beginner9 min read1 of 24

    What AI Actually Is in a Trading Context

    Four different technologies wear the same label. Knowing which one you are holding decides what it is safe to use it for.

    Rohit Singh

    Mr. Chartist · SEBI RA INH000015297

    Module

    Open almost any trading product page in India today and you will find the same two words near the top: AI-powered. Open ten of them and you will find the same two words ten times. What sits behind the phrase, though, is not one technology. It is at least four, and they have very little in common — different inputs, different outputs, and completely different ways of being wrong. The label has collapsed a distinction that matters. Your job, before you trust any of it with money, is to reopen that distinction.

    The four are these. A large language model — the machine behind a chat box — which completes text. A vision model, which describes what an image looks like. A supervised machine-learning predictor, which is fitted to columns of historical numeric market data. And plain rule-based automation: a set of if-this-then-that conditions that never learned anything and never will. A surprising amount of software sold under the AI label is the fourth one.

    None of these is fake, and none of them is useless. Each is a real tool with a real job. The damage comes from mismatching them — expecting a chat model to know yesterday’s closing price, expecting a vision model to measure a level off a chart image, expecting a fixed rule to notice that the market regime has changed underneath it. This article takes each of the four in turn: what it genuinely does, the specific way it fails, and how you work out which one you are actually holding.

    The one thing to remember

    AI is a single label stretched across four unrelated technologies — until you can name which one a product is running, you cannot know what it is safe to trust it with.

    The Label Is Marketing. The Machine Underneath Is Not.

    There is no technical bar a piece of software has to clear before its marketing page may use the word AI. The phrase carries commercial weight, so it gets applied broadly — to a chat assistant, to a screener that has run the same three conditions since 2019, and to everything in between. That is not necessarily dishonesty. It is a word doing too much work.

    The practical cost lands on you. When you cannot name the technology, you calibrate your trust to the label instead of to the machine. You end up expecting a summariser to forecast, or dismissing a perfectly sound rule engine because it is not clever enough to be exciting. Both mistakes cost you — one in money, the other in a tool you should have used and did not.

    Think of the word engine. A moped has one and a truck has one. Knowing that a vehicle has an engine tells you nothing about what it can pull. You would never buy a vehicle on that word alone, yet that is roughly what happens every time somebody signs up for a product because the page said AI-powered.

    Once you can name the technology, three things become easy. You know what output to expect. You know the specific failure mode to watch for, because each of the four fails differently. And you know the one question that will get a straight answer out of a vendor, because vague questions get vague answers and specific ones do not.

    Four Technologies Wearing One Label

    The fastest way to keep the four separate is to look at what goes in and what comes out. A language model takes text and returns text. A vision model takes an image and returns text describing it. A supervised predictor takes rows of numbers and returns a number or a category. Rule-based automation takes a data feed and returns a fixed action — an alert, a flag, a sorted list.

    That input-output pairing is not a simplification for beginners. It is the actual architecture. A language model has no arithmetic engine bolted on; it produces the text that most plausibly follows your question. A vision model has no price series behind the picture it is looking at. A trained predictor has no idea what any of its columns mean. A rule has no capacity to change itself.

    Two of the four learn from data during training and then stop: the language model and the supervised predictor. The vision model is a trained model too, but it learned to describe pictures, not to measure them. The fourth learned nothing at all, because nobody trained it — somebody wrote conditions, which is a different activity entirely.

    The table below is worth returning to. Read the failure column first. It is the column that decides how much of your process you are willing to hand over, and it is the column vendors are least likely to write out for you.

    One label, four different machines

    Tell them apart by what goes in and what comes out — not by the word on the marketing page.

