The importance of past MJO activity for empirical predictions of midlatitude weather

The Madden Julian oscillation (MJO) is one of the most important sources of predictability on subseasonal to seasonal (S2S) timescales. While our previous work (and that of others) has demonstrated the utility of using empirical models based on today’s MJO to predict weather over the United States multiple weeks in advance, it is still unclear whether the accumulated influence of past MJO activity can improve these predictions. Here, we demonstrate the importance of past MJO activity in determining future midlatitude weather using neural networks of varying complexity. We show that the past 15 days of MJO information plays an important role in determining the state of MJO teleconnections and weather extremes 15 days into the future. That is, the skill in midlatitude predictions increases significantly with these more complex empirical models. Furthermore, we show that the past 15 days of additional information is only important for some MJO phases, and not others, and the physical basis for this result is discussed.