MANDATORY DISCLAIMER - READ FIRST. This article is for educational and informational purposes only. Nothing here is financial advice, investment advice, a solicitation, or a recommendation to buy or sell any security, option, future, or other instrument. ChatGPT and other large language models (LLMs) hallucinate. They will state figures that sound plausible and are wrong. They will fabricate ticker-specific news, misquote earnings transcripts, and confidently assert relationships in the data that do not exist. Always verify any number, quote, or fact against a primary source (the SEC filing, the company press release, the broker statement) before you act on it. Past performance does not guarantee future results. Trading options and leveraged products involves substantial risk of loss. Consult a registered investment professional and review your own objectives, risk tolerance, and tax situation before trading.
The short answer: ChatGPT will not make you a better trader by predicting markets. It will make you a more disciplined trader if you use it as a structured journal, a checklist engine, a scenario tester, and an emotional-audit mirror. This guide gives you 10 prompts that keep AI on the process side of the line, plus the regulatory and ethical facts you need to stay compliant in 2026.
I write this in mid-2026, after three full years of retail traders pasting earnings transcripts into chatbots and asking whether to buy. Some of them got rich. Many of them got burned. The ones who got rich almost always used AI the way a pilot uses a checklist, not the way a lottery player uses a fortune cookie. That’s the only use of ChatGPT for trading that I can stand behind in writing.
Below is the playbook: what AI actually does well in trading, the ten prompts I keep coming back to, the use cases (journaling, post-trade review, earnings summaries, sector screening, risk sizing, scenario planning, options comparison, tax-loss harvesting), the regulatory line (FINRA Rule 3110, FINRA Rule 2210, SEC Marketing Rule, OpenAI’s own usage policies, and the FCA’s UK stance), and the questions everyone actually asks in 2026.
TL;DR
- Process, not prediction. ChatGPT is a checklist, journal, and scenario tester - not a stock picker.
- Regulators are watching. FINRA Rule 3110 supervision, FINRA Rule 2210 communications, and the SEC’s Marketing Rule all apply to AI-generated content. The SEC/FINRA/NASAA joint AI fraud alert (Jan 25, 2024) is still the baseline.
- **OpenAI itself prohibits using ChatGPT to “provide tailored advice that requires a license” without a licensed professional and prohibits “automation of high-stakes decisions” in financial activities without human review (Usage Policies, effective Oct 29, 2025).
- The 10 prompts in this guide cover: pre-trade plan, risk sizing, bull/bear case, earnings summary, sector screen, post-trade review, emotional audit, scenario stress-test, options strategy comparison, and tax-loss harvesting plan.
- Never paste trade secrets, customer data, or non-public information into a public chatbot. Prompt-injection and data-exfiltration risks are real and regulator-acknowledged (FINRA, March 6, 2026).
The 2026 Reality: AI in Retail Trading
AI is now part of the trading stack. It’s just not part of the trade signal.
The numbers paint a clear picture. The FCA reports that 75% of UK financial firms have adopted some form of AI, and 84% have a named individual accountable for the firm’s AI approach, per the FCA’s “AI in financial services” page in 2026. FINRA’s resource “How FINRA Member Firms Use GenAI” (July 14, 2025) catalogs the most common member-firm GenAI use cases: drafting research summaries, locating policies, surveilling communications, summarizing earnings transcripts. The 2026 FINRA Annual Regulatory Oversight Report dedicates an entire GenAI topic to how firms are embedding these tools (December 9, 2025 release).
Retail adoption has tracked the institutions, just messier. OpenAI’s ChatGPT usage policies (effective October 29, 2025) explicitly prohibit “the provision of tailored advice that requires a license, such as legal or medical advice, without appropriate involvement by a licensed professional” and “automation of high-stakes decisions in sensitive areas without human review,” with “financial activities and credit” listed as one of those sensitive areas. Read that twice. The company that makes the tool is telling you: do not let the tool make the call.
Brokerages have started shipping AI features of their own. Schwab, Fidelity, Interactive Brokers, and Robinhood have all rolled out or expanded AI-assisted research summaries, earnings digests, and (in some cases) paper-trading strategy builders during 2025 and 2026. I won’t name specific product feature names in this article because vendor features change quarter to quarter and I don’t want to misstate them. The point stands: even the platforms are using AI on the process side, not on the buy/sell side.
What hasn’t changed: regulators still expect humans in the loop. FINRA’s Regulatory Notice 24-09 (June 27, 2024) is unambiguous: “FINRA’s rules - which are intended to be technology neutral - and the securities laws more generally, continue to apply when member firms use Gen AI or similar technologies in the course of their businesses.” That means FINRA Rule 3110 (Supervision), FINRA Rule 2210 (Communications with the Public), the SEC’s Marketing Rule (Rule 206(4)-1), and Reg BI’s obligations all keep applying to AI-generated content the same way they apply to content written by a person.
