FinBERT and Domain-Adapted Models

A general-purpose language model trained on Wikipedia and news will misread a lot of financial language. "The company guided lower" sounds negative in everyday English but is a routine, sometimes neutral disclosure phrase in earnings calls; "shares were volatile" carries no sentiment at all to a finance reader but might read as vaguely negative to a general model. FinBERT takes the general-purpose BERT language model and continues training it on financial text — earnings calls, analyst reports, financial news — so its internal representations shift toward how financial language is actually used.

This process, called domain adaptation, does not require retraining from scratch — it continues training an already-capable general model on a narrower corpus, which is far cheaper and typically outperforms both a general model used as-is and a finance-specific model trained from zero data.

Worked example. Given the sentence "the company missed on revenue but beat on EPS due to buybacks," a generic sentiment model might average the mixed signals into a neutral score. FinBERT, having seen many earnings-call transcripts, more reliably recognizes that "beat on EPS due to buybacks" is a known qualifier finance analysts treat with skepticism, and can weight the revenue miss more heavily in a way that better matches how professional analysts read the same sentence.

Domain adaptation is now the default recipe for financial NLP tasks generally: take a strong general-purpose model and fine-tune it on in-domain text rather than building a specialized architecture from scratch.