What is trading?
CROSS-CUTTING: COURSE 1 | LESSON 2
Prompting fundamentals for market research
Learning objectives
Build prompts using the four-part pattern: role, context/data, task, output format.
Use iteration and follow-up questions to sharpen a first draft answer instead of accepting it.
Apply a verification pass to every AI research output before it touches a trading decision.
Garbage prompt in, confident garbage out
Most disappointing AI sessions fail in the first message. "Tell me about the euro" produces a Wikipedia-flavoured essay because that's exactly what was asked for. An AI assistant has no idea who you are, what you already know, what data you're looking at or what you'll do with the answer — unless you tell it. Prompting is not a magic incantation; it's just briefing a very fast, very literal research assistant who never asks clarifying questions unless invited to.
The good news: one reusable pattern covers almost everything a trader needs.
The four-part prompt pattern
1. Role — tell the model what hat to wear. This tunes the vocabulary, depth and skepticism of the answer.
2. Context and data — paste what it needs to know: your experience level, the instrument, and crucially the actual data, because (as Lesson X1.1 established) the model has no live feed and a stale memory.
3. Task — one specific, bounded job. "Summarise what changed versus the previous statement" beats "analyse this".
4. Output format — name the sections, the length, the structure. Models follow formatting instructions well, and a fixed format makes hallucinations easier to spot because every claim has a place you expect to find it.
Compare the two versions:
Weak: "What do you think about the Fed?"
Strong: "You are a macro research assistant. I trade EUR/USD on 1–4 hour timeframes; I understand the basics of rate policy. Below I've pasted the full text of today's FOMC statement and the one from the previous meeting. Task: list every sentence that changed between the two, and for each change explain in one line why a currency trader might care. Format: a two-column list — 'What changed' / 'Why it might matter'. If a change is ambiguous, say so rather than guessing. Do not tell me what the market will do."
The strong version does four things the weak one can't: it supplies current data (pasted statements), bounds the task (diff the two texts), fixes a format (two columns), and pre-blocks the failure mode (no forecasting).
Notice the last two sentences. Explicitly instructing the model to flag uncertainty instead of filling gaps, and telling it what not to do, are two of the highest-value lines you can put in any trading prompt. Make them a habit:
- "If the data I've provided doesn't cover something, say 'not in the provided data' — do not estimate."
- "Do not give buy/sell recommendations or price predictions."
Iteration: the first answer is a draft
Treat the first response as an opening bid, not a verdict. Good follow-ups do more work than good first prompts:
- "Which of those points is weakest, and why?" — forces the model to critique itself, often surfacing the hand-waviest claim.
- "You said services inflation 'remains elevated' — quote the exact sentence from the statement I pasted." — grounding check. If it can't quote it, it invented it.
- "Now argue the opposite case." — LLMs are agreeable by default; they mirror your framing. If you sound bullish, the answer leans bullish. Deliberately requesting the counter-case corrects for this sycophancy.
- "Rewrite that as three bullets I could paste into my journal." — compression for actual use.
A practical session for market research therefore looks like: brief → draft → challenge → ground → compress. Five messages, ten minutes, and you end with something checkable instead of an essay.
The verification pass — non-negotiable
Everything from Lesson X1.1 applies double for research prompts, because research output feels authoritative. Before any AI-generated research influences a decision, run this checklist:
- Every number (rates, dates, percentages, price levels) — check against the primary source: the central bank's site, the official statistics release, your own platform.
- Every quote — search the source document you pasted. If the "quote" isn't there verbatim, discard the claim built on it.
- Every claim about "recent" events — is it actually recent, or a knowledge-cutoff fossil? If you didn't paste it in this session, treat it as stale.
- The overall direction — did the model just mirror the bias in your question? Re-ask neutrally ("summarise both bullish and bearish readings") and compare.
Example — grounding check in action: You: "You listed 'the RBA dropped its tightening bias' as a change. Quote the exact sentence from each statement showing this." AI: "Previous statement: '…further tightening may be required…'. Current statement: I cannot find an equivalent sentence in the text you provided — the phrase appears to have been removed, but I should note I inferred 'dropped its bias' from the absence, not from an explicit sentence."
That's a useful answer: the model has separated what's in the text from what it inferred. When a model won't or can't make that separation, downgrade everything it said.
Prompts as reusable assets
Once a prompt works, don't retype it — save it. A tested prompt with a role, a data slot ("paste statement below"), a task and a format is a small piece of intellectual property you refine over months. That's exactly what the AI Lab skills library is: eight pre-built, safety-railed prompt packs (trade-journal analyst, morning macro briefing, strategy stress-tester and more) that you copy, paste and feed with your own data. Lessons X1.3–X1.5 walk through the three most important ones.
One last calibration: prompting skill raises the usefulness of answers, not their truthfulness. A beautifully briefed model still hallucinates; it just hallucinates in a tidier format. The verification pass never comes off the checklist.
Key takeaways
Brief the model like a literal-minded assistant: role, context + pasted data, one bounded task, explicit output format.
Always include the two guard lines: flag missing data instead of inventing it; no predictions or buy/sell calls.
Iterate: challenge the draft, demand exact quotes from your pasted source, and ask for the opposite case to counter the model's agreeableness.
Verify every number, quote and "recent" claim against the primary source before it touches a decision.
Save working prompts — they become reusable skills, which is the basis of the rest of this course.