What you can take to the bank in 2026: the trader who uses ChatGPT to plan, journal, and audit is operating in the regulator’s good graces. The trader who uses ChatGPT to generate a stock recommendation they then post on social media is asking for a FINRA or SEC inquiry.
How to Use This Guide
Every prompt below follows the same template so you can copy-paste and adapt:
- The full prompt text. Use it as-is or modify for your instrument (stocks, ETFs, options, futures).
- Use case. Which part of your process it plugs into.
- Before/after example. A short worked example with redacted detail.
- Risk tip. A specific guardrail that prevents the prompt from doing something dumb.
A note on privacy before we start: never paste account numbers, social security numbers, brokerage statements, or non-public information into a public chatbot. FINRA’s prompt-injection guidance (March 6, 2026) explicitly flags that GenAI tools “cannot reliably distinguish between legitimate operational instructions and…hidden malicious commands.” If a document you ingest mentions a friend’s account, an employer’s confidential code, or anything covered by a non-disclosure agreement, redact first.
Use Case Matrix: Which Prompt for Which Job
This is the comparison table. Pick the row that matches what you’re trying to do tonight.
| Use case | Prompt goal | Data input you provide | What the prompt outputs | Top risk to manage |
|---|---|---|---|---|
| Pre-trade plan | Force a written thesis before entry | Ticker, setup type, account size, risk % | Thesis, entry trigger, invalidation, target, position size | ChatGPT adding a “buy” recommendation you didn’t ask for |
| Position sizing | Turn account size and stop distance into shares | Account equity, risk %, entry, stop | Share count, dollar risk, R-multiple target | Calculator math errors on small numbers |
| Bull/bear case | Steelman both sides | Ticker, your current thesis | 3 bullish reasons, 3 bearish reasons, what would flip you | Confirmation bias dressed up as analysis |
| Earnings-call summary | Pull the 3 numbers that actually matter | Pasted transcript or 10-K excerpt | Revenue, margin, guidance, sentiment, red flags | Hallucinated quotes or figures |
| Sector screen checklist | Build a repeatable stock screener | Sector, factors you care about | Filter list, weight rationale, sample names | Overfitting on a backtest that won’t repeat |
| Post-trade review | Catch what you missed in the trade | Trade log entry: entry/exit, reason, P&L | What went right, what went wrong, pattern tag | Self-justifying narrative |
| Emotional audit | Spot revenge trading and tilt | Journal entries or feelings you describe | Behavioral pattern call-out, circuit-breaker prompt | Privacy leak if you over-share |
| Scenario planning | Stress-test a thesis | Position, base case, three macro scenarios | P&L and action per scenario, hedging ideas | Treating scenarios as predictions |
| Options strategy comparison | Compare two structures side-by-side | Underlying, view, two strategy structures | Risk graph, breakevens, max loss/gain, Greeks, decay | Greeks oversimplified into a single number |
| Tax-loss harvesting plan | Map losses to gains without crossing wash-sale | Lots with cost basis, intent | Match table, wash-sale warnings, replacement ideas | Wash-sale violations, fee math errors |
The 10 Prompts
Each prompt is structured as: role (who the AI is acting as), task (what to do), constraints (what not to do), inputs (what to paste in), and output format (the shape of the answer). This is the framework the FINRA Foundation’s “Understanding Generative AI and Prompt Injection Fundamentals” guidance (March 6, 2026) implicitly endorses when it tells firms to put guardrails around prompts.
Prompt 1 - Pre-Trade Plan Checklist
Use case: Pre-trade journal / decision journal Why this works: The hardest part of any trade is writing the thesis before you click buy. The prompt forces the boring prep.
You are a trade journal reviewer, not an investment advisor. Do not recommend buying, selling, or holding any security. I will paste details of a trade I am considering. Ask me to fill in any missing field. Then output a structured plan with: (1) Thesis in one sentence; (2) Entry trigger - what price or signal gets me in; (3) Stop loss - exact price and the rule that triggers it; (4) Target - price and the rule for taking profit; (5) Position size in shares and as a percent of account equity; (6) Time stop - when I close the trade if nothing happens; (7) What would make me wrong; (8) Three things I am uncertain about. After the plan, list three things I should check against a primary source before placing the order. Do not predict price movement.
Before / after:
Before: “Thinking about NVDA. Feels like it wants to run. Earnings next week.”
After: A structured plan with a written thesis, an invalidation level, a hard position size, and three explicit uncertainties. You may decide the trade is too thin to take. That’s the win.
Risk tip: If the model adds a “verdict” line (e.g., “looks bullish”), delete it. That is the model role-playing a financial advisor. The prompt above tells it not to.
