Correlograms for non-stationary autoregressions
Analysis of time series often involves correlograms and partial correlograms as graphical descriptions of temporal dependence. Two methods are available for computing these statistics: one based on autocorrelations and the other on scaled autocovariances. For a stationary time series the resulting plots are nearly identical. When it comes to time series exhibiting non-stationary features these methods can lead to very different results. This has two consequences: incorrect inferences can be drawn when confusing these concepts; better discrimination between stationary and non-stationarity is achieved when using autocorrelations instead of, or along with, the autocovariances which are commonly used in statistical software.