Trading &amp; AI

Strong atStrong atStrong at+4 more

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

A large language model is a research analyst, never an oracle. It is genuinely good at three things: compressing long documents into structured summaries, holding a consistent checklist across dozens of names so you don't skip steps under time pressure, and drafting the first version of an analysis you then verify against primary sources. It is genuinely bad at three other things: predicting where a stock trades tomorrow, knowing anything that happened after its training cutoff without being given the data, and telling you when it doesn't know something — it will produce a fluent, confident-sounding answer whether or not the underlying fact is correct. Draw the line there and the tool becomes useful instead of dangerous. The practical test is simple: would you accept this output from a junior analyst without checking their sources? If yes, the model has probably done its job — compression, structure, a first draft. If the output is a specific number, a specific claim about a specific filing, or anything that will move a real position, the model's job is done and yours begins: open the source document and confirm it yourself.

Pro Tip Example

02Beginner

Prompting for Market Research — Getting a Useful Answer, Not a Confident One

Vague prompts produce vague, overconfident answers because the model fills gaps with the most statistically likely completion, not the most accurate one. A research prompt that actually works gives the model four things: the exact scope (which company, which quarter, which metric — not 'analyze Reliance'), the source material to reason over (paste the filing excerpt or transcript, don't rely on the model's memory of it), the output format you want (a table, a five-bullet brief, a comparison — structure forces the model to be specific), and explicit permission to say 'not stated in the source' instead of inventing a number to fill the gap. That last instruction matters more than the other three combined. By default, a model treats 'I don't know' as a failure state to be avoided, so it will produce a plausible-sounding figure rather than admit the source didn't contain one. Telling it directly — 'if the document doesn't state this, say so, do not estimate' — changes that default and is the single highest-leverage sentence you can add to any research prompt.

Pro Tip

9 min

03Intermediate

Filings & Concalls with AI — Turning 300 Pages into Ten Sourced Claims

An annual report, a set of quarterly results and a concall transcript together can run past 300 pages a quarter. Reading all of it for every stock you track isn't realistic, and skipping it is how red flags get missed — this is the exact gap AI compression is built for. The workflow that works: feed the model one document at a time (not the whole bundle at once — accuracy drops as context gets crowded), ask it to extract the SAME fixed set of ten to twelve items every time — revenue and margin trend, debt movement, promoter holding change, related-party transactions, auditor remarks, management's own stated risks, and any one-line item that changed materially versus last quarter — and require a page or line reference next to every extracted claim. The page reference is not optional. It is what turns an AI summary from 'something I read once' into 'something I can defend,' because it gives you a two-minute path back to the source for the handful of claims that actually matter to your decision, instead of forcing a full re-read of a document you already summarized once.

Pro Tip Warning

10 min

04Intermediate

Hallucination & the Audit Trail — Trust, But Verify Everything That Moves Money

A hallucination is a model stating something false with the same fluent confidence as something true — not a glitch or an error message, which would at least be honest about failing. It happens most often in exactly the situations traders reach for AI help: a specific number that wasn't in the source material, a citation or filing reference that sounds plausible but doesn't exist, or a date, quarter or figure quietly transposed from a different company the model has seen more of in training. The defense isn't distrust of the tool — it's a habit: nothing that influences a real position moves forward without an audit trail back to a primary source, every time, no exceptions for outputs that 'seem obviously right.' Build the audit trail as you go, not after the fact. Every AI-assisted note should carry, alongside the conclusion, the specific source it was checked against and the date it was checked — a discipline that costs thirty seconds per claim and is the difference between a research note you can defend six months later and one you can't.

Warning

9 min

05Intermediate

AI for Charts & Pattern Recognition — What a Vision Model Sees, and What It Misses

A vision-capable model can look at a candlestick chart screenshot and describe what it sees — a rising trendline, a cluster of green candles, a level that's been tested three times. What it is doing is pattern-matching pixels to the kind of chart descriptions it saw during training, not measuring price and volume the way a scanner or a human trained on live data does. It has no sense of the actual price scale unless you tell it, no memory of what happened on the candles just off the edge of the screenshot, and no access to volume unless the volume panel is visibly included and legible in the image you gave it. Used correctly, this is a second pair of eyes for a first pass across a large watchlist — flagging charts that plausibly show a pattern worth a closer manual look — never a substitute for actually opening the chart and confirming the setup against real price and volume data yourself. The moment a chart-reading answer is used to size or place a trade without that manual confirmation, the tool has been asked to do a job it was never built for.

