Current seasonal forecasting systems are dependent on projections from general circulation models (GCMs) of varying complexity, designed to exploit predictability associated with memory and interactions between different elements of the climate system. These models demonstrate a degree of skill in predicting large-scale features and primary modes of variability in the climate system, but there remains significant room for improvement. Upgrades to these systems typically involve increasing model resolution, model complexity, or the size of the forecast ensemble. However, processing the raw model forecasts with computationally efficient machine learning techniques also has the potential to significantly increase forecast skill. Nonlinear machine learning methods such as neural networks and kernel methods using observed data as predictors have proven useful in forecasting the state of teleconnection patterns and extreme climate indices. By using GCM output to generate an additional set of predictors, the skill of GCM-based forecasts can be passed to the machine learning techniques. This approach is used here to forecast seasonal extreme (instead of the more common seasonal mean) climate indices for stations in the U.S. and Canada. Results are compared against raw GCM forecasts and observationally derived forecasts.