Forecasting extreme weather events on sub-seasonal timescales (10 days to 3 months) is challenging due to the complex nature of atmospheric phenomena. This paper explores the use of artificial neural networks (ANNs) – specifically, a feed-forward neural network (FFNN) and an autoregressive recurrent neural network (AR-RNN) – to predict the Madden-Julian Oscillation (MJO), a dominant mode of tropical atmospheric variability. The ANNs achieved a competitive MJO prediction skill of approximately 26-27 days on average, comparable to state-of-the-art climate models. However, for specific initial phases and seasons, the prediction skill extended to 60 days or more. While the ANNs accurately predicted the MJO phase, they underestimated its amplitude.
