FinFakeBERT: financial fake news detection

The intentional use of fake news for financial manipulation or the disruption of financial markets is a serious concern, particularly with the rise of generative artificial intelligence, which is expected to significantly increase its dissemination. A lack of open-access, labeled financial fake news data poses challenges when training effective models for financial fake news detection. To address these challenges, we present FinFakeBERT, a family of models trained using newly curated fake news data. We demonstrate that fine-tuning BERT with a small set of actual fake financial news, following fine-tuning with a large cross-domain fake news dataset and accurate financial news articles, leads to high fake news detection accuracy and significantly reduces the false positive rate (FPR) when tested on several large sets of real financial news articles. Our best model achieves a 2.1% false positive rate (FPR) on real financial news, whereas available benchmark fake-news detectors exhibit FPRs that are more than three to ten times higher.