This document proposes using a deep neural network to predict currency exchange rates. It discusses using DNN to directly predict future exchange rates or to perform binary classification to predict if the rate will increase or decrease. The model takes in features like past exchange rates, moving averages and volatility indicators. Experiments show the model can predict trend transitions with over 75% accuracy on closed tests and over 60% on open tests by classifying trend direction changes. Pre-training is done using restricted Boltzmann machines to initialize weights before fine-tuning with backpropagation.