    Language modelthe chat boxGoes inText you pasteComes outText backFails by
    Completes rather than retrieves
    Learned from data
    Vision modelimage readerGoes inAn imageComes outWords describing itFails by
    Describes, cannot measure
    Learned from data
    Supervised predictorfitted to columnsGoes inRows of numbersComes outA number or classFails by
    Knows nothing of what the columns mean
    Learned from data
    Rule automationif-this-then-thatGoes inA data feedComes outA fixed actionFails by
    Cannot notice the regime changed
    Learned nothing
    Until you can name which of these a product runs, you cannot know what it is safe to trust it with.
    The four technologies sold under one label, separated by what each one actually takes in and gives back. Only the first three learned anything from data; the fourth is conditions somebody wrote by hand.
    TechnologyWhat it doesWhere it failsHonest use for a trader
    Large language model (chat)Completes text — summarises, restructures, draftsStates things it never saw, fluently and confidentlyCompressing a filing or concall into notes you then verify
    Vision model (image reading)Describes what a picture resemblesReads pixels; never measures price, volume or datesFirst-pass triage of chart screenshots you check yourself
    Supervised ML predictorFits historical numeric patterns to an outcome labelMemorises noise; decays quietly when the regime shiftsRanking a universe whose drivers you already understand
    Rule-based automationExecutes fixed if-then conditions on a data feedCannot adapt; keeps firing after the setup stops workingAlerts, scans and repeatable mechanical checks
    The four technologies, side by side

    Language Models — Machines That Complete Text

    A large language model is trained on an enormous quantity of written text and learns, in effect, one skill: given some text, produce the text that plausibly comes next. That is the whole mechanism. When you ask it a question, it is not looking anything up. It is generating the most likely continuation of a conversation in which somebody asked that question.

    This is why it is so good at the jobs it is good at. Summarising a forty-page annual report section, restructuring a rambling concall answer into three clean bullets, drafting a comparison table from text you supplied — these are all text-to-text transformations, and that is exactly what the machine does. Give it the source material in the prompt and it works on what is in front of it.

    The failure mode is specific and it is the one that catches everybody. Because it produces plausible continuations rather than retrieved facts, a question it has no information about still produces an answer. Ask for a company’s segment margin for a quarter it never saw and you will get a number that looks exactly like a real one — same format, same confident tone, same decimal places. Fluency is not accuracy, and the machine has no way to signal the difference.

    So the honest use is narrow and genuinely valuable: give it the text, ask it to work on the text, and check anything numeric against the source. The moment you are relying on its memory of a document rather than on a document you pasted in, you have left the zone where this technology is reliable.

    It is not retrieving an answer. It is producing the most plausible continuation of your question — and plausible is not the same as true.

    Text in, text out — the job a language model is actually for
    You are working only from the transcript text I paste below. Do not use anything you know about this company from anywhere else.
    
    Transcript:
    [paste the concall section here]
    
    Return exactly this:
    1. Three claims management made about demand, each followed by the sentence it came from.
    2. Three claims about costs or margins, each followed by the sentence it came from.
    3. Any analyst question that management did not answer directly.
    4. A list headed "Not stated in this transcript" for anything above you could not find.
    
    Do not estimate. Do not fill a gap with a plausible number. If it is not in the text, it belongs in list 4.

    When to use — After an earnings call, when you have the transcript open and want a structured read of it in two minutes instead of forty.

    A good answer — Every claim carries the sentence it came from, so you can find it in the transcript in seconds. Section 4 is not empty — a real transcript never answers everything, and a model that returns nothing there is filling gaps instead of reporting them.

    Vision Models — Machines That Describe Pictures

    A vision-capable model takes an image and produces a description of it. It learned by seeing enormous numbers of images paired with captions, so what it has really learned is resemblance: this arrangement of shapes tends to be described using these words. Point it at a chart screenshot and it will tell you what that picture looks like to something that has seen a great many chart pictures.

    That is genuinely useful at the triage stage. Hand it a screenshot of a daily chart and ask whether price appears to be consolidating in a range or trending, whether the recent candles show long upper wicks, whether there is a visible gap. Those are descriptive questions about the picture, and description is the job.

    The failure mode is that there is no price data behind the image. The model is not reading a series of open, high, low and close values. It is reading pixels. It cannot tell you the exact level a breakout occurred at, it cannot compare the breakout candle’s volume to the previous twenty candles, and it cannot tell you how many candles ago the base formed unless it can literally count the bars in the picture — which it does badly.

    It will also read axis labels wrongly on a compressed screenshot, invent a level that sits between two gridlines, and describe a pattern with more conviction than the picture supports. All of that is consistent with what it is: a describer of images, not a measurer of markets.

    A vision model is answering "what does this picture look like", never "what did price actually do" — the difference is the entire safety boundary.
    It has no access to the underlying OHLC series, so any specific level, gap size or volume comparison it states is read off pixels, not data.
    Screenshot quality changes the answer: a compressed, low-resolution image produces more confident errors, not fewer.
    It is genuinely useful for describing structure — range versus trend, long wicks, visible gaps — which is a real first-pass filter on a long watchlist.
    Every level it names has to be confirmed on your own chart before it enters a note, let alone a decision.

