•A hybrid model for predicting exchange rate return volatility is proposed.•The model is an artificial neural network – Glosten, Jagannathan, and Runkle model.•The model improves forecasting accuracy over applied benchmark models by 90%.•Accurate forecasting with respect to time horizon and model configuration is achieved.•Commodity price series as input financial variables can improve model performance. The study examines the integration of an asymmetric Glosten, Jagannathan, and Runkle (GJR) model into an artificial neural network (ANN) comprising of a NARX (Nonlinear Autoregressive model with eXogenous inputs) augmented by a NAR network. The empirical results obtained from the study of five (5) major currency pairings reveal that, compared to the benchmark generalized autoregressive conditional heteroskedasticity (GARCH), GJR and Asymmetric Power Generalised Autoregressive Conditional Heteroskedasticity (APGARCH) models, the proposed hybrid ANN-GJR model provides an overall substantial improvement in exchange rate volatility forecasting precision by effectively capturing asymmetric volatility as well as volatility clusters. In terms of the MSE, the MAD and the MAPE measures, a significant improvement of the measured forecast accuracy was found when using the ANN hybrid model compared to the benchmark models. Applying measurements from a heteroscedasticity-adjusted mean squared error (HMSE) model, the hybrid ANN-GJR model can significantly reduce the error associated with using only the GJR model by 90%. The study also investigated the effect of incorporating commodity prices series of oil and gold as additional input variables. It was found that depending on the specific currency pair understudy, the inclusion of commodity price series potentially improves model performance over different forecast horizons.