Linear prediction of ARMA processes with infinite variance (Q1059970)

For ARMA, AR, MA processes with infinite variance and noise having regular variation (regularly varying tails), the authors develop predictors based on the criterion of minimizing a naturally defined 'relative dispersion' of the error distribution (this coincides with minimizing the size of the error tails probabilities). It is proved that the best predictor exists, is unique and satisfies a recursive relation. For the stationary ARMA(1,1) process the predictor and the error dispersion are derived. For the ARMA(p,q), \(q>1\), process, the predictor determination involves infinite sums, requiring truncations. Some indications are given for this case to derive close to optimal predictors. As infinite variance ARMA processes have some technical applications, the paper can have sounding consequences.