Financial sentiment analysis meets algorithmic trading: a performance-based approach

This study investigates the use of financial sentiment analysis (FSA) to build trading strategies. We evaluate tweets from 25 large-cap stocks with three linguistic models (BERTweet, RoBERTa, FinBERT), aggregate daily sentiment, and transform it into algorithmic trading signals. The strategies are compared with different buy-and-hold strategies. The results show that sentiment-based portfolios outperform the benchmarks. Despite near-random directional accuracy, the signals capture large movements that improve risk-adjusted performance. Our results highlight the gap between classification metrics and economic value, demonstrating that sentiment can generate alpha under realistic conditions and motivating application-driven evaluation of natural language processing in finance.