News Sentiment and Equity Returns – BERT ML Model
Log in to collectAcademic paper
News Sentiment and Equity Returns
Thomas Dangl; Stefan Salbrechter
- ATTU Wien
- ?Vienna University of Technology
Strategy in a nutshell
This is a daily sentiment-driven trading strategy on S&P 500 stocks using financial news from Refinitiv. Positive-news stocks are bought and negative-news stocks are sold, forming long-only or long-short portfolios rebalanced daily. Sentiment signals are generated by FinNewsBERT combined with topic classification and a deep neural network, which classifies news as positive, neutral, or negative. By predicting news impact on stock returns, the strategy aims to generate alpha by capturing price moves driven by market sentiment.
Economic rationale
The strategy works because FinNewsBERT, a domain-specific, smaller BERT model (18.95M parameters) pre-trained exclusively on Thomson Reuters financial news, effectively captures sentiment relevant to stocks. Its compact design reduces computing costs while maintaining accuracy. Out-of-sample backtests (2002–2020) confirm that the extracted sentiment signals are robust for guiding investment decisions.