Dynamic neural networks have been shown as an encouraging alternative to traditional approaches for nonlinear temporal predictions. We use partially recurrent neural networks to study solar wind-magnetosphere coupling by predicting geomagnetic storms. The solar wind and Dst data used in this study are selected from the period 1963 to 1992. Statistical cross-correlation analyses and neural networks are applied to finding the best coupling functions. It is found that the results from both studies are consistent and that the coupling functions P1/3VBs, P1/2VBs, V2Bs, and VBz are well suited for predicting geomagnetic storms. Also, the relative importance of solar wind parameters is investigated in detail, aimed at finding the optimal combinations of solar wind parameters for predicting geomagnetic storms. The two basic combinations giving accurate prediction are Bs, n, V and Bz, n, V. Addition of B, F, or By can further improve predictions. We find that the best combination outperforms the best coupling functions in terms of prediction accuracy. Geomagnetic storms are very accurately predicted 1--2 hours in advance from the solar wind alone. The accurate predictions 1--2 hours ahead might well imply that the internal dynamic magnetospheric processes in forming geomagnetic storms occur on a timescale of about 1 hour. The predictions 3--5 hours ahead are useful in practical operation according to their acceptable accuracy. The prediction goodness is decreased with increasing prediction time. We consider this as a result of the fact that the internal magnetospheric dynamics limits the short-term prediction accuracy. The short-term predictability of geomagnetic storms and the validity of the results from cross-correlation analyses are discussed.¿ 1997 American Geophysical Union