SHORT-TERM FINANCIAL TIME SERIES ANALYSIS WITH LONG SHORT-TERM MEMORY NEURAL NETWORKS

The process of creating a long short-term memory neural network for high-frequency financial time series analyzing and forecasting is considered in this article. The purpose of the research is to carry out an instrument for modelling high-frequency financial time series, that could also be used as a decision support tool while trading in the stock market. The research base, presented by 1-minute stock index prices, is compiled in the beginning. Then the main groups of models, applied for analyzing time series as such, are described. Long short-term memory neural networks are highlighted in the group of artificial intelligence models, their advantages over models from other groups are stated. Further a set of rules, necessary for creating a long short-term memory neural network, is listed, the estimation of long short-term memory neural network parameters is carried out on the learning subsample and the test of its modelling quality is made on the testing subsample. In addition, the forecast of future returns signs is made for the horizon of 90 minutes with the estimated neural network. In conclusion, the trading strategy, which is combined of the estimated neural network and the automated trading system, is formulated. This strategy is utilized on returns series, which were not used while learning or testing phases. The analysis of financial results from trading completes the article. The corresponding finding are made. They concern the unequal results of data modelling, inverse relationship between index volatility and its modelling precision on the data from the test subsample, inverse relationship between index volatility and accuracy of forecasting its future values.

high-frequency financial data, automated trading system, artificial intelligence, long short-term memory neural network, stock market index