Example

06Advanced

Building a Research Workflow — Context Windows, Your Own Notes & Repeatability

A single clever prompt doesn't scale — a repeatable workflow does. The difference between someone who uses AI for research once and someone who uses it every week is a small set of saved assets: a fixed prompt template per task (screening, filing summary, concall summary, thesis draft), a place to paste source material so the model reasons over what you gave it instead of what it remembers, and a running file of your own verified notes that you feed back in as context so the model builds on your prior work instead of starting cold every session. The technical term for the third piece is retrieval-augmented generation — giving a model your own documents as grounding before it answers, rather than relying on what it happened to absorb in training. You don't need custom software to do a version of this: a folder of your past verified summaries that you paste relevant excerpts from into new prompts already gets you most of the benefit, and it compounds — every verified note you add makes next month's research faster and more consistent than this month's.

Pro Tip

10 min

07Intermediate

AI-Assisted Screening & Idea Generation — Casting a Wider Net, Not Picking the Winner

Traditional screeners filter on numbers — P/E under a threshold, ROE above a level, debt-to-equity below a cap. AI-assisted screening adds a layer traditional filters can't: reading qualitative language across dozens of concalls and management commentaries at once to surface phrases and shifts in tone that a pure numeric filter would never catch — a management team suddenly hedging on guidance language, a recurring mention of one input cost across a whole sector's transcripts, an unusual cluster of exchange announcements around one theme in a single week. The output of this kind of screening is a longer, wider shortlist to investigate manually, not a ranked buy list. Feeding it a vague goal like 'find me good stocks' returns generic, low-signal noise; feeding it a specific, checkable pattern — 'across these twenty concall transcripts, which managements mentioned margin pressure from raw material costs in the last quarter but not the one before' — returns something you can act on by opening each flagged transcript and confirming it yourself.

08Intermediate

News & Sentiment at Machine Speed — Filtering the Noise Without Trading the Headline

Markets move on headlines faster than any human can read a full wire feed, and an AI-assisted news filter can watch far more sources than you can — the actual value is triage, not prediction. A well-built filter clusters duplicate coverage of the same event (so twenty headlines about one RBI decision don't read as twenty separate events), flags which stocks or sectors a piece of news is actually relevant to versus tangentially mentions, and separates confirmed facts (an exchange filing, an official announcement) from unconfirmed reports and social-media chatter, which move price on rumor alone and reverse just as fast. The trap is letting a sentiment score substitute for reading the actual news. 'This headline is 80% negative sentiment' is a compression of language, not an assessment of whether the underlying event actually matters to the stock's fundamentals or is noise that the market will forget by the next session — that judgment still has to be yours, made after reading the source the filter pointed you to, not instead of reading it.

Warning

09Beginner

Guardrails, SEBI & the Human in the Loop — Where AI Ends and Advice Begins

In India, investment advice and research recommendations are regulated activities — a SEBI Registered Research Analyst or Investment Adviser stands behind a recommendation with disclosures, a track record and regulatory accountability. An AI model has none of that: no license, no disclosed conflicts of interest, no accountability for a bad call, and no ability to know your personal risk profile, capital, or existing positions unless you type them into a prompt window that forgets everything the moment the session ends. Treating a model's output as regulated investment advice is a category error, not a shortcut — the guardrail here isn't optional caution, it's the actual, correct description of what the tool is and isn't. The healthy mental model: AI drafts, checklists and compresses; the licensed human (you, or the analyst/adviser you follow) verifies, decides and takes accountability for the decision. That division of labor is what every topic in this module has pointed toward — capability line, verification habit, audit trail, human judgment on materiality — and it's the one rule that, if followed, makes every other AI-assisted technique in this curriculum genuinely safe to use.

Pro Tip