Economic series prediction basing on internet users sentiment analysis

Abstract

A set of economic time series forecasting models (based on objective and subjective internet content analysis) is proposed in the article. The Bitcoin (BTC) system, the first and most popular cryptocurrency today, was chosen for the analysis. Exponential increase of the cryptocurrency market stipulates relevance of BTC rate forecasting in recent years. Theoretical framework of the research is based on the behavioral finance concept supposing that traders’ behavior is irrational, and the character of their decisions largely depends on psychological factors. The aim of the research is to develop currency rate forecasting methodology and a set of models based on the analysis of factual (objective) and conceptual (subjective) internet content. The source of factual information is relevant newsfeed of specialized news portals, the source of conceptual information is users’ records in microblogs. The basis of the models’ set includes: 1) parsing scripts of news portals and microblogs; 2) algorithm of factual and conceptual exogenous variables generation on the basis of latent-semantic analysis, sentiment analysis, Granger causality analysis; 3) chosen mathematic forecasting tools such as feedforward neural networks and recurrent networks with long short-term memory (LSTM) the set of inputs of which was formed by applying genetic algorithms. As a result of news feeds and tweets database processing, the set of exogenous factors including four out of fourteen factual variables (infrastructure, activity, dissemination and expect) and two out of eight conceptual ones (calm and confusion) was worked out. Automation of neural networks architecture optimization was conducted with the use of genetic algorithms: chromosome length equaled the number of variables; species were subject to hybridization, mutation and selection based on their fit function, which is MSE for the validation dataset. Comparative analysis of different neural networks architectures allowed proving the expediency of Internet content application for economic time series forecasting and demonstrated high appropriateness of the developed models.