Measuring Sentiment News with Transformer-Based Language Models

Battle Count: 94

The paper provides a robust methodology for constructing daily sentiment indices from financial news that could serve as alpha signals or risk indicators in quantitative trading strategies. The transformer-based measures show stronger agreement with human sentiment judgments and produce more polarized signals than dictionary-based alternatives, potentially offering better discrimination for trading decisions. However, the paper does not directly test predictive power for asset returns or trading performance. The daily frequency and the focus on macro-level mood rather than stock-specific sentiment limit direct applicability to high-frequency trading. The indices could be most useful for regime detection, risk management overlays, and medium-frequency strategy signals.

The core pipeline (FinBERT sentence classification → document aggregation → daily fixed-effects regression) is moderately complex but well-documented. FinBERT is publicly available and requires no additional fine-tuning. The six aggregation schemes are straightforward mathematical formulas. The fixed-effects regression for daily mood construction is standard econometrics. However, the full pipeline requires: (1) access to a large news corpus (Factiva or equivalent), (2) sentence segmentation, (3) batch inference with FinBERT, (4) multiple aggregation computations, (5) weighted fixed-effects regression with clustered standard errors, and (6) optional human annotation validation. The computational cost of running FinBERT on 143,755 articles (averaging ~49 sentences each) is non-trivial but manageable with GPU resources.

The paper provides detailed methodology, formula definitions for all six aggregation schemes, and preregisters the human annotation experiment on OSF (https://osf.io/5ruva/overview). However, the primary data source (Dow Jones Factiva) is proprietary and not publicly available. The FinBERT model is publicly available. No GitHub repository is mentioned. The fixed-effects regression specification and aggregation pipeline are described in sufficient detail for replication given access to similar news corpora.