From autocovariances to moving average: An algorithm comparison (Q1297872)

Let us consider the model \(y_{t}=\theta(B)a_{t}=(1+\theta_{1}B+\ldots+ \theta_{q}B^{q})a_{t}\), where \(B\) is a backshift operator; \(a_{t}\) is a white noise process with variance \(\sigma_{a}^{2}\). The autocovariances \(\gamma_{0},\gamma_{1},\ldots,\gamma_{q}\) can be obtained through the autocovariance generating function \(\gamma_{0}+\sum_{k=1}^{q} \gamma_{k}(B^{k}+B^{-k})=\sigma_{a}^{2}\theta(B)\theta(B^{-1})\). The authors compare the Wilson's Newton-Raphson procedure and the Robinson's polynomial root search method to solve the obtained system of nonlinear equations with respect to \(\theta_{1},\ldots,\theta_{q}\) and \(\sigma_{a}^{2}\). For each algorithm the authors measure the speed of convergence in term of mean computing time for convergence and measure the precision of the solutions with a root mean squared error criterion.