Neural networks are developed for reconstructing the chaotic attractor in the nonlinear dynamics of the solar wind driven, coupled magnetosphere-ionosphere (MI) system. Two new methods which improve predictive ability are considered: a gating method which accounts for different levels of activity and a preconditioning algorithm which allows the network to ignore very short time fluctuations during training. The two networks are constructed using the Bargatze et al. [1985> substorm database that contains solar wind speed and interplanetary magnetic field (IMF) along with ionospheric electrojet index, AL. Both networks are found to produce improvements in predictability, and the significance of the performance increase of the gated network is demonstrated using the bootstrap model testing method. ¿ 1999 American Geophysical Union