A practical model order selection criterion for autoregressive processes is presented in this paper. The criterion was developed based on the normalised error between sample data and data generated by the model. The method is an excellent indicator of how well an autoregressive model fits the available data and is therefore a useful tool when selecting the optimum model order needed to accurately and efficiently model the underlying process of the sampled data. The procedure is different from the existing criteria that are in common use such as the Akaike information criterion, the minimum descriptive length and Hannan's criterion. The paper shows that the criterion developed here performs well over a broad range of data sample lengths and signal-to-noise ratios.