The Capability Line — What AI Can (and Can’t) Do for a Trader

The line is not really about difficulty. People assume the model must be better at easy questions and worse at hard ones, but that is not how it divides. It is excellent at a genuinely hard job — reading forty pages of dense management commentary and returning the six claims that matter — and useless at a question a child can phrase, which is what a stock does tomorrow.

The line is about whether the answer already exists in front of it. Compression, restructuring, comparison and drafting all work on material you supplied. The information is present; the job is transformation, and transformation is exactly what a text-completion machine does well. Prediction, recall of recent events and self-assessment all require something the model does not have in the room with it.

This is why the same tool can feel brilliant on Monday and reckless on Tuesday. Nothing changed about the tool. What changed was whether you handed it the material it needed. Once you start noticing that, the variability disappears and the machine becomes predictable — which for a research process matters far more than being impressive.

The map below is the whole article in one view. It is worth reading the third column first, because that column is where the work actually lives. Nothing on the strong side is trust-free; it is simply the side where your verification is cheap and quick rather than impossible.

The most valuable thing a language model does for a working trader is turn something long into something structured. A three-hundred-page annual report, a ninety-minute concall transcript, a twenty-page scheme document — these contain perhaps a dozen facts you actually need, buried in language designed to be thorough rather than readable. Extracting those dozen facts is real work, and the machine does it in seconds.

This works because nothing is being invented. You supplied the text. The model is rearranging material that is already in front of it, which is the operation it is genuinely built for. Ask for the same six headings every time — demand drivers, cost commentary, capital expenditure, debt, related-party items, anything management declined to answer — and you get a scannable brief where you previously had a wall.

What it buys you is not accuracy. It is coverage. Most retail research fails not because a number was wrong but because a document was never opened. Compression makes it realistic to actually read the filings for a watchlist of twenty names in an evening, which is a different quality of process from reading three of them properly and skimming headlines for the rest.

The discipline that makes it safe is a page reference beside every claim. Once the model must name where something came from, two useful things happen: your verification takes seconds instead of minutes, and a claim with no locatable source becomes visibly suspect rather than blending into the rest of the summary.

Human screening degrades under repetition, and it degrades invisibly. By the twelfth name in an evening you are not applying the same eight checks you applied to the first. You are skipping the two that are tedious, weighting the one you find interesting, and quietly forming a view before the checks are done. Everyone does this. It is not a discipline failure; it is what attention does.

A model does not get tired and does not get bored. Give it the same eight questions against twelve different documents and it will ask all eight, twelve times, in the same order, with the same phrasing. That consistency is worth more than most people credit, because the value of a checklist comes entirely from its being applied uniformly. A checklist applied selectively is just a list of things you already believed.

This is also where the model complements rather than replaces you. Judgement about which eight checks matter is entirely yours — it depends on the sector, the setup, your own process and what you have been burnt by before. The machine contributes no view on that. What it contributes is the guarantee that check number seven gets asked at eleven at night with the same care as at eight in the evening.

The practical shape is one document per run, the same prompt each time, and output in a fixed format you can stack side by side. Twelve briefs with identical headings can be compared in one screen. Twelve briefs with different structures cannot be compared at all, which quietly defeats the purpose.

A blank page is expensive. Getting a comparison table started, roughing out the structure of a thesis, listing the ways an idea could be wrong, converting your own scattered notes into something with headings — these are all jobs where getting to a bad first version quickly is worth more than getting to a good version slowly. Editing is faster than originating, for almost everyone.

The model is well suited to this precisely because a draft carries no authority. Nobody acts on a draft. Its errors are supposed to be found and fixed, which means the technology’s central weakness — confident wrongness — is contained by the format itself. You were always going to rewrite it.

The most useful drafting job in market work is the inversion. You have formed a view; ask the model to construct the strongest case against it using only the document you supplied. It has no ego and no position, so it will do this cleanly. You keep the argument, discard the rest, and you have stress-tested a thesis in two minutes rather than waiting for the market to do it for you.

The failure mode here is subtle and worth naming. A fluent draft is persuasive by construction, and after two edits it starts to feel like yours. Keep drafts visibly separate from verified notes — a different file, a different colour, a header that says DRAFT. Anything that quietly graduates from draft to research without a verification step in between is a problem waiting for a bad quarter.

