Machine Learning-Based Analysis of the Association Between Online Texts and Stock Price Movements

Abstract

The paper presents the result of experiments that were designed with the goal of revealing the association between texts published in online environments (Yahoo! Finance, Facebook, and Twitter) and changes in stock prices of the corresponding companies at a micro level. The association between lexicon detected sentiment and stock price movements was not confirmed. It was, however, possible to reveal and quantify such association with the application of machine learning-based classification. From the experiments it was obvious that the data preparation procedure had a substantial impact on the results. Thus, different stock price smoothing, lags between the release of documents and related stock price changes, five levels of a minimal stock price change, three different weighting schemes for structured document representation, and six classifiers were studied. It has been shown that at least part of the movement of stock prices is associated with the textual content if a proper combination of processing parameters is selected.

Analysis of the Association between Topics in Online Documents and Stock Price Movements

“…However, focusing simply on the sentiment (positive and negative) dimensions does not always bring useful predictions (Li et al, 2014). In our previous research (Dařena et al, 2018), we revealed that it was possible to uncover and quantify a relationship between unstructured texts and stock price movements at a micro level with the application of machine learning. The data to be classified included the texts converted to their structured representation (bag-of-words).…”

“…Moving averages based on a larger number of days (20 days) had more positive impact than moving averages based on a smaller number of days. The difference between the Simple Moving Average and the Exponentially Weighted Moving Average (NIST / SEMATECH, 2016) were negligible (Dařena et al, 2018).…”

“…We considered 1 % minimal price difference to have sufficient amount of data. The analysis also considered a one-day lag between the publication of a text and a stock price movement as the association for this lag was found to be the strongest for this specific data (Dařena et al, 2018).…”

Text-Mining in Streams of Textual Data Using Time Series Applied to Stock Market

“…Lee et al (2014) used a minimal change of 1 % and Mittermayer (2004) worked with 1 % average change and 3 % extremes in the change. And last but not least, Darena et al (2018) used values 1 to 5 %. Regarding the window sizes it was more of a guess because there is no recommendation.…”

“…Texts contain objective or subjective information (Darena et al, 2018). Behavioural finance theory says that emotions may deeply influence behaviour and decision making of individuals as well as whole human societies (Kearney, 2014).…”