Every reporting season, the same inexpensive object is manufactured thousands of times over: the quarterly earnings release, four to ten pages long, drafted to be skimmed and engineered to be quoted. Its tables travel instantly; its sentences travel slowly and matter more.
Read the way most people read it — top to bottom, once — a release feels dense with fact. Read the way a language pipeline reads it, as a stream of tokens set beside the same company’s previous releases, it turns into a tractable object: mostly repetition, a little novelty, a handful of engineered emphases. This piece walks through that object section by section and asks what a machine can honestly do with each.
Four registers in one file
A typical release mixes four registers under one letterhead. Headline tables state results — revenue, margins, per-share figures — each beside its year-ago comparative. A narrative paragraph explains why the numbers moved, in prose polished by investor relations. An outlook sentence, sometimes stretched to a paragraph, describes the coming period using deliberately soft verbs. And a safe-harbor note at the foot of the page legally un-promises everything above it.
Each register defeats a different kind of automated reading. Tables hold clean numbers but demand entity resolution: which “net revenue” is this one, net of what, for which fiscal period. The narrative explains and persuades while committing to almost nothing, which makes it grammatical and unmeasurable at once. The outlook is the rare case of a company describing the future in words thin enough to mean several outcomes. And the safe-harbor paragraph — which no human reads and every machine should — teaches the register of caution that the whole document performs in.
The hedge census that caught a copying habit
Set two adjacent quarters beside each other and the machine sees what a fast reader cannot: when nothing new was said. Consider the pair below, written to be illustrative rather than attributed:
This year’s third quarter: “We expect margins to improve sequentially during the second half.”
One year earlier: “We expect margins to improve sequentially during the second half.”
The years advance; the promise does not. A hedge census — a count of expect, approximately, anticipate, around, about and their relatives across quarters — will report that pair for what it is: a copying habit in the guidance paragraph. A careful analyst reaches the same conclusion eventually. The census reaches it every quarter, for every filing, at effectively no cost.
Most releases are subtler than that, and the finer signal is substitution. A cost called primarily logistical last quarter returns this quarter as primarily logistical and partly structural — and the added name quietly reclassifies the expense from temporary to probably permanent. Extraction and classification tasks turn substitution patterns into a table the eye was never going to build by hand.
Where the numbers hide in the words
Earnings prose scatters numbers with unusual density: targets, drift guesses (“around 8 per cent”), comparatives, the occasional currency translation in a footnote. The first useful pipeline pass is recognition at clause scale. It lifts adjusted operating income, binds it to the third quarter, treats adjusted as a flag demanding a footnote check, and tags the verb grew with its own softening adverb modestly. What falls out is not a forecast but a ledger of commitments a company made in prose — each checkable next season, which is the only ledger an editorial publication can keep honestly.
Adjectives, meanwhile, are the release’s costume jewellery. Strong, resilient, prudent, disciplined: their arrivals and departures are countable, and counting them is more honest than believing them. A corpus reader holds them apart from the arithmetic, refuses to let them inflate it, and notes the drift when a word that defended last year’s margin is retired without explanation.
What this reading is good for
Set honestly, the machine’s quarterly read is not prediction. It is a consistent second pass: a record of what was said this season, clause by clause, that makes change visible instead of merely memorable. Analysts quote it against their own reading; writers borrow its patterns with attribution; sceptics use its failure cases to keep vendors precise.
Next season the same pipeline will read the same paragraphs again and diff them. That repetition is the point. While the company’s prose works to make every quarter feel new, the machine works at making newness measurable — and both of those efforts deserve the reading this column exists to give them.