Agregação de índices de análise de sentimentos com conjuntos nebulosos hesitantes para previsão de séries temporais no mercado financeiro

Abstract

Sentiment analysis is an automatic technique to extract subjective information from texts, such as opinions and sentiments. For providing a time series forecasting using sentiment analysis, sentiment classifications of news and social media posts have to be aggregated into a single value to produce a time series with the same periodicity of the stock market prices, for example daily or hourly. In this paper, we adopt fuzzy linguistic values (and corresponding fuzzy sets) to represent prices and sentiments. Given the fuzzified sentiment index of each tweet, we proceed to an aggregation based on hesitant fuzzy sets, which aim to model the uncertainty caused by the hesitation that may arise in the attribution of degrees of membership of the elements to a fuzzy set. Having fuzzified the sentiment index and aggregated them within the same time period, we produce a fuzzified time series of sentiment data, which can be used as additional information for forecasting models. In this work, we employ a multivariate fuzzy time series (FTS) method, namely Weighted Multivariate FTS (WMVFTS), as the machine learning model. For the experiments we collected tweets posted by Bloomberg and the closing prices of Standard & Poor's 500 Index and Nasdaq Composite Index. The main feature delivered by the proposed method is the capability of improving an FTS method by using hesitant information, such as the news posted on Twitter.