Measuring Sentiment News with Transformer-Based Language Models
创建于 更新于
摘要
Using transformer-based language models to measure sentiment in 143,755 Dow Jones Factiva financial news articles, this study shows that context-aware sentiment scores (FinBERT) align more closely with human judgments than dictionary-based approaches, producing more dispersed and polarized sentiment distributions. The authors construct multiple article- and day-level sentiment indices, validate them against incentivized human annotations, and demonstrate superior in-sample and out-of-sample performance relative to lexicon-based benchmarks. The framework remains robust to alternative aggregation schemes and highlights the value of context in financial sentiment measurement for macro-finance applications [page::1]
