Portfolio Construction with News Sentiment using Large Language Model

Most existing text-based sentiment measures in finance are lexicon-based which are effectively based on word counts of positive and negative sentiment dictionaries, and naturally lose most information. We measure news sentiment using BERT, a state-of-the-art large language model, which reads and comprehends the whole text, and explore return predictability based on Refinitiv Machine Readable News. The resulting portfolio achieves annualized Sharpe ratios of 2.79, 3.09, and 3.87 when considering news alerts, news alerts and articles’ headlines, and article body contents, respectively, significantly higher than passive investment as proxied by S&P 500 index’s Sharpe ratio of 0.32 and dictionary method of 1.59, 2.94, and 0.04, suggesting that large language models are much better at capturing sentiment, and dictionary methods struggle to extract information from complicated texts. Our results also imply that reacting too fast on incomplete textual news information may yield suboptimal performance. An interesting finding is that news of positive sentiment is tailored to fewer audiences, contain fewer topics, and are generally shorter. 【Note: This manuscript is subject to freuqent updates】