Introducing Sentiment Trading with Large Language Models for after-work quants

Dictionary sentiment is still taught in many finance courses. Large language models (LLMs) are now used to score news tone. Which approach actually lines up better with next-day returns in a careful study?

Sentiment Trading with Large Language Models (Kirtac and Germano, 2024) compares several NLP stacks on a large U.S. financial-news sample, including BERT, OPT, FinBERT, and the classic Loughran–McDonald dictionary.

This Hub note is orientation for after-work learners. It is not investment advice and not a signal to copy.

What they did (plain English)

• Large news corpus spanning many years (paper reports on the order of ~965k articles, 2010–mid-2023).
• Score sentiment with modern LLMs / transformers and with a traditional finance word list.
• Relate scores to subsequent daily stock returns and discuss strategy-style evaluation (read the PDF for costs and construction details).

Headline comparison from the paper’s abstract: OPT (GPT-3 family) shows the highest reported sentiment-classification accuracy among the models they highlight, with BERT and FinBERT close behind, and the Loughran–McDonald dictionary much weaker on their accuracy metric. Treat numbers as paper results, not a live edge.

Why it is useful after work

You will see Hub and GitHub posts that say “just use FinBERT.” This paper is a reminder to compare baselines, including old dictionaries, and to ask what “accuracy” even means for your task. For evening study:

• reproduce a tiny sentiment histogram on public headlines
• compare two scorers on the same 50 headlines
• do not scale to portfolio code until your Python basics are solid

One-evening reading checklist

1. Read abstract + introduction.
2. Note which models are compared and what “accuracy” refers to.
3. Find where they discuss returns / trading-style tests—and any caveats on costs.
4. Write one paragraph: what you would not claim after one reading.

Download the PDF

Download PDF

Landing: arXiv:2412.19245

Important disclaimer

Not investment advice. LLM scores and paper backtests are research outputs. Markets change; costs, latency, and capacity can erase paper-looking edges. Do not trade from this summary.

What this is not

• Not CEF or funded-certificate content
• Not a Quant/Macro enrollment CTA
• Not a broker or API sales page

Next step

Only course CTA on this page: Introduction to Python Programming at SoloQuant Academy—short after-work lessons so you can load text files and run small scripts confidently before you touch LLM sentiment stacks.