Who this is for — Anyone turning words (news, 10-K, transcripts) into numbers for models or dashboards. NLP powers much of modern sentiment work.
NLP (Natural Language Processing) in finance applies linguistic techniques to market documents — tokenization, entity recognition (tickers, CEOs), sentiment classification, topic modeling, LLM summarization — to extract signal from high-frequency text streams.
In plain terms — Teach the machine to «read» headlines and reports like an analyst — at scale and with explicit rules.
Typical pipeline
| Step | Output |
|---|---|
| Ingest & dedupe | Clean stream, UTC timestamps |
| Normalization | Lemmas, boilerplate removal |
| Scoring | Polarity (−1…+1), urgency, novelty |
| Aggregation | Per ticker, sector, time window |
| Features | Model inputs with lags |
Finance-specific lexicons (Loughran-McDonald) beat generic dictionaries on filings and corporate news. LLMs help summarization but need hallucination and cost controls.
Risks
- Look-ahead: news timestamp vs real availability (wire delay)
- Survivorship: only names with media coverage
- Overfitting on small corpora
- Regime: bear language differs from bull — static models decay
Common mistake — LLM sentiment on headlines without verifying publication timestamp — data leakage in backtest.
Example — Earnings call: NLP extracts tone on «guidance» and «margin» → T+1 aggregate score as feature for weekly directional model.
Card
- Output: score, topic, entities.
- Validate: out-of-sample correlation with returns.
- Hub: AI & markets.
Sources
- Araci, D., "FinBERT: Financial Sentiment Analysis with Pre-trained Language Models" (2019) - arXiv.