Prompt 2 - Position Sizing Calculator
Use case: Risk sizing / R-multiple planning Why this works: Most retail blow-ups are sizing errors. A consistent 1% rule beats most strategies.
You are a position-sizing calculator. Do not give trading advice. Inputs I will provide: account equity in dollars, risk percent per trade, entry price, stop price, target price. Compute and output: (1) Dollar risk per share = entry − stop; (2) Shares to buy = floor((equity × risk%) / dollar risk per share); (3) Total position size in dollars; (4) Position size as percent of equity; (5) R-multiple to target = (target − entry) / (entry − stop); (6) Expected reward if target hits. Round shares down to whole numbers. If the result is zero or negative shares, say so and stop. If position size exceeds 20% of equity, warn me. Show your math.
Before / after:
Before: “Buying 200 shares at $50, stop at $48, account is $40,000.”
After: “Dollar risk per share: $2. Risk at 1% ($400) → max 200 shares. R-multiple to target $56 = 3.0R. Reward if target hits: $1,200.” A second iteration with risk tightened to 0.5% produces 100 shares, which fits the prompt’s math without you having to think.
Risk tip: Always double-check the model’s arithmetic by hand for two trades out of ten. LLMs can do math, but they are not calculators. Use a spreadsheet for actual order entry.
Prompt 3 - Bull / Bear Case Steelman
Use case: Decision journal / cognitive debiasing Why this works: The single best antidote to confirmation bias is writing the opposite case.
You are a debiasing assistant. Do not recommend an action. I am considering [TICKER] and currently believe [YOUR THESIS IN ONE SENTENCE]. Produce: (A) Three distinct reasons the thesis could be right, with the strongest single piece of evidence for each. (B) Three distinct reasons the thesis could be wrong, with the strongest single piece of evidence for each. (C) One fact or data point that, if it appeared next week, would flip my view. (D) One fact or data point that would strengthen my view. (E) Two questions I should ask a registered investment professional before sizing this position.
Before / after:
Before: “Long XOM, oil’s tightening, OPEC cutting.”
After: Three reasons it works (supply cuts, demand resilience, valuation gap). Three reasons it doesn’t (demand destruction in China, SPR refilling, capex discipline breaking). One flip signal (China PMI sub-48 for two months). Two professional questions (margin sensitivity to crude, hedge accounting on 2027 production).
Risk tip: If you can’t honestly write the bearish case, you don’t have enough information to take the trade. That’s the prompt’s verdict.
Prompt 4 - Earnings-Call Summary (With a Hallucination Guard)
Use case: Earnings review / fundamental note Why this works: Earnings transcripts are long. The model is great at summarizing - but terrible at quoting numbers exactly. The guardrail forces verification.
You are an earnings transcript summarizer, not an analyst. I will paste an earnings press release or 10-Q excerpt. Output: (1) Headline numbers reported: revenue, EPS, gross margin, free cash flow - only if explicitly stated in the text; (2) Year-over-year change you can compute from the text; (3) Guidance range if given; (4) Two notable quotes, in quotation marks, with the speaker’s name; (5) Two risks management called out; (6) Two things management did not say but I might assume; (7) One number I should verify against the original SEC filing before using in any decision. Do not infer numbers. If a figure is unclear, write “unclear in source” and stop.
Before / after:
Before: Skim the transcript. Remember 30% of it. Misquote guidance two days later.
After: Five explicit numbers, two direct quotes with attribution, two flagged omissions, and one number flagged for SEC EDGAR verification. The “did not say but I might assume” line is the most useful - it catches the analyst’s voice pretending to be the CEO’s.
Risk tip: Always cross-check the model’s numbers against the actual filing on SEC EDGAR (sec.gov/edgar). The hallucination rate for specific dollar figures in transcripts is non-trivial.
Prompt 5 - Sector Screening Checklist
Use case: Idea generation / watchlist maintenance Why this works: Screening prompts fail when they’re too greedy. The checklist forces narrow filters.
You are a stock screener builder, not a stock recommender. I want to find stocks in [SECTOR] that meet: (1) Market cap between [MIN] and [MAX]; (2) Revenue growth above [X]% in the last reported fiscal year; (3) Operating margin above [Y]%; (4) Net debt / EBITDA below [Z]; (5) Average daily dollar volume above [V] (liquidity filter). Output: (A) The filter as a JSON or text checklist I can paste into a screener; (B) Three example names that might fit, but verify against a primary source before relying on them; (C) Three risks of using a backtested screen in the next 12 months; (D) Three sectors where this screen historically fails (style drift). Do not rank or recommend any specific stock.
Before / after:
Before: “Find me mid-cap AI infrastructure stocks.”
After: A 5-line screener, three example tickers flagged for verification, three failure modes (mean reversion in momentum factors, index inclusion effects, sector reclassification), and three sectors where the screen historically breaks.
