Read this before anything else
This session is a survey of research-assistant uses of AI in markets - nothing more. It is not investment advice. It is not a trading system. It is not a stock-picking or recommendation engine. We will summarise public filings, compare public figures, and screen on criteria you define - all things that help a human think faster. We will never tell you, or anyone, what to buy, sell, or hold.
The deep material - portfolio construction math, factor models, algorithmic and high-frequency trading, order execution - is deliberately deferred to future dedicated courses. It is a craft that deserves its own program, taught properly, not squeezed into one evening. Tonight is the honest, bounded version: AI as a fast, tireless reader and organiser of public information, with a person owning every single conclusion.
Assist, never advise: the boundary 6 min live
Every use of AI in markets sits on one side of a hard line. On the left, it helps a human read, organise, and compare public information faster - assist. On the right, it makes or implies a decision - advise. The left is fair game with discipline. The right is off-limits for a tool, full stop.
LiveWhich side of the line am I on?3 min▶
The test is simple. Ask: does the output help a person read and reason, or does it stand in for their judgement? If it is a summary, a table, or a filtered list that a human then evaluates, you are assisting. The moment the output implies "do this", you have crossed into advice - and a tool has no business there.
- Assist keeps a human in the loop. The AI drafts; the analyst reads, checks the sources, and forms the view. The AI never gets the last word.
- Advise removes the human. A recommendation, a personalised "should", an order placed automatically - these hand the decision to a machine. That is the line we do not cross.
- When in doubt, add a human step. If a workflow feels like it is drifting toward a decision, insert an explicit "a person reviews and signs" gate. That gate is not bureaucracy - it is the rail.
Legit research-assistant uses 6 min live
Stay on the left of the line and there is real, everyday value here. The through-line is always the same: AI does the fast reading and organising of public information; you do the verifying and the thinking. Here is where it earns its keep - and the check that makes each one safe.
LiveFive uses, each with its verify step3 min▶
| Assist use | What you feed it | Verify step - non-negotiable |
|---|---|---|
| Summarise a filing | public 10-K, 10-Q, annual report | open cited sections, confirm the summary matches |
| Summarise an earnings call | public transcript excerpt | trace each bull/bear point back to the words said |
| Comparison table | public reported figures you supply | tie every cell to the source number and period |
| Screen on your criteria | a list plus rules you define | re-check the data behind each pass and fail |
| Sentiment on public text | public commentary, press, transcripts | read the passages; sentiment is a signal, not a fact |
Two of these are draft-a-memo and screen-a-list, and both hide the same trap: the model will confidently produce figures. Screening is only ever as good as the data behind it, so define the criteria yourself and then verify the numbers each pass and fail rests on - do not let the tool source the data and apply the rule unchecked.
Self-studyPublic data only - the privacy rail, applied to markets2 min read▶
Everything in this session runs on public information: published filings, released transcripts, exchange data, company press. That is not an accident - it is the rail. The instant you are tempted to feed something non-public into a tool, stop.
- Public filings and transcripts - a company's own 10-K, a posted earnings call, published market data - are fine to summarise and compare. They are already public.
- Draft research, internal notes, client holdings go only into an enterprise tool that does not train on your data, and even then with care. A client's portfolio is confidential; treat it that way.
- Material non-public information never touches a tool at all. That is not a privacy preference, it is the law - covered head-on in Part 3.
The hard lines 5 min live
The assist side is safe only if you respect the boundaries around it. These are not style preferences - several are legal lines with real consequences. Know them cold, because "the AI did it" is never a defence. The analyst owns the call.
LiveThe five you must not cross3 min▶
- No personalised advice. An AI telling a specific person what to buy, sell, or hold for their situation is regulated advice a tool is not licensed to give. Keep outputs general, sourced, and framed as research a human evaluates.
- No automated trading. This course does not build order-placing systems, and you should not wire an LLM to a trading account. Execution without a human gate is exactly the decision-making we bar. Trading mechanics are a future dedicated course.
- MNPI and insider-trading rules. Material non-public information must never go into an AI tool, and AI must never be used to act on it. This is the sharpest legal line in finance - insider trading is a crime. Public data only, always.
- SEC "AI-washing". Regulators are scrutinising firms that overstate their AI capability. Do not claim an AI-driven process, edge, or product it does not have. Describe what the tool actually does - assist, not decide.
