Sentiment Spin: Attacking Financial Sentiment with GPT-3

29 Pages Posted: 30 Jan 2023 Last revised: 13 Apr 2023

See all articles by Markus Leippold

Markus Leippold

University of Zurich; Swiss Finance Institute

Multiple version iconThere are 2 versions of this paper

Date Written: January 25, 2023

Abstract

The use of dictionaries in financial sentiment analysis and other financial and economic applications remains widespread because keyword-based methods appear more transparent and explainable than more advanced techniques commonly used in computer science. However, this paper demonstrates the vulnerability of using dictionaries by exploiting the eloquence of GPT-3, a sophisticated transformer model, to generate successful adversarial attacks on keyword-based approaches with a success rate close to 99% for negative sentences in the financial phrase base, a well-known human-annotated database for financial sentiment analysis. In contrast, more advanced methods, such as those using context-aware approaches like BERT, remain robust.

Keywords: sentiment analysis in financial markets, keyword-based approach, FinBERT, GPT-3

JEL Classification: G2, G38, C8, M48

Suggested Citation

Leippold, Markus, Sentiment Spin: Attacking Financial Sentiment with GPT-3 (January 25, 2023). Swiss Finance Institute Research Paper No. 23-11, Available at SSRN: https://ssrn.com/abstract=4337182 or http://dx.doi.org/10.2139/ssrn.4337182

Markus Leippold (Contact Author)

University of Zurich ( email )

Rämistrasse 71
Zürich, CH-8006
Switzerland

Swiss Finance Institute ( email )

c/o University of Geneva
40, Bd du Pont-d'Arve
CH-1211 Geneva 4
Switzerland

Do you have negative results from your research you’d like to share?

Paper statistics

Downloads
2,169
Abstract Views
6,149
Rank
12,722
PlumX Metrics