Beyond Sentiment: Detecting Sarcasm in Financial Cash Tag Discourse Using BERT

Abstract

As microblogging grows rapidly and significantly affects online communication, it creates new challenges for sentiment analysis systems. One of them is the sarcasm gap, in which the intended message of the user has a very different meaning from what they actually write. This introduces noise and prediction errors in automated trading systems. This paper proposes a BERT-based framework designed to detect sarcasm within financial cashtag discourse, specifically targeting high-volume cashtags such as $AAPL and $TSLA on platforms including X (formerly Twitter) and StockTwits. The framework leverages bidirectional context through Masked Language Modelling (MLM) to generate deep contextual embeddings capable of identifying polarity reversal—instances where ostensibly positive language conceals negative sentiment. WordPiece tokenization is employed to manage out-of-vocabulary (OOV) financial terminology while preserving domain-critical cashtag prefixes. Experimental evaluation on a Gold Standard dataset comprising 12,500 manually annotated posts (sourced 60% from X and 40% from StockTwits) demonstrates that the proposed BERT-based model achieves an F1-score of 0.87 and an AUPRC of 0.91, representing improvements of 14.5% and 18.2% respectively, over a Bi-LSTM baseline with 128 hidden units and GloVe embeddings. All improvements are statistically significant (p < 0.01, paired bootstrap test). Multi-head attention visualizations via BertViz confirm that the model’s internal reasoning aligns with linguistically meaningful sarcasm indicators. These findings demonstrate that Transformer-based architectures constitute a critical advancement for minimizing false positives and enhancing predictive signal quality in automated financial trading systems.