Reading between models
Does a language model see something FinBERT misses?
What this explores
A study of large language model and FinBERT sentiment on UK regulatory disclosures, separating shared sentiment from model disagreement.
How it comes together
Score the same disclosures with both models, separate common sentiment from disagreement, then examine subsequent returns and portfolio performance.
Evidence before claims.
Model disagreement did not show reliable next-month predictive value. Transaction costs are central to interpreting portfolio results.
This is the design preview. Reports, repository links and demos will be connected when the publication-ready materials are supplied.
Technical notes +
Score the same disclosures with both models, separate common sentiment from disagreement, then examine subsequent returns and portfolio performance. The full implementation notes will sit here, alongside the underlying evidence.
Limitations and next steps +
Model disagreement did not show reliable next-month predictive value. Transaction costs are central to interpreting portfolio results.