The experimental results show that this new neural model is more accurate than SOM-based models and other neural networks for predicting non-stationary FX rates time series. They also show that the improved performance of NGMAR model is significant under the given criteria.