Machine learning sentiment analysis, COVID-19 news and stock market reactions

Abstract

The recent COVID-19 pandemic represents an unprecedented worldwide event to study the influence of related news on the financial markets, especially during the early stage of the pandemic when information on the new threat came rapidly and was complex for investors to process. In this paper, we investigate whether the flow of news on COVID-19 had an impact on forming market expectations. We analyze 203,886 online articles dealing with COVID-19 and published on three news platforms (MarketWatch.com, NYTimes.com, and Reuters.com) in the period from January to June 2020. Using machine learning techniques, we extract the news sentiment through a financial market-adapted BERT model that enables recognizing the context of each word in a given item. Our results show that there is a statistically significant and positive relationship between sentiment scores and S&P 500 market. Furthermore, we provide evidence that sentiment components and news categories on NYTimes.com were differently related to market returns.

Keywords: COVID-19 news; Sentiment analysis; Stock markets.

Graphical abstract

Fig. 1

Weekly number of articles collected from the three news sources over time: MarketWatch.com, NYTimes.com, and Reuters.com.

Fig. 2

The Preprocessing steps using the Natural Language Toolkit (NLTK). Notes: Text refers to the raw text, which consists of all articles collected from NYTimes, Reuters, and MarketWatch; Sentence Tokenizer (NLTK) indicates the splitting of the text into single sentences using the NLTK; BERT Subword Tokenization refers to the splitting of sentences into words known to the model (or subwords and characters if the entire word is unknown); Conversion of Tokens into IDs indicates the creation of readable IDs; BERT Model is the final step, where BERT is able to train the model based on the preprocessing input.

Fig. 3

FinBERT architecture developed by Araci (2019). Notes: The Financial Phrasebank consists of 4840 sentences from financial news originally developed by Malo et al. (2014); [CLS] stands for Classification and represents the token at the beginning of each sequence, which contains one or two sentences; Token $1$ to $k$ refer to the tokens created by the model. Each token represents a word that is known to the model’s vocabulary (or subwords and characters if the entire word is not known to the model); [SEP] stands for separation and represents the token at the end of each sequence; Dense refers to the dense layer, a neural network layer in which the final classification takes place; Sentiment prediction represents the output of the FinBERT model, which is a positive, negative, or neutral sentiment value.

Fig. 4

Rolling seven-day average of the sentiment indicators over time according to the three news sources: MarketWatch.com, NYTimes.com, and Reuters.com. Dashed red lines indicate numbered significant events related to the COVID-19, as reported in Table 3. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 5

Market returns, realized volatility, and $\Delta log$ trading volumes of the S&P 500 over the period from 23 January 2020 to 22 June 2020.