Financial News Sentiment Learned by BERT: A Strict Out-of-Sample Study

46 Pages Posted: 10 Jan 2022 Last revised: 19 Oct 2023

See all articles by Stefan Salbrechter

Stefan Salbrechter

Deka Investment GmbH; Vienna University of Technology

Date Written: November 25, 2021

Abstract

I investigate the impact of financial news on equity returns and introduce a non-parametric model to generate a sentiment signal, which is then used as a predictor for short-term, single-stock equity return forecasts. I build on Google’s BERT model and sequentially pre-train and fine-tune it using Thomson Reuters financial news data covering the period from 1996 to 2020. With daily return
data of S&P 500 constituents, the analysis shows that financial news carry information that is not immediately reflected in equity prices. News is largely priced-in within one day, with diffusion varying across industries. A trading strategy that leverages the sentiment signal generates an average return per trade of 24.06 bps over an 18 year out-of-sample period.

Keywords: Sentiment Analysis, BERT, NLP, Machine Learning, Return Predictability, Text Mining

JEL Classification: G10, G11, G12, G14

Suggested Citation

Salbrechter, Stefan, Financial News Sentiment Learned by BERT: A Strict Out-of-Sample Study (November 25, 2021). Available at SSRN: https://ssrn.com/abstract=3971880 or http://dx.doi.org/10.2139/ssrn.3971880

Stefan Salbrechter (Contact Author)

Deka Investment GmbH ( email )

Mainzer Landstrasse 16
Frankfurt am Main, 60325
Germany

Vienna University of Technology ( email )

Karlsplatz 13
Vienna
Austria

Do you have a job opening that you would like to promote on SSRN?

Paper statistics

Downloads
519
Abstract Views
4,251
Rank
101,699
PlumX Metrics