Natural language processing, or NLP, converts unstructured financial text such as news, filings, and earnings transcripts into quantifiable signals. It tokenises text, identifies relevant entities, classifies sentiment, and aggregates individual scores for an instrument or sector.

NLP does not comprehend news like a human analyst. Its value is processing large volumes of text consistently and at scale, with sentiment data acting as one layer alongside price, volume, and regime data.

This explainer is based on the NLP mechanism, approaches, applications, and limitations described in the source article.

What natural language processing does in financial markets

Natural language processing is a branch of machine learning that extracts structured information from unstructured text. Financial markets generate large amounts of unstructured text, including earnings call transcripts, central bank statements, regulatory filings, news wire items, analyst reports, and social media commentary.

NLP converts this text into numerical representations and scores it against patterns learned from prior examples. The resulting output is a quantifiable signal that a systematic model can process alongside price and volume data.

How NLP turns news into a signal

1. Tokenisation

Tokenisation breaks continuous text into discrete units, usually words or subword segments, that a model can process numerically. Each token is mapped to a vector in a high-dimensional space, where words with related meanings occupy nearby positions.

2. Entity recognition

Entity recognition identifies the companies, instruments, currencies, central banks, and economic indicators mentioned in a text. This helps route a sentiment signal to the relevant instruments. For example, a Federal Reserve decision may relate to dollar-denominated instruments, fixed income markets, and rate-sensitive equities, but not to every asset in the same way.

3. Sentiment classification

Sentiment classification assigns a directional score, typically positive, negative, or neutral, based on patterns learned from labelled financial text. Modern systems can use transformer-based architectures fine-tuned on financial corpora to account for financial vocabulary and context.

4. Signal aggregation

Signal aggregation combines article-level sentiment scores across a defined time window. The composite reading for an instrument or sector can be weighted by source credibility, recency, and relevance.

Rule-based and model-based NLP

Early financial NLP systems used predefined lexicons of positive and negative financial terms. The Loughran-McDonald sentiment dictionary is an example of this rule-based approach. These systems are transparent and computationally cheap, but they can struggle with negation, context-dependent terms, and domain-specific vocabulary.

Model-based systems are trained on labelled datasets and can handle more complex relationships between words and context. A transformer model fine-tuned on earnings call transcripts can learn that a phrase may carry different market-relevant implications depending on the sector, market cycle, and surrounding text.

The trade-off is opacity. A deep learning classification does not necessarily provide a human-readable explanation of why a particular score was assigned. Production financial NLP systems can combine rule-based filters for efficiency and error detection with model-based classifiers for more nuanced sentiment scoring.

Why sentiment works alongside market data

NLP-derived sentiment is most useful as one layer among several rather than as an isolated source of information. Sentiment data captures information flow, while price and volume data captures how the market is responding.

Combining these data types can help classify situations where the information environment and market behaviour are aligned, or where they diverge. This is the design logic behind the Opes Borsa platform's architecture, in which the Sentiment Layer feeds into a broader model alongside price, volume, and regime data.

What NLP cannot do

NLP classifies text against patterns in its training data. It does not understand a central bank statement in the full context of a specific economic cycle or determine what that statement means for the yield curve. It also does not reliably weigh sarcasm, deliberate ambiguity, or the significance of what was not said.

The system may not recognise that a CEO's usually bullish tone makes cautious phrasing unusually significant. Its output is a classification, not human comprehension. This limitation matters when interpreting sentiment signals.

Within its defined task, NLP can process thousands of news items per day in real time without fatigue or emotional response. That makes it suited to converting text flow into consistent, quantifiable directional signals at scale, while leaving interpretation and broader model context to the surrounding system.

The development of financial NLP

NLP has been applied to financial text since at least the early 2000s, when researchers explored computational text analysis for earnings announcements. The technology advanced considerably with transformer architectures after 2017, but its fundamental logic remains the same: extracting structured signals from unstructured text and processing them at a scale and consistency that manual analysis cannot match.

Frequently asked questions

Does NLP understand financial news like a human analyst?

No. NLP classifies text against patterns in its training data rather than understanding context, irony, omissions, or the significance of a speaker's tone in the way a human analyst may.

What does NLP read in financial markets?

NLP can process financial news, earnings call transcripts, central bank statements, regulatory filings, analyst reports, and social media commentary.

How does NLP create a financial signal?

It tokenises text, identifies relevant entities, assigns a sentiment score, and aggregates article-level scores across a defined time window for an instrument or sector.

What is the difference between rule-based and model-based NLP?

Rule-based NLP applies predefined financial terms and scoring rules, while model-based NLP learns patterns from labelled datasets and can handle more context. Rule-based systems are more transparent, while model-based systems can be harder to explain.

Can sentiment data be used on its own?

The source describes sentiment data as working best alongside price, volume, and regime data. Sentiment captures information flow, while price and volume capture market behaviour.

How does Opes Borsa use NLP?

The Opes Borsa architecture includes a Sentiment Layer that classifies incoming market-relevant news in real time and feeds into a broader model alongside price, volume, and regime data.

Key terms

• Natural Language Processing (NLP): A branch of machine learning that extracts structured information from unstructured text and converts it into quantifiable signals.

• Tokenisation: The process of breaking continuous text into discrete units that a model can analyse computationally.

• Entity Recognition: The identification of companies, instruments, currencies, central banks, and economic indicators mentioned in text.

• Sentiment Classification: The assignment of a positive, negative, or neutral directional score to financial text.

• Sentiment Layer: The Opes Borsa component that classifies incoming market-relevant news as part of a broader model.

• Transformer Architecture: A neural network design developed in 2017 that underpins modern large language models and financial NLP systems.

• Signal Aggregation: The process of combining individual sentiment scores over a defined time window for an instrument or sector.