    Watch out — If a product reads chart images and then states precise support and resistance levels, ask where those numbers came from. If the only input was a screenshot, they were estimated from pixels — not measured from price data.

    Supervised Learning — Machines Fitted to Numbers

    This is the technology most people are actually picturing when they hear that AI can analyse markets. You assemble a table of historical data. Each row is one stock on one date. The columns — the features — might be percentage distance from a recent high, volume relative to its own recent average, days since a filing. Then you add a label: what happened over the following twenty candles. The model learns a mapping from features to label.

    It is a completely different machine from a chat model. There is no language in it and no reasoning in it. It does not know that a column called volume means volume. It is a fitted function, and its only skill is reproducing relationships that were present in the data it was fitted on.

    The first failure mode is overfitting: with enough columns and enough freedom, a model will latch onto coincidences in the training data and reproduce them beautifully. The backtest looks superb. The relationship was never real, so live performance looks nothing like it. This failure is invisible unless you deliberately test on data the model never saw.

    The second is drift. Even a genuine relationship can stop holding — liquidity changes, a sector re-rates, participation shifts. The model does not error out when this happens. It keeps producing confident outputs against a world that has moved. Nothing goes red. The quality just decays.

    The honest ceiling is worth stating plainly. On liquid, well-studied markets like NSE large caps, a carefully built model may add a small statistical tilt on top of a process you already understand. It does not produce certainty, it does not survive being left unattended, and any product describing it in stronger terms than that is describing marketing, not machine learning.

    Rule-Based Automation — The One That Never Learned Anything

    Here is a rule. If today’s close is above the highest close of the previous twenty candles, and today’s volume is above the average volume of those twenty candles, raise an alert. That is a complete, working piece of trading software. It contains no learning of any kind. Somebody wrote the conditions, and the machine checks them.

    A large share of the screeners, alert engines and dashboards marketed as AI-powered are exactly this, sometimes with a language model bolted on at the end to phrase the output as a sentence. That combination is common enough to be worth naming: the intelligence in the marketing sits in the wrapper, and the logic sits in a list of thresholds somebody typed.

    This is not a scandal, and it is often the right answer. Deterministic rules are the most auditable software on this list. You can read them, argue with them, change one number and watch the effect. Neither a language model nor a fitted predictor gives you that. Given the choice between a rule you fully understand and a model you do not, a serious process usually takes the rule.

    Its failure mode is rigidity. The rule cannot notice that the market changed. It fires the same breakout alert in a choppy, range-bound tape as it did in a trending one, because the conditions are met either way and the conditions are all it has. It also breaks silently when an upstream data feed changes format — the scan simply returns nothing, and nothing looks the same as no setups today.

    Pro tip — Rule-based is frequently the honest answer to a trading problem. A condition you can read and edit is easier to trust, and far easier to debug at 9:20 in the morning, than a model whose reasoning you cannot inspect.

    How to Tell Which One You Are Actually Holding

    The quickest tell is what the product asks you to supply. A chat box means a language model. A screenshot upload means a vision model. A per-stock score or probability derived from history means a trained predictor. A settings page that is a list of thresholds — periods, percentages, volume multiples — means rules.

    The second tell is repeatability. Run the identical input twice. A language model samples as it generates, so the wording will differ and sometimes the substance will too. A rule engine and a fitted predictor return the same output every time, because there is nothing random in them. This one test separates the generative half of the list from the deterministic half in about ten seconds.

    The third tell is whether a training cutoff exists. Trained models have a date beyond which they know nothing. Ask a vendor when the model was trained and on what period. A product built on rules has no meaningful answer to that question, and the answer you get will show you which it is.

    None of this requires you to be technical. It requires you to ask what the software eats, what it returns, and whether it does the same thing twice. Vendors who are building something real will answer those questions directly. The ones who cannot are usually selling the wrapper.

    What the product asks you for — text, an image, a data feed, or threshold settings — usually names the technology on its own.
    Identical input run twice: varying output means generative sampling; identical output means fixed rules or a fitted model.
    Only trained models have a training cutoff date, so asking for one splits the list immediately.
    Editable thresholds in a settings page are the signature of rule-based automation, however the marketing describes it.
    A language model wrapped around a rule engine is common — the phrasing is generated, the logic is not.
    A vendor who will not describe the mechanism is asking you to trust the label instead of the machine.
    Vendor claim decoder
    Below is the marketing copy from a trading software product page. Do not tell me whether the product is good or worth buying.
    
