Does Twitter sentiment permit to forecast the intraday price evolution of quoted financial stocks on the US market

Abstract

Forecasting market data has always been a major concern for financial professionals and investors, both on a daily basis and on an intraday basis. Since the apparition of the first theories, models have been built in order to achieve this very objective. With the growing number of publicly available data, the models began to integrate data more and more precise and with a higher level of granularity. In the late 2000s, a new source of textual data began to appear, the microblogs. With regards to their growing popularity among individuals, and especially investors, research and papers emerged from the financial words to analyse if their contents, usually represented through sentiment, could be used to predict market data.

We propose to continue the research done in that field and analyse the new intraday opportunities they may offer, by focusing on one of the most popular microblogs, Twitter. Using a supervised text classification algorithm, we extracted the intraday sentiment of 805,652 English-language tweets containing a reference (called Cashtag on Twitter) of one of the companies forming the S&P 500. Then, using traditional statistical tools, we analysed its predictable power when aggregated in the time periods of 15, 30 and 60 minutes.

Studying the aggregated impact on our 500 companies, we find that an increase in the sentiment extracted from Twitter during the two and three last 60 minutes’ time periods leads to an increase in the current logarithm return. Furthermore, an increase in the last 30 minutes’ time period of the agreement leads to an increase in the next period volatility. Both effects were statistically significant, although economically small. However, unlike the results of previous research, we did not found that an increase in the number of posted tweets lead to an increase in the trading volume. On the contrary, it seems that the number of tweets is predicted by the trading volume, but not vice-versa.

We also studied the same relationships on an individual basis, focusing on the 10 most heavily discussed companies of the S&P 500. Our analyses did not highlight any predictive relationships between their sentiment and market data time series. It therefore seems that the Twitter feed of popular companies cannot be used to predict changes in their market prices, volatility or trading volume.