Non-stationary predictive deconvolution is necessary for inverse filtering with variable wavelet. In most cases we mean the variation of wavelet shape related with frequency-dependent attenuation. The procedure is also suitable for elimination of reverberation on the marine shallow data with variable interval between arrival times. This procedure is alternative to the multi-windows deconvolution, which limitations are well-known and related to the insufficient volume of selections for statistical estimations. The model of non-stationary, which describes the wide class of variable signal processes is used in non-stationary deconvolution approach.