Trading & AI
Use AI as a research analyst, never as an oracle — prompting for market work, turning filings and concalls into sourced claims, and the guardrails that keep the output defensible.
Explore the topics
24 of 24 topicsWhat AI Actually Is in a Trading Context
Chat models, vision models, machine-learning predictors and plain rule-based automation all get sold under the same two letters. They are four different technologies with four different failure modes, and knowing which one you are actually holding decides what it is safe to use it for.
How an LLM Actually Works — Tokens, Training Cutoffs and Why It Guesses
You cannot judge an answer until you understand where it came from. Next-token prediction, training cutoffs, context windows and temperature — four mechanics, no mathematics, and together they explain every confident wrong answer you will ever be given.
The Capability Line — What AI Can (and Can't) Do for a Trader
A large language model is a research analyst, never an oracle. Three things it is genuinely good at, three things it is genuinely bad at, and the single question that tells you which side of the line the task in front of you belongs on.
Choosing Your AI Stack — Chat, Vision, Reasoning and Local Models
Four model types, four different jobs. Reaching for the wrong one is the most common reason AI research output disappoints — and the fix is usually a different tool, not a cleverer prompt.
Context, Cost and Limits — What AI Research Actually Costs You
Context windows fill, attention degrades in the middle of long inputs, and a chat that has run all afternoon is quietly worse than the one you started this morning. The constraints that decide how a research session should be structured.
Prompting for Market Research — Getting a Useful Answer, Not a Confident One
Vague prompts produce vague, overconfident answers. Scope, source, structure, and explicit permission to answer 'not stated' — the four-part prompt that separates research you can defend from fiction that reads well.
Filings and Concalls with AI — Turning 300 Pages into Ten Sourced Claims
One document at a time, the same fixed checklist every time, and a page reference beside every number. The workflow that turns compression into something you can still defend six months later.
Hallucination and the Audit Trail — Verify Everything That Moves Money
A wrong answer and a right answer arrive in identical, fluent, confident tone. Tone is therefore not a reliability signal — which means the defence has to be procedural, and it has to cost you thirty seconds every single time.
Data Privacy — What You Must Never Paste Into a Prompt
Broker credentials, client data, live holdings and unpublished research. A prompt box looks like a private notebook and is not one — and for anyone handling someone else's money, the difference is legal, not stylistic.
Building a Research Workflow — Templates, Notes and Repeatability
A clever one-off prompt does not scale. A saved template, used the same way every week and grounded in your own verified notes, compounds — and you can build a working version of it with a folder and no software at all.
AI for Charts and Pattern Recognition — What a Vision Model Misses
A vision model matches pixels to chart descriptions it saw in training. It does not measure price, it cannot see past the edge of your screenshot, and it has no volume unless the panel is legible in the image. That distinction governs every safe use of it.
AI-Assisted Screening — Casting a Wider Net, Not Picking the Winner
Numeric filters still belong in a real scanner running on structured data. What AI adds is the qualitative layer no ratio can reach — a guidance tone that shifted, one input cost complained about across a whole sector's transcripts in a single week.
News and Sentiment at Machine Speed — Filtering Noise Without Trading the Headline
Triage and deduplication are the real value: twenty headlines about one RBI decision are one event, not twenty. A sentiment score is a compression of the language used — never a judgement about whether the event actually matters.
AI-Assisted Coding — Pine Script and Python You Can Actually Audit
Generated code that looks right and quietly repaints is worse than no code at all, because it produces a backtest you believe. How to specify precisely, where AI reliably introduces look-ahead bugs, and how to test before you trust.
An AI Trading Journal — Reviewing Your Own Decisions Without Flattery
Models agree with whatever premise you hand them, which makes the default output useless for reviewing a losing trade. Structuring the prompt so it argues with you instead of echoing you is the entire skill.
AI for Portfolio Review — Exposure, Concentration and Correlation
Ten positions can be one bet wearing ten different names. Surfacing hidden sector overlap and correlated exposure is structural work AI does well — provided you supply the numbers rather than letting it recall them.
Machine Learning vs LLMs — Where Prediction Genuinely Fits
Language models and predictive machine learning are different technologies solving different problems. Conflating them is why most retail 'AI trading' projects fail before the first backtest — and where supervised learning honestly does earn its keep.
Features, Labels and Overfitting — Why Most ML Backtests Are Fiction
Look-ahead bias, survivorship bias and a model that simply memorised the past. Three defects that each produce a beautiful equity curve and none of the returns, and how to test for every one of them honestly.
Backtesting an AI-Generated Strategy — The Honest Test
The model wrote it in thirty seconds. Proving it is a real edge rather than curve-fitted noise takes considerably longer — out-of-sample splits, walk-forward windows, and Indian transaction costs modelled properly rather than assumed away.
Model Drift and Silent Failure — When the Tool Changes Under You
Nothing throws an error. A regime turns, a provider updates the model behind the same name, a data feed starts arriving late — and the output just gets quietly worse. A process that never checks is a process that never notices.
AI Agents and Automation — When to Let a Model Take an Action
Reading a filing is reversible. Placing an order is not. Irreversibility — not cleverness, not accuracy — is the right test for how much autonomy is defensible, and the kill switches have to exist before the autonomy does.
Evaluating an AI Trading Tool Before You Pay — Claims and Red Flags
An accuracy percentage with no sample size, no period and no cost assumption is marketing, not evidence. The questions a genuine vendor can answer in one line, the SEBI registration check that comes first, and the claims that should end the conversation.
Guardrails, SEBI and the Human in the Loop — Where AI Ends and Advice Begins
In India, investment advice is a regulated activity with a registered, disclosed, accountable person behind it. A model has none of that and cannot acquire it. The division of labour that keeps every other technique in this module safe to use.
Your First 30 Days — A Practical Plan That Doesn't Break Your Process
One task, one template, four weeks. The capstone that turns twenty-three articles of theory into a habit you actually keep — with the verification checks built in from week one rather than bolted on after the first expensive mistake.