In this work, a robust and accurate neural predictive model based on a randomized neural learning scheme is developed for foreign exchange market modelling and forecasting purpose. In our predictive model, a dynamic single-hidden layer feedforward neural network (SLFN) is constructed with tapped-delay-memories applied at its input layer. A modified sigmoid function is designed and input weights and hidden biases are randomly assigned in such a way that highly coupled financial input patterns can be represented in the hidden feature space in a clearer way and sensitivities of the network's hidden outputs to the changes in the financial input signals are enhanced. Also, a large number of hidden nodes in the hidden layer is used to improve the clarity of input patterns' representation in the hidden feature space. Output weights of the network are optimized using regularised batch-learning type of least square method to improve robustness of the predictive model against external and internal disturbances. Simulation results show excellent performance of the developed model in both target deviation and directional performance measurements.