Artificial neural networks (NN) have been used to model solar wind-driven auroral electrojet dynamics, with emphasis on the reliable real-time forecasting of auroral electrojet activity (the AE index) from solar wind input. Practical limitations of the NN-based models used earlier are clarified. These include the inability to accurately predict large-amplitude substorm events, which is the most important feature for many applications. A novel technique for improving predictions is suggested based on application-specific threshold mapping and symbolic encoding of the AE index. This approach allows us to disregard relatively unimportant details of small-amplitude perturbations and effectively improve forecasting of large-amplitude events. Results from our new model imply that application-oriented optimization of the real-time substorm forecasting system can be an important factor in the overall improvement of the prediction accuracy. ¿ 2001 American Geophysical Union