    Copy:
    [paste the vendor's description here]
    
    For every claim that mentions AI, machine learning, automation or prediction, tell me:
    - Which of these four it most likely describes: a language model, a vision model, a trained numeric predictor, or fixed rules.
    - The exact words in the copy that led you to that conclusion.
    - One specific question I should ask the vendor to confirm it.
    
    If a claim is too vague to classify, write "too vague to classify" rather than guessing. Do not add anything about the company that is not in the copy I pasted.

    When to use — Before a trial or a subscription, when the product page is long on adjectives and short on mechanism.

    A good answer — Several claims marked "too vague to classify" — that is the honest result for most marketing copy, and a decoder that confidently classifies every line is inventing certainty. The questions it produces should be answerable with a sentence, not an essay.

    Six questions that identify the technology

    Run these against any product page or vendor call before you pay. You are not trying to catch anyone out — you are trying to find out which failure mode you will be living with.

    • What exactly do I supply — pasted text, an uploaded image, a live data feed, or numbers in a settings form?
    • If I run the same input twice, do I get the same output? Variation means a generative model is in the loop.
    • Is there a training cutoff date, and what period was the model trained on?
    • Can I see and edit the conditions that produce a signal? If yes, it is rules, not learning.
    • What happens with a stock the system has little history for — does it degrade, or does it output as if nothing were different?
    • Is the number I am shown generated fresh each time, or looked up from something already computed?

    What the Distinction Actually Buys You

    Naming the technology changes what you expect, and expectation is most of the risk. If you know you are holding a language model, you stop asking it where NIFTY closes tomorrow and start feeding it the documents you have not had time to read. If you know you are holding a rule engine, you stop waiting for it to adapt and you review the thresholds yourself when the tape changes character.

    It also changes what you check. Each technology fails in its own direction: the language model invents, the vision model estimates from pixels, the fitted model decays, the rule engine keeps firing after the setup stopped working. A verification habit built for one of those does nothing for the others. You cannot design the check until you know the failure.

    And it changes what you pay for. Once you can recognise a rule engine wearing a language-model coat, the pricing question becomes clear rather than emotional. You may still want it — a well-built rule engine with clean alerts is worth money. But you are buying a known thing.

    Everything else in this module builds on this one separation. The next article opens up the language model in particular, because it is the one you will use most and the one whose mechanics explain almost every strange answer you will ever get from it.

    The label tells you what a vendor wants you to feel. The mechanism tells you what you can actually rely on.

    Working with the label

    Do

    • Ask what goes in and what comes out before asking anything else about a product.
    • Match the task to the technology — compression to a language model, structure and description to a vision model, ranking to a fitted model, mechanical checks to rules.
    • Assume each of the four will fail, and learn the direction each one fails in.
    • Treat a clearly described rule engine as a serious tool, not a downgrade.

    Don't

    • Do not let AI-powered on a product page stand in for an explanation of the mechanism.
    • Do not ask a language model for live prices, or a vision model for exact levels, and then blame the tool for answering.
    • Do not assume that because software learned something once, it keeps learning from your usage.
    • Do not pay a premium for the word without first identifying which of the four is underneath it.

    Common questions

    No. A chat assistant is a language model: text goes in, text comes out, and it has no market data feed and no ability to place orders. An AI trading algorithm usually means a fitted numeric model or a set of coded rules executing against live data. They are different technologies solving different problems, and neither can do the other’s job.

    Knowledge Check

    Question 1 of 3Score: 0

    A screener lets you set a lookback period, a percentage threshold and a volume multiple, then emails you matches. Its homepage calls it AI-powered. What is it, most likely?

    Rohit Singh — Mr. Chartist

    Written By

    Rohit Singh

    Mr. Chartist

    With 14+ years of experience in Indian financial markets, Rohit Singh (Mr. Chartist) is a SEBI Registered Research Analyst, Amazon #1 bestselling author, and the founder of Investology — a premium trading ecosystem trusted by a 1.5 Lakh+ strong community across India.

    INH000015297Full Bio

    Keep reading