The use of recurrent neural networks to solve a regression type problem

Abstract

© 2019, Research Trend. All rights reserved. The development of the cryptocurrency market and its integration into the system of economic, operational, financial and other processes determine the need for a comprehensive study of this phenomenon of particular relevance to this is the fact that in recent months, discussions on the prospects for legalizing the cryptocurrency market and the possibilities of using its tools in the economic activity of economic agents have been intensified at the state level. Despite sometimes controversial views and approaches that have emerged at the moment among Russian experts regarding the solution of this issue, the development of the crypto market is independent of its state regulation. This causes and actualizes the conduct of scientific research in the field of studying the main parameters and prospects for the crypto market development, including through the use of mathematical analysis methods. The paper deals with the problem of predicting the trend of financial time series using the LSTM neural network. The time series compiled from the BTC/USD currency pair is analyzed, the timeframe is a day. The authors analyzed the neural network architecture, built a neural network model taking into account the heterogeneity and random volatility of the time series, developed and implemented an algorithm for solving the problem in the Python system. For training the neural network, data for the period from 24/09/2013 to 17/03/2019 (a total of 2002 data sets) were used. The experiment boils down to the fact that the constructed neural network model tries to determine the trend of the time series for one next timeframe. The training with a "teacher" was conducted. To determine the prediction error, the root-mean-square error (RMSE) was calculated. The research results are represented in tabular and graphical form.