Natural language processing (NLP) on filings and transcripts is the practice of converting unstructured corporate text — SEC filings (10-K, 10-Q, 8-K), earnings-call transcripts, proxy statements, conference-call Q&A — into structured numerical signals: a tone score, an uncertainty measure, a topic distribution, or a year-over-year change metric. The core tension is that this text is simultaneously information-rich (managers reveal a great deal in word choice, hedging, and what they change versus leave alone) and adversarially constructed (the same text is drafted by lawyers and IR teams precisely to manage impressions). The discipline is therefore an arms race between extracting genuine signal and being misled by deliberate spin, boilerplate, and — increasingly — text that issuers know is being machine-read.

How it's calculated / formed

There are three broad generations of method, often layered together:

1. Dictionary / lexicon (bag-of-words) tone. Count words against domain word lists and normalize by document length. The foundational resource is the Loughran-McDonald (LM) Master Dictionary from Notre Dame, built from the LM 2011 Journal of Finance paper "When Is a Liability Not a Liability?". The dictionary tags words into categories — negative, positive, uncertainty, litigious, strong/weak modal, constraining (and a later complexity lexicon) (Notre Dame SRAF). Its central insight: roughly three-fourths of words flagged negative by the general-purpose Harvard dictionary are not negative in a financial context — words like "liability," "tax," "cost," "mine," "cancer," and "capital" are industry vocabulary, not sentiment (LM 2011, Journal of Finance). Tone is typically (positive − negative) / total words, or just the negative ratio, which LM found more reliable.

2. Document-change / similarity metrics. Rather than tone, measure how much a filing changed from the prior period using cosine similarity, Jaccard similarity, or edit distance on the raw text. This is the engine behind the "Lazy Prices" work (Cohen, Malloy & Nguyen, 2020 Journal of Finance): firms that change their 10-K/10-Q language signal trouble; those that leave it unchanged signal stability (Lazy Prices, Wiley).

3. Transformer / LLM models. Context-aware models read sentences rather than counting words. FinBERT (Araci, 2019) fine-tunes BERT on financial text and the Financial PhraseBank, outputting positive/negative/neutral probabilities (FinBERT paper; ProsusAI/finbert). Modern pipelines use FinBERT, FinVADER, or general LLMs (GPT-4-class) for classification, summarization, and Q&A-vs-prepared-remarks comparison. Hybrid stacks — LM word lists to label, BERT to capture nuance — are now common.

How it's used in practice

Buy-side and sell-side users extract several distinct signals from the same documents:

• Aggregate tone of management discussion (MD&A) and the prepared earnings-call remarks, as a continuous sentiment factor.
• Uncertainty / litigious / risk-factor language — often more predictive of volatility than tone is of returns.
• Prepared-remarks vs. Q&A divergence — analysts probe; a confident script followed by evasive answers is a tell.
• Filing-change detection — flagging modified risk factors, litigation language, or CEO/CFO references.
• Topic/event extraction — pulling guidance, M&A, restructuring mentions into structured fields.

These feed quant cross-sectional factors, event-driven overlays, risk dashboards, and fundamental-analyst triage tools that surface "read this filing first." For longer documents, NLP is also a time-saver: surfacing what changed and what's notable rather than predicting price directly.

Adoption, debate & evidence

The academic evidence is genuinely positive but should not be over-read. LM's dictionaries are the de facto standard in accounting/finance research. Lazy Prices reports a long-short portfolio (long non-changers, short changers) earning up to ~188 bps/month of alpha in 1995–2014, with the striking feature that returns accrue with a delay — there is little announcement-day reaction — consistent with investor inattention (NBER w25084). On transcripts, an NYU Courant study reports an NLP-derived signal delivering a double-digit annualized alpha, stronger for small, lightly-covered stocks (NYU Courant seminar), and Nordic research finds FinBERT/LLM methods outperform pure dictionaries for abnormal-return prediction (Aalto thesis).

The honest caveats: (1) these are published, in-sample academic results; live, after-cost, capacity-constrained returns are far lower and rarely disclosed. (2) The signal is decaying through crowding — practitioner sources for 2025-26 describe simple text-based earnings-call NLP as increasingly commoditized, pushing sophisticated funds toward multimodal (text + audio + other) approaches (CFA Institute; Resonanz Capital). (3) Reflexivity: once managers know filings are machine-scored, word choice adapts — a moving target that erodes static lexicons.

Strengths & limitations

It works best where the text is informationally underexploited: small caps, dense risk-factor sections, subtle year-over-year filing edits, and the unscripted Q&A. It fails when (a) the general-purpose-dictionary trap is not corrected — the single most common misuse is applying a generic sentiment lexicon to finance text and inheriting LM's three-fourths misclassification rate; (b) tone is treated as causal when it is correlated with already-public news; (c) boilerplate and legal templating drown out signal; and (d) deliberate impression management is read at face value. Dictionary methods are transparent and cheap but blind to context and negation; transformer methods capture context but are opaque, costly, and can hallucinate or overfit small labeled sets.

Sources

• Loughran & McDonald (2011), "When Is a Liability Not a Liability? Textual Analysis, Dictionaries, and 10-Ks," Journal of Finance — Wiley
• Notre Dame SRAF, Loughran-McDonald Master Dictionary — sraf.nd.edu
• Cohen, Malloy & Nguyen (2020), "Lazy Prices," Journal of Finance — Wiley; NBER w25084
• Araci (2019), "FinBERT: Financial Sentiment Analysis with Pre-trained Language Models" — HF papers; model: ProsusAI/finbert
• NYU Courant Mathematical Finance Seminar, NLP earnings-call transcript predictiveness — math.nyu.edu
• Aalto thesis, FinBERT/FinVADER earnings-call sentiment, Nordics — aaltodoc
• CFA Institute, "Using NLP to unlock a treasure trove of alternative data" — cfainstitute.org
• Resonanz Capital, on crowded alt-data alpha decay — resonanzcapital.com

Dispute flag: published alphas (Lazy Prices ~188 bps/mo; transcript double-digit annualized alpha) are in-sample academic figures; live net-of-cost returns are unverified and, per 2025-26 practitioner reporting, materially eroded by crowding.