Will the stock go up after earnings? What AI can honestly tell you

Why prediction fails, structurally

The post-earnings move is a function of the gap between the report and expectations, and expectations are only partly visible. The official consensus is public; the whisper number, positioning, and options flow are not, and they shift up to the closing bell. Two identical reports can produce opposite moves depending on what was priced in, which no transcript contains.

There is also a harder problem: if a model could reliably predict the move from public transcripts, the prediction would be arbitraged away the moment it worked. The signal destroys itself. This is not a temporary limitation of current AI; it is what efficient-enough markets do to public information.

What the call genuinely reveals

The call is the company's own account of its state, on the record, in language chosen under pressure. That yields real information: whether guidance moved and how it was hedged, whether the margin story is improving or deteriorating, which questions management answered fluently and which it deflected, and how all of that compares to the previous quarter's wording.

None of this predicts Friday's price. All of it feeds the only question you can actually answer: is this business developing the way my thesis requires? That is a slower question than "up or down tomorrow", and a far more valuable one.

Questions with real answers

• "What changed in the guidance language versus last quarter?" Answerable, verbatim, from two transcripts.
• "Did management address the margin pressure analysts asked about in Q1?" Answerable: yes, no, or evasively, with quotes.
• "What is the bear case in management's own words?" Answerable: risks and hedges are in the transcript.
• "Is the sector confirming or contradicting this company's story?" Answerable across the peer group's calls.
• "What would I expect them to say next quarter if the thesis is right?" Answerable as a written test you check in three months.

The pre-registration trick

The most useful habit an AI chat enables: before the call, write down what the thesis predicts management will say, in one message. After the call, ask what they actually said and compare. This is pre-registration, the same discipline that keeps science honest, applied to a portfolio. It converts hindsight bias into a measurable hit rate, and it costs two questions per quarter.

Over a few seasons you learn something no predictor can sell you: whether your model of the company is any good, and where it drifts wrong.

The honest division of labour

Let the AI do what it can do perfectly: read everything, quote exactly, compare quarters, surface the sector context, refuse to invent. Keep what is yours: the thesis, the position size, the decision. earnings.chat draws that line deliberately, every figure from a query or a quote, every gap reported as a gap, because a tool that pretended to know Friday's price would be worse than useless. The market pays for judgment; the transcripts are how judgment gets fed.