsilon_i\) is an augmented GARCH(1,1) process. The size and power of the procedure are investigated via simulations.