The European Central Bank published a working paper proposing daily financial news mood indices based on FinBERT, a transformer language model that assesses sentence context rather than counting positive and negative words. Using 143,755 English language financial news articles from 2025, the authors find that transformer based measures capture more polarized sentiment and align more closely with human assessments than dictionary based alternatives. The validation exercise involved 444 participants assessing 588 articles. Transformer based indices recorded correlations of about 0.75 to 0.76 with consensus human ratings, compared with about 0.70 to 0.73 for vocabulary based measures. They also achieved an in-sample macro F1 classification score of up to 0.687 and an out-of-sample score of 0.638, versus 0.187 and 0.162, respectively, for vocabulary based indices, which tended to classify articles as neutral. The paper concludes that contextual models can produce more accurate financial news sentiment indicators for economic monitoring and research, although its findings do not represent the ECB’s views.
2026-09-17European Central Bank
European Central Bank working paper finds transformer models better match human judgments of financial news sentiment
An European Central Bank working paper finds that FinBERT based news mood indices match human assessments of financial news sentiment more closely than dictionary based measures. The analysis covers 143,755 articles and validates the measures against ratings from 444 participants, with transformer based indices substantially outperforming vocabulary based alternatives in sentiment classification. The findings do not represent the ECB’s views.