Ask what natural language processing actually does for investment analysis, and the answers drift toward the theatrical: engines that feel the market, summaries that know what matters. The work underneath is more modest and more useful, and four tasks do most of the load-bearing when a pipeline meets a quarterly report. Each has a real record of success; each fails in one signature way that matters.
Task one: extraction, or sorting the named things
For a model, an earnings document is first of all a pile of named objects: metrics, periods, units, currencies, products, jurisdictions. Named-entity recognition tags free cash flow as a metric and the third quarter as a period; relation extraction ties them into triples — metric, period, value — that downstream code can lay into a table a human will then audit.
The specialty failure is aliasing. A model that happily binds revenue and net revenue as one entity will invite adjusted revenue into the same family — and adjusted is precisely where accounting keeps its favourite stories. The defence is a short human pass over every aliased metric before the table is trusted, which is why this review quotes filings instead of re-tabulating them.
Task two: guidance comparison, or philology with numbers
Guidance paragraphs are short futures written in modal verbs. The comparison task isolates sentences that bind an expectation to a figure — expect, anticipate, foresee, continue to — and sets this quarter’s sentence against the one it replaced. The discipline is closer to philology than to statistics: which qualifiers entered, which quietly left, whether the range stayed as wide, whether approximately moved.
Its failure mode is complacency, and the complacency belongs to the reader rather than the code. A softened paragraph can still precede strong results; language lags arithmetic more often than it leads it. The comparison tells you what the company allowed itself to say — that is a record, not a prediction, and the best practitioners repeat that sentence like a metronome.
Task three: tone, measured where it is least rehearsed
The earnings call splits in two. Prepared remarks are written by the hands that wrote the release. The question section is where sentences arrive unrehearsed, and tone scoring — hedge density, hesitation markers, sentence-length collapse under a hard question — is most informative exactly where it is least flattering: in answers that begin with look and end earlier than the answerer intended.
The signature failure is over-reading a single call. Speakers hedge happily under pressure; sarcasm survives transcription; a chief financial officer with a cold sounds like a chief financial officer with doubts. Tone earns trust as a distribution measured across many calls, one company against its own history — never as a verdict on one afternoon.
Task four: summarization that may not round
A summary of financial prose answers to a stricter law than a summary of the news: the numbers must survive contact with the paraphrase. A summary that renders a 2.6 per cent margin improvement as “about 3 per cent” has passed a language test and failed a financial one, and the failure is invisible from the prose alone.
Current best practice keeps the original numbers verbatim, marks anything recomputed as recomputed, and accepts that a summary of an earnings document without its tables is a summary of half a document. The residual failure is fluent drift — the more confident the compressed prose, the more exactly the source must sit beside it for anyone to check.
The honest division of labour
None of the four divines the future. Extraction and guidance comparison document commitments; tone registers the texture of explanation; summarization compresses without corrupting, when it behaves. Assembled, they mechanize the marginal note a careful analyst writes in the margin — nothing more, and nothing less worth having.