This paper introduces a new approach to the modelling of a stationary long run component, which is an autoregressive process with near unit root and small sigma innovation. We show that a combination of a noise and a long run component can explain the long run predictability puzzle pointed out...