A quarterly results release contains two kinds of information. The financial statements, which are structured, comparable and easy to process at scale. And the earnings call, which is an hour of unstructured speech.
The second has always been understood to carry information. Managers describe conditions the statements do not capture, and the question-and-answer section produces unrehearsed responses to analysts who have done their own work.
Processing that at scale was the problem. A human can read a handful of transcripts a week. There are thousands per quarter, and the useful comparisons are across many of them.
Why Text Was Hard Before
Most descriptions of an AI investment strategy focus on price data, where the input is already numeric and the modelling question is well defined.
Text is different. Earlier approaches counted words against predefined dictionaries, classifying terms as positive or negative and aggregating the result. That works to a point and misses most of what language does:
Contemporary approaches represent text as vectors rather than word counts, which allows a model to distinguish uses of the same word and to compare documents on meaning rather than vocabulary.
What the Research Actually Measures
The academic literature on this has grown quickly, and a recent review sets out what these applications currently address.
The review notes that transcripts provide information about executive tone, forward-looking statements and market sentiment, and can be useful in predictive modelling and sentiment analysis, and describes datasets combining text, audio and metadata so that models can use tone and intensity alongside the words themselves.
That multimodal direction matters. Risk and sentiment are often inferred from how something is said as much as from what is said, and a transcript alone discards that.
The Shift From Documents to Changes
The more interesting development moves away from scoring individual calls.
Research in this area notes that while studies analysing individual transcripts have shown managerial language such as tone or readability can predict excess returns, more recent work shifts focus from analysing individual documents to examining changes in content across earnings call transcripts, showing that textual modifications over time carry incremental information for return predictability.
The same work constructs a measure of target shifting, testing whether firms that change the objectives they emphasise across successive calls experience different subsequent returns, using a standard cross-sectional protocol.
This is the shift worth understanding. A single transcript scored for positivity is a weak signal, partly because managers optimise their language and partly because the level of positive language varies by company and sector. What one company says relative to what it said last quarter is a comparison against itself, which removes much of that noise.
Where the Adversarial Problem Sits
One structural difficulty deserves stating plainly, because it limits how far any of this goes.
The people producing this text know it is being analysed. Corporate communications teams are aware that language models score their calls, and the literature on strategic disclosure documents that managers obfuscate negative news, producing a readability gap that simple models fail to penetrate.
That makes this an adversarial environment rather than a static one. A measure that works well enough to be noticed creates an incentive to write against it, which is the same dynamic that erodes any published signal.
What This Means for an Individual
Most of this research operates at a scale individuals cannot replicate, and the honest implications are narrower than the field’s enthusiasm suggests:
The first point is the practical takeaway. It requires no tooling, and it applies the same logic the research is converging on.
What Still Requires Judgment
These tools are good at coverage. They can process every transcript in an index in the time a person reads three, and that breadth is a genuine capability rather than a marketing claim.
What they do less well is decide which differences matter. A change in emphasis might reflect a deteriorating business, a new strategy, a different executive speaking, or nothing at all. Distinguishing between those requires knowing the company, and that knowledge is not in the transcript.
The reasonable division is that scale handles the search and judgment handles the interpretation. Neither substitutes for the other, and the research showing what text contains does not show that text alone is sufficient.