- Backtest illusion and overfitting. A strategy that looks brilliant on past data often means the model memorised noise, not signal. Impressive backtests are the easiest way to fool yourself; healthy skepticism, out-of-sample thinking, and a human's judgement are the guard. The deep version of this lives in the future markets and trading courses.
The confident price that was never real. An analyst asked a chatbot for a company's latest reported quarterly revenue to seed a comparison. It returned a precise figure with a decimal - and it was wrong, invented from patterns, not the filing. Pasted into a memo it would have poisoned every downstream number and any conclusion drawn from it. One check against the 10-Q caught it in under a minute. The model assisted; the human, by verifying, decided what was true.
Structure a public earnings-call excerpt - then verify it ★ 13 min · everyone builds
Paste an excerpt from a public earnings-call transcript and turn it into a structured read: bull points, bear points, and open questions for management. The value is the structure and speed - the discipline is tracing every point back to the words that were actually said.
Grab a public earnings-call transcript - most companies post them in investor relations, and several sites publish them free. Copy a real excerpt of a few paragraphs. It is public, so any chat tool is fine (rail one).
Paste the excerpt and set the ground rule: work only from this text, quote the line behind each point, and never add outside knowledge or a view on the stock.
Ask for a structured summary - bull points, bear points, and questions a human analyst should put to management - each tied to a quoted line from the excerpt.
Now the real work: for each bull and bear point, find the quoted line in the transcript and confirm it actually supports the framing. Watch for the model editorialising beyond what was said.
Delete any point whose quote does not hold up, and flag any figure the model mentions to check against the filing. The output is a draft for a human, not a verdict on the company.
Build a comparison table from supplied public figures ★ 8 min · build your own
Compare three companies on figures you supply from public sources. The point is to see AI as a fast organiser of numbers you already trust - not a source of numbers. So we feed it verified figures, then run a verify pass anyway, because even copied numbers can get transposed.
Pull a few public metrics for three companies from their own filings or releases - revenue, operating margin, and headcount, say - and write them down with the source and period next to each.
Give the model only those figures and ask it to build a clean, aligned comparison table. Tell it explicitly not to add, estimate, or "improve" any number.
Ask for one short, neutral observation about the differences - purely descriptive, no view on which is a better investment. Description, not recommendation.
Verify pass: check every cell against the figures you supplied. Confirm nothing was changed, transposed, or invented, and that periods line up. Any mismatch, fix it and note where the tool drifted.
The transposed margin. An IR team used a chat tool to tidy a peer comparison from figures they had pulled by hand. The table looked perfect - except one company's operating margin had been swapped with its net margin, a subtle slip the model made while "aligning" the data. The final-line "which number went in which cell" check surfaced it instantly. The lesson stuck: even when you supply the numbers, verify the output. The tool organised; the humans owned the table.
Try it yourself - this week ◐ 30-45 min total
- Finish both demos if you did not complete them live. The verify pass is the point - never skip it, and keep any figure the tool got wrong as evidence for why.
- Take one public earnings call you would normally read end to end and run the structured-summary prompt on an excerpt. Time the verified read against reading the whole thing, and note where the model editorialised.
- Write the assist/advise boundary in your own words, in one sentence, and stick it somewhere you will see it. If a colleague asks AI to "pick a stock", you should be able to explain the line in ten seconds.
- Find one market figure an AI gives you that turns out to be wrong when you check the primary source. You will not have to look hard - and it is the most convincing lesson in this whole course.
- Note which uses on your desk are genuinely assist, and flag any that quietly drift toward advice or unchecked figures. Those are the ones to redesign with a human gate.
Official sources covered
This session surveys markets and investing AI at research-assistant depth only, drawing on the Columbia and Wharton finance material and the SEC/FINRA marketing guidance. Portfolio construction math, factor models, and trading mechanics are deferred to future dedicated courses - said honestly. This session covers:
Three questions before you go 🎯 ◐ 90 seconds
1 · Which of these is a legitimate assist use of AI on a research desk?
Summarising public text into a structured draft a human verifies is assist. Personalised advice and automated trading are on the advise side of the line - off-limits for a tool.
2 · An AI gives you a company's latest revenue figure, stated confidently. What do you do first?
Markets-data hallucination is severe. Models invent plausible figures with total confidence. Every number must be matched to the primary source before it goes anywhere - treat unverified figures as wrong.
3 · A colleague wants to paste a not-yet-public earnings number into a chatbot to "get a head start". What is the issue?
MNPI is the sharpest legal line in finance. Material non-public information must never touch an AI tool, and using AI to act on it is insider trading - a crime. Public data only, always.