Cryptocurrency rate forecasting system based on methods of recurrent neural networks

Abstract

The object of the study is a sample of prices of cryptocurrency "Litecoin" for each day from 2017 to 2021. The subject of research is the methods of data mining based on recurrent neural networks and the method of GMDH. Python was chosen as the programming language. Litecoin cryptocurrency forecasting models are built on the basis of the latest modifications of recurrent neural networks LSTM and GRU, such as bidirectional LSTM and GRU, extended LSTM, dependent (dependent) Bi-RNN, these models with the addition of attention mechanism attention-based mechanism), machine learning methods, and the method of group argumentation (GMDH). A software product for predicting the value of cryptocurrencies has been developed. Training of models of recurrent neural networks is executed. Experiments were conducted to select the parameters of these models. Using a set of criteria, estimates of the quality of the constructed models and the obtained forecasts are obtained. The results obtained by GMDH and models of recurrent neural networks are compared. When performing the work, two methods were established that give the best results that are quite close to the real ones. The direction of development of work is in expansion of functionality, reduction of errors of forecasting of the price and time of training of model.