NLP in finance is not new. It is twenty years old. Researchers were extracting sentiment from 10-K filings using bag-of-words models before the iPhone existed. The Loughran-McDonald financial sentiment dictionary was published in 2011. FinBERT, a BERT model fine-tuned on financial text, shipped in 2019. These tools still work. Many production systems at hedge funds and asset managers still run them.
What changed with LLMs is real but more specific than the marketing suggests. Understanding where the actual value sits requires separating three distinct NLP tasks that have different maturity levels, different evidence bases, and different commercial landscapes.
Task 1: Sentiment Extraction from Earnings Calls
This is the oldest and most studied NLP application in finance. The premise is simple. Management tone during earnings calls contains information not captured by the reported numbers. Hedging language, deviations from prepared remarks, changes in word choice quarter-over-quarter. These signals carry predictive content.
The academic literature is substantial. Jegadeesh and Wu (2013) showed that the tone of 10-K filings predicts future earnings and stock returns. Price, Doran, and colleagues found that vocal cues during earnings calls, things a human might call "nervousness," predicted post-call price movement. The effect was statistically significant and economically meaningful.
Here is the problem. The signal has decayed. As more funds adopted sentiment-based NLP, the edge compressed. A strategy that generated 200 basis points of alpha in 2010 might generate 30 today. This is exactly what you would expect from a well-documented anomaly in a market full of quantitative participants. The information gets priced in faster.
That does not mean sentiment analysis is worthless. It means the marginal value has shifted. Sentiment NLP now functions less as a standalone alpha signal and more as a screening and monitoring tool. Portfolio managers use it to flag calls that deviate from expected tone, prioritizing their limited attention. Risk teams use it to detect early warning signs in credit portfolios. The value is real but operational, not directly tradeable at scale.
Task 2: Information Extraction from Filings
This is where NLP delivers the clearest, most measurable productivity gain today. SEC filings are long, standardized, and information-dense. A typical 10-K runs 80 to 150 pages. Pulling specific data points from these documents, revenue by segment, contract terms, covenant details, risk factor changes, has historically required armies of analysts.
Daloopa has built its business specifically around this problem. Their system extracts financial statement data from filings and earnings releases, structuring it into analyst-ready models. The pitch is straightforward: instead of an analyst spending two hours manually updating a model after an earnings release, Daloopa delivers the structured data in minutes. The quality claim is that their extraction accuracy exceeds 99%, verified against source documents. For sell-side analysts covering 15 to 20 companies, this is a genuine time saver.
9fin applies similar extraction capabilities to the leveraged finance market, pulling covenant terms, pricing details, and structural features from credit agreements and bond indentures. These documents are notoriously dense. A single credit agreement can run 300 pages of legal text. 9fin's NLP parses them into structured, searchable data that credit analysts can query without reading every page.
The LLM improvement here is meaningful. Pre-LLM extraction systems were brittle. They worked well on standardized formats and broke on edge cases. LLMs handle variation in document structure more gracefully. They can extract a revenue figure whether it appears in a table, a paragraph, or a footnote. This flexibility reduces the manual cleanup that older systems required.
Task 3: Semantic Search Across Document Corpora
Keyword search fails in finance for obvious reasons. A search for "recession risk" will miss a paragraph discussing "deteriorating macroeconomic conditions and rising unemployment." Semantic search, which matches on meaning rather than exact terms, solves this.
AlphaSense is the largest player in this space. Their platform indexes earnings call transcripts, SEC filings, broker research, news, and proprietary content sets. Users query in natural language and get results ranked by semantic relevance. The company has raised over $650 million and reportedly serves the majority of S&P 500 companies. What AlphaSense understood early is that the value of semantic search scales with corpus size. Searching one document intelligently is nice. Searching across 15 years of earnings calls for every company in a sector is a different category of capability.
Tegus built a large proprietary dataset of expert call transcripts, conversations with former employees, customers, and industry participants. Their NLP layer lets investors search and synthesize across thousands of these transcripts. The data moat matters as much as the technology here. NLP without differentiated content is a commodity.
