Non Stationary Series and ARMA model

[SnakeS2k]

Reading some papers as "Forecasting the term structure of government bond yields" by Diebold or "The Impact of the sovereign debt crisis on the activity of italian bank" by albertazzi and other, I noticed that they do autoregression on series that are not stationary. This sound a little bit strange to me.
For example Diebold, after estimation of Nelson and Siegel factors, models them with an AR(1), but parameters are not stable (autocorrelation is evident). So forecasting should not be accurate.

Could someone give to me an explanation.

thanks