Building Stock Sentiment Factors from Analyst Reports with BERT

Summary

This study turns Chinese analyst-report summaries into equity signals using a fine-tuned BERT sentiment model. It starts with a pretrained Chinese model, adapts it on labeled financial opinion text, and applies it to unlabeled report summaries. Sentence-level sentiment probabilities are aggregated over a rolling 90-day window with linearly declining weights. Because analyst commentary tends to skew positive, an adjusted factor gives greater weight to negative sentiment. The study compares these measures with report-rating and report-count factors.

In the reported tests, the basic sentiment factor performs well, but much of its information overlaps with ratings and report counts. The adjusted factor and its residual after neutralizing those comparison factors perform similarly, which the authors interpret as evidence of more incremental information. A monthly large-cap portfolio selected by sentiment also shows strong historical returns in the reported period. These are backtest results, not evidence of future performance. The study flags possible differences between news and analyst-report language, along with limits in the interpretability tool used to inspect model decisions.

Key ideas

• A Chinese BERT model can be fine-tuned on labeled financial text and applied to unlabeled analyst-report summaries.
• The study aggregates sentence sentiment over a rolling window using linearly declining weights.
• An adjusted sentiment factor places greater weight on negative statements to account for analysts’ generally positive tone.
• The basic factor overlaps with report ratings and report counts, while the adjusted factor appears to add more distinct information in the reported tests.
• Historical portfolio results may not persist, and the study’s language-transfer assumption and interpretability approach have limitations.

Tags

This summary was written by Stratmill's research agent from the original; it is not a copy of the source.