Risk tip: LLMs can hallucinate ticker symbols and financials. Always run the screen through a real screener (Finviz, TradingView, your broker’s screener) and verify financials on SEC EDGAR or the company’s investor relations page.
Prompt 6 - Post-Trade Review (Blame-Free Autopsy)
Use case: Trade journaling / pattern detection Why this works: The point of a post-trade review is not to feel good or bad. It’s to tag the trade so the next 100 trades improve.
You are a trade journal auditor. I will paste a closed trade: ticker, side (long/short), entry date and price, exit date and price, position size, the reason I took it, the reason I exited, and P&L. Output: (1) What went right - cite a specific decision, not luck; (2) What went wrong - cite a specific decision; (3) Process tag: planned entry, unplanned entry, scaled in, scaled out, stopped out, target hit, time stop, news-driven exit, emotional exit; (4) Pattern tag: trend, mean-reversion, breakout, earnings, macro; (5) One thing I should add to my pre-trade checklist based on this trade; (6) One thing I should remove from my checklist. Do not judge the outcome. Do not tell me to trade differently next time - only show me the data.
Before / after:
Before: “Lost $400 on TSLA. Should have held.”
After: “What went right: respected the stop. What went wrong: entered before the news event with no plan for the gap. Process tag: news-driven exit. Pattern tag: event-driven. Add to checklist: no new entries 24 hours before scheduled catalysts. Remove from checklist: ‘trust the trend’ as a stand-alone reason.”
Risk tip: Keep this output in a structured journal (Notion, Obsidian, even a spreadsheet). Tag counts over 30+ trades reveal your real edge - or absence of one.
Prompt 7 - Emotional Audit (Tilt & Revenge Detector)
Use case: Behavioral risk / mental game Why this works: Most retail losses come from emotional decisions, not bad analysis. The model can spot the words faster than you can.
You are a behavioral finance coach. I will paste journal entries from the last 5 trading days. Do not give investment advice. Identify: (1) Any signs of revenge trading (entering a new position within 30 minutes of a loss, position size larger than my usual); (2) Any signs of FOMO (mentions of specific tickers trending on social media, urgency language); (3) Any signs of disposition effect (reluctance to take a loss, premature profit-taking); (4) Sleep, health, or external-stress mentions that correlate with worse decisions; (5) Three circuit-breaker rules I should write down for myself before the next session. Be specific. Quote my own language back to me where helpful.
Before / after:
Before: Stare at the screen after a loss, reopen the position at 1.5x size, lose again.
After: A pattern: “On [date], you used the word ‘stupid’ three times within an hour of a loss. Your next entry was 150% of your average size. On [date 2], the same pattern appeared.” Three circuit-breaker rules: (1) No new positions for 2 hours after a loss > 1R, (2) Position size capped at 1R for the rest of the day after two losses, (3) Phone in another room during the first 30 minutes of the session.
Risk tip: Don’t share anything with the model you wouldn’t put on a postcard home. Use the prompt on a journal you’ve already cleaned of identifying information. If you work at a broker-dealer, do this only on a personal device, not on a firm-managed endpoint.
Prompt 8 - Scenario Stress-Test (Three Worlds)
Use case: Risk management / scenario planning Why this works: A single base-case P&L is a lie. You need to know what happens in three different worlds.
You are a scenario planning assistant, not a forecaster. I will describe my position and three macro scenarios. For each scenario, output: (1) Estimated P&L range, with explicit assumptions about price, volatility, and time-to-resolution; (2) One action I would take (roll, hedge, add, trim, close); (3) One signal that the scenario is unfolding; (4) One signal that the scenario is no longer credible. Do not predict which scenario is most likely. Do not recommend an overall action.
Before / after:
Before: “I’m long semis, paid $180.”
After: Scenario A (cyclical recovery): P&L +12-18%, hold, signal: ISM back above 50. Scenario B (rates stay higher for longer): P&L −8 to −15%, hedge with SPY puts, signal: 10-year above 4.75%. Scenario C (China shock): P&L −20 to −30%, trim to half size, signal: USD/CNY above 7.40.
Risk tip: Use this prompt quarterly. The world changes; your scenarios should too.
Prompt 9 - Options Strategy Comparison (Side-by-Side)
Use case: Options strategy review / payoff comparison Why this works: Two option structures can look similar and behave very differently. The matrix exposes it.
You are an options comparison assistant, not an options strategist. I will provide two strategies on [UNDERLYING] with strikes, expirations, and net debit/credit. Output a comparison table with columns: Max gain, Max loss, Breakeven prices, Probability of profit at expiration (estimate only, flag as estimate), Theta cost per day, Vega exposure at entry, key risk to monitor. After the table, list three reasons strategy A might be chosen over B, and three reasons B might be chosen over A. Do not recommend one strategy over the other.