•
Ke, Kelly, and Xiu (2019) show a supervised way to turn news text into return-prediction signals with careful screening. Plain-language intro for after-work Python NLP learners. Not investment advice.
News is messy. Bags of words are huge. It is easy to overfit a story that “sounds smart” in sample and fails out of sample.
Predicting Returns with Text Data (Ke, Kelly, and Xiu, 2019) is a research paper about doing news-to-return prediction more carefully: supervised learning from text with screening so you do not keep every noisy word.
This Hub note is a plain-language intro for after-work Python learners who want a serious NLP-in-finance reading path. It is not investment advice.
The authors study how to map financial news text into signals related to future returns, with methods designed for high-dimensional text. The practical lesson for learners is not a magic dictionary—it is a workflow mindset:
Read the PDF for the exact estimators and empirical design. Use this page only as orientation.
Evening learners often jump from FinBERT demos to “sentiment = alpha.” This paper pushes you toward a harder, healthier habit: treat text prediction like any other supervised problem, with leakage checks and humility about costs and capacity.
If you code later, start smaller than production:
Open PDF (Hub does not host an uploaded copy—external link only):
Paper page: NBER w26186
Educational summary only. Not investment advice. Paper results are research evidence, not a recommendation to trade. You are responsible for your own learning experiments and any real-money decisions (we recommend none from this page).
Build Python fluency first if news pipelines still feel heavy. SoloQuant Academy’s Introduction to Python Programming is the only course link here—short after-work lessons so you can own a .py file before you chase text models.