Kensho, acquired by S&P Global in 2018 for $550 million, provides NLP infrastructure across S&P's data products. Their models process millions of documents to extract entities, events, and relationships that feed into S&P's analytics platforms. Kensho operates at the infrastructure layer, powering capabilities that end users access through S&P Capital IQ and other products.
Rogo has taken a different approach, building what they describe as an AI analyst specifically for finance. Rather than providing a search interface, Rogo attempts to answer complex financial questions directly, synthesizing information from filings, transcripts, and market data. The output looks more like a research note than a list of search results. This is a harder problem and the quality bar is higher. A wrong search result is annoying. A wrong answer presented with confidence is dangerous.
Earnest Research occupies a related but distinct space. They use NLP and alternative data, primarily consumer transaction data, to generate insights on company performance ahead of reported earnings. Their models parse and structure transaction-level data at scale, a task that combines NLP with statistical estimation.
What LLMs Actually Changed
The GPT-era models introduced three capabilities that earlier NLP could not deliver reliably.
Multi-document reasoning. Ask a pre-LLM system to compare how three companies discuss supply chain risk across their most recent 10-Ks and you get nothing useful. Ask a modern LLM with the right context window and retrieval setup and you get a coherent comparison. This is qualitatively new.
Natural language Q&A over financial documents. Instead of constructing boolean search queries, an analyst can ask "What did the CFO say about margin guidance in the Q3 call?" and get a direct answer with source attribution. The interaction model changes from "search and read" to "ask and verify."
Draft generation. LLMs can produce first drafts of research notes, earnings summaries, and credit memos. The output requires editing. It is not publishable as-is. But it reduces the time from raw information to structured analysis. Junior analysts spend less time on formatting and summarization, more time on judgment and interpretation.
What LLMs Did Not Change
The need for verified data. This cannot be stated strongly enough. Hallucination in financial NLP is not a minor inconvenience. It is a material risk. An LLM that fabricates a revenue figure, invents a covenant term, or misattributes a quote from an earnings call can cause real financial harm. Every serious vendor in this space implements source attribution and verification layers on top of their LLM outputs. The ones that do not are building products that institutional investors will not trust.
The need for structured data. LLMs are good at unstructured text. Financial analysis requires structured, comparable, time-series data. Extracting a revenue number from a paragraph is useful only if that number lands in the right cell of a financial model with the right units, the right time period, and the right accounting treatment. This translation from unstructured extraction to structured data remains hard and is where companies like Daloopa differentiate.
The need for domain expertise in model evaluation. A general-purpose LLM does not know that "adjusted EBITDA" means something different at every company. It does not know which non-GAAP adjustments are aggressive and which are standard. It does not know that a covenant described as "Total Net Leverage Ratio" in one credit agreement is economically identical to "Senior Secured Leverage Ratio" in another, despite different names. Domain knowledge still matters. The best systems encode it explicitly.
Where Value Is Real vs. Overhyped
Semantic search across large financial document sets is real, delivered, and widely adopted. AlphaSense's growth validates this. The value proposition is clear and the failure modes are manageable. A bad search result wastes a few seconds.
Automated data extraction from filings is real and getting better. Daloopa, 9fin, and others are shipping products that save measurable analyst hours. The accuracy is high enough for production use with appropriate verification.
Sentiment-based trading signals are mature and largely arbitraged. New entrants claiming alpha from earnings call sentiment are selling something the market priced in years ago. The operational use case, flagging anomalies for human review, remains valid.
LLM-generated research and analysis is promising but early. The hallucination problem is not solved. The liability question, who is responsible when an AI-generated research note contains a material error, is not settled. Institutional adoption is cautious and appropriate.
The vendors worth watching are those solving specific, well-defined extraction and retrieval problems with verifiable outputs. The vendors worth questioning are those promising general-purpose "AI analysts" that claim to replace human judgment. The former is engineering. The latter is marketing.
Financial text is abundant, structured enough to evaluate against ground truth, and high-stakes enough that accuracy matters. That combination makes it one of the better domains for applied NLP. It also makes it unforgiving of sloppy implementations. The technology is good. The question is whether the products built on it are honest about what they can and cannot do.