The winter Arctic Oscillation (AO) is important for understanding the Northern Hemisphere climate variability and predictability. However, ENSEMBLES models produce inconsistent predictions when applied to the interannual variability of the 1962–2006 winter AO. In this study, the interannual increment of the winter AO index (DY_AOI) during 1962–2006 is first improved by a dynamical‐statistical model with two predictors: the preceding autumn Arctic sea ice and the concurrent winter ENSEMBLES‐predicted sea surface temperature over the North Pacific. Next, the improved final AOI is obtained by adding the improved DY_AOI to the preceding observed AOI. Because the interannual increment approach can amplify prediction signals and takes advantage from the previous observed AOI, this method shows promise for significantly improving the interannual variability prediction capabilities of the winter AO during 1962–2006 in the ENSEMBLES models. Therefore, this study offers important insights for AO predictions, even other climate variables predictions.