On estimation and prediction procedures for AR(1) models with power transformation

The power transformation of Box and Cox has been shown to be quite useful in short-term forecasting for the linear regression model with AR(1) dependence structure. It is crucial to have good estimates of the power transformation and serial correlation parameters, because they form the basis for estimating other parameters and predicting future observations. The prediction of future observations is the main focus of this paper. The authors propose to estimate these two parameters by minimizing the mean squared prediction errors. These estimates and the corresponding predictions compare favourably, via revs and simulated data, with those obtained by the maximum likelihood method. Similar results are also demonstrated in the repeated measurements setting.