Competitor Earnings Call Analysis Prompts

An earnings call is the most carefully worded document a public competitor produces. A finance chief saying a number will be roughly flat has chosen roughly, and flat, and has considered what happens if either turns out wrong. Analysts are listening for exactly those words.

The results release that goes out alongside it is a different kind of document, written to be re-reported rather than questioned. Working from it instead is quicker and gives you the numbers without the hedging, which is why a competitor's own announcements are the weaker half of the same event.

Summarising works by taking qualification out. That is what makes a summary shorter than the thing it summarises. On most documents the trade is worth it. On this one it removes the part that was written with the most care.

“Growth remained solid, though we expect continued pressure in the second half” becomes “growth remained solid”. The clause that went is the one the company argued about internally. And nothing in the summary tells you a clause is missing, because a good summary never reads like it left something out.

What a summariser removes first

The deletions are not random, which means you can predict them. Hedges go first. Then conditionals, then anything about timing, then any sentence whose main verb is expect rather than is. What survives is the declarative sentences, and those are the ones from the prepared script.

So the default output of a summarising prompt is the scripted half of the call, cleaned up. You can get everything else, but you have to ask for it by name: the qualifiers, the questions that were sidestepped, the words that are new this quarter.

Two prompts rather than three, and the second one exists because of a problem the first cannot fix. Both take the transcript whole.

The line that does the most work in either prompt is the shortest one. Quote, do not characterise. “Management sounded confident” cannot be checked by anybody, and it will be repeated in three meetings as though it can.

Three inputs, and the third one is unusual enough to explain. You write down what you expect to find before you read anything, because of what a long document does to a question.

Keep the analyst questions whole. An answer means almost nothing on its own, and the pair means a great deal. A model given only the answers will report what was said; a model given the pairs can report what was asked and not addressed, which is the thing worth carrying out of an hour of audio.

Log the date you read the transcript as well as the date of the call. The two are rarely the same, transcripts appear on their own schedule, and how current your intelligence is is a property of when you read something rather than when it was said.

We took one quarter of a real transcript and asked two ordinary questions of it. The first is the one everybody asks. The second is the one competitive teams ask, and it is the more dangerous of the two.

The question that was not answered

One exchange in that call is worth reading twice. An analyst asked what customers are choosing to spend money on instead, which is as direct a competitive question as anybody gets to ask. The answer opened: “I wish I had like a very clear crystal ball to give you exact answers there.”

That is not evasion and it is not a concession. It is a company saying it does not know, on the record, to somebody who will ask again in ninety days. A summary records that the call discussed budget scrutiny. The pair records that the question was put and not answered, and only one of those two is worth putting in a brief.

Quotes like that one are what a brief should be made of, and they are worth holding on to beyond the week. Four calls a year is not enough to build a picture on its own, so the exchanges worth keeping are the ones that answer a question your weekly competitor digest keeps raising and never settling.

Four calls a year is not a monitoring rhythm. It is four fixed dates, and between them a public competitor says nothing that has been through this much preparation.

The gap is the problem rather than the calls. A competitor repositions, reprices or ships something in February, and the next chance to hear management describe it in their own words is April. By then the sentence has been polished and the original wording is gone. What they said first is often more informative than what they settle on.

A call is also strictly about the public parent rather than the product you compete with. A company reporting one revenue line for six products tells you nothing about the one in your deals, and no prompt recovers a number that was never disclosed.

Their annual filing is the companion document, and it moves at a different speed. What management says on a call in August has usually been through legal review in a filing months earlier, so comparing two years of filings shows whether a phrase is new or simply newly spoken aloud.

Flares tracks what competitors publish and say between quarters, and dates each change. A phrase that turns up on a call in August can then be read against the week it first appeared on their site, which is usually months earlier.