This paper investigates the issues of short-term forecasting of exchange rates using deep learning models, which is relevant for both the academia and traders and investors. The purpose of this study is to build a model to forecast the direction of changes in the price movement of currency quotes based on deep neural networks. The developed architecture is based on the gated recurrent unit model, which is a modification of the “long short-term memory” model, but is simpler as to the number of parameters and learning time. Forecast calculations of the dynamics of quotations of the euro/dollar currency pair and the most capitalised cryptocurrency – bitcoin/dollar are carried out using daily, four-hour, and hourly observations. The obtained results of binary classification (forecasting the direction of trend change) when applying daily and hourly quotes turned out to be generally better than those given by time series models, or models of neural networks of other architecture (namely multilayer perceptron, or models based on “long short-term memory”). According to the results obtained, the highest classification accuracy was found for the model of daily quotes for both EUR/USD – about 72%, and for BTC/USD – about 69%. When using the four-hour and hourly time series, the classification accuracy decreased, which can be explained both by an increase in the impact of “market noise” and by the likely retraining of models. As a result of computer experiments, it was found that models better predict an upward trend than a downward one. The conducted research confirmed the prospects of using deep learning models for short-term forecasting of time series of currency quotes. At the same time, the use of the developed models proved to be effective for both fiat and cryptocurrencies. The proposed system of models based on deep neural networks can be used as a basis for the development of an automated trading system in the foreign exchange market