The usual assumption in the analysis of DPCM systems is that the predictor is, at least initially, matched to the source. In this paper we analyze the DPCM system without that assumption. Assuming that the source is an AR process of unknown order we obtain statistical descriptions of the prediction error residuals. Using these descriptions we obtain a simple expression for the Signal-to-Prediction Error Ratio (SPER) and the noiseless channel signal to noise ratio (SNR) for both the case where the predictor is matched to the source and the case where the predictor is not matched to the source. This formulation points out some interesting interactions between the predictor and the quantizer.