Skip to content
All library documents

Using Large Language Models to Trade on Financial News Sentiment

Article BigQuant
Read the original

Summary

This research summary compares large language models with a financial word-list method for classifying news sentiment and assessing stock returns. It describes analysis of 965,375 US financial news articles matched with daily stock returns. Models classify articles using labels based on subsequent three-day excess returns; regressions then test whether sentiment scores predict next-day returns.

The document reports that OPT had the strongest classification accuracy and return-prediction results, while BERT and FinBERT also outperformed the Loughran–McDonald dictionary. It describes long, short, and long-short portfolios that account for trading costs and adjust execution to publication time. The reported long-short results favor OPT over the other approaches. These findings are limited to the study’s stated sample and evaluation period; the brief does not provide enough detail to assess robustness, portfolio construction choices, or whether results generalize beyond the tested setting.

Key ideas

  • The study compares OPT, BERT, QuantAgent, and FinBERT with a finance-specific word-list method.
  • News sentiment labels are based on the sign of subsequent three-day excess returns.
  • Regression analysis evaluates whether sentiment scores predict next-day stock returns.
  • The reported OPT-based long-short strategy leads the compared strategies after accounting for stated trading costs.
  • The summary does not establish that the reported performance generalizes to other periods or implementations.

Tags

EquitiesMachine learningSentimentBacktestingUS markets

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.