An approach to the nonstationary process analysis (Q5904017)

The author develops a procedure for estimating the spectral density of a nonstationary process. The procedure first divides the data into P blocks of equal length. Next an AR(M) model is fitted to each block with the restriction that the first difference between one-step-ahead predictors for adjacent blocks be \(N(0,\sigma^ 2/\eta^ 2)\). A normal prior is assigned to the AR parameter in blocks 2 to P and the expected likelihood given the AR parameters in block I is maximized. The ABIC criterion is then used to choose \(\eta\) and the order of autoregression in block 1. Finally, the spectrum is computed from this fitted model on a block by block basis. An algorithm and numerical examples are given.