The aim of this paper is to show methodology of forecasting with various horizon of prediction using grey system theory, basing on practical application to vibration condition monitoring problems. The method of forecasting was based on GM(1,1) prognostic models with various window lengths for estimating model parameters. The model GM(1,1) is very effective where we have only few data, incomplete, and with low accuracy. The moving window method applied to GM(1,1) model enables to adapt to changes in data trend. However, selecting an inappropriate window length can result in excessive forecast errors. The applied algorithm is based on tracking the current prediction error for models having various window lengths, and then eliminating the models for which the error of prediction caused by the loss of adequacy of the model to the data increases excessively.