Applying Deep Learning Algorithms to Predict Asset Prices

Most work on time series analysis is based on the use of deterministic linear models. This article discusses the possibility of effective use of modern machine learning methods in the analysis of financial time series and compares their results with linear models. The paper implements an algorithm that predicts an asset price change based on time-series variables and non-linear deep learning methods. Models of deep learning, including classical and recurrent neural networks, serve as a prediction algorithm. The subject of the simulation is the data on the stock prices of the Russian and American securities markets, as well as the main tools of the cryptocurrency market. This problem is a regression problem, and the measure in it is a measure of deviation from the true values. Despite the low coefficient of determination, some models of deep learning show a predictive ability significantly higher than the results of linear algorithms and different from a random one, based on which we can conclude about the big potential of deep learning in analyzing time series.