Nothing on this planet can tell you where NIFTY closes tomorrow, and a language model is further from being able to than most things, because it has no market data, no model of price and no forecasting mechanism of any kind. Ask it anyway and you will get an answer, because producing an answer is what it does. The answer is a plausible-sounding sentence, and that is all it is.

It is worth being precise about why, since the reason is not merely that markets are hard. A language model has learned the patterns of how people write about markets. So when you ask where a stock goes next, the machine reproduces the shape of market commentary — a level, a condition, a hedge, a target. It is a very good imitation of an analyst’s sentence, generated with no information about the stock at all.

This also covers the questions that are price prediction in disguise, which is most of the ones people actually type. Is this a good entry. Should I hold this through results. Will this breakout hold. Each of those requires knowledge of the future, and each will receive a confident reply. The reply tells you what market commentary usually sounds like. It tells you nothing about your position.

The reframe that works is to convert every prediction question into a fact question. Not "will this breakout hold" but "list every statement in this filing about order book and capacity, with references". Not "should I hold through results" but "what did management guide to last quarter, in their own words". The model can genuinely help with the second version of each. The decision stays where it belongs, which is with you.

The model knows what it learned up to its training cutoff, plus whatever you paste into the conversation. That is the complete list. It has no live feed, no access to the exchange website, no view of your holdings, and no knowledge of the filing that came out this morning unless you put it there yourself.

The failure is not the ignorance. Ignorance is fine and expected. The failure is that the machine cannot perceive its own boundary. There is no internal signal that says this question lands outside what I have. So a question about last week’s board meeting outcome gets the same fluent treatment as a question about a document sitting in the prompt, and the two answers look identical on the page.

This is also true within a document. If you paste chapter four and ask about something discussed in chapter nine, the model will not say that chapter nine is missing. It will answer from general knowledge of how such things usually read, which is exactly the situation where a plausible fabrication is most likely and hardest to catch, because the answer is in the right register and about the right company.

The habit that solves it is unfashionably simple. For anything time-sensitive or document-specific, supply the material. Then ask, in the same prompt, for the source sentence behind each claim. A model working over pasted text with a citation requirement is a substantially different tool from a model asked to recall — same machine, completely different reliability.

You are used to reading confidence as information. A colleague who is sure of a figure says it flatly; one who is guessing hedges, slows down, or offers to check. That signal is genuinely useful with people and completely absent here. The model has no internal measure of whether a statement corresponds to anything real, so it has nothing to hedge with.

What you get is a uniform surface. A correctly extracted figure and an invented one are written with the same assurance, the same supporting detail, the same willingness to elaborate if you ask a follow-up. Ask it directly how confident it is and you get another generated sentence — the text that plausibly follows a question about confidence, not a reading from any gauge.

The consequence is more practical than it first appears. You cannot triage. With a human analyst you check the claims they seemed shaky on and take the rest on reasonable trust, which is what makes working with people efficient. Here there is no shakiness to detect, so triage by feel is not merely unreliable — it is actively misleading, because the fabricated claims are often the most fluent ones.

So verification has to be organised by category rather than by impression. Every number that will touch a position gets checked. Every citation gets opened. Every claim about anything after the cutoff gets sourced or deleted. That sounds heavy until you notice it applies to a handful of lines in a page-long brief, and that the alternative is trusting a signal that does not exist.

When you are unsure which side of the line a task sits on, ask yourself one question: would I accept this from a junior analyst without checking their sources? Not a bad analyst — a capable, hard-working one who has been with you a few weeks and whose judgement you have not yet had a chance to calibrate.

If the answer is yes, the model has almost certainly done its job. A structured summary, a first-pass comparison, a consistently applied checklist, a rough draft — you would take all of these from a new junior and work with them, precisely because their errors are the kind you would catch while working. That is the strong side of the line.

If the answer is no — if it is a specific figure, a claim about a specific filing, or anything that will move a real position — then the model’s job is finished and yours begins. You would not let a new junior’s number go into a decision unchecked, and you have far more reason to check this one, because at least the junior would have told you they were unsure.

There is a shorter version for when even that feels like too much thought. Ask: if this is wrong, how would I know? If you have a concrete answer — page eighty-two of the annual report, the filing on the exchange website, the transcript you have open — proceed. If you have no way to find out, you are not looking at research yet. You are looking at a well-written suggestion of what research might say.