Forecasting financial time series with LSTM recurrent neural networks

The paper deals with the possibility of forecasting the closing prices of a volatile financial instrument (Close) by means of the special architecture of recurrent neural networks (Long Short-Term Memory, LSTM). The set of data for the study is a sample of the time series of the Sberbank futures over a 2-year period in 5-minute intervals between observations. Based on the selected time series, sequences with a fixed offset window are formed, and the data used is normalized at [0 : 1] interval. In relation to the generated data, neural network models consisting of two recurrent layers as well as two aggregation layers in forward propagation are applied. At the end of training in the LSTM model, the predicted data and historical closing prices are compared. Their comparison demonstrates that the recurrent neural network model based on the LSTM architecture is able to predict the behavior of the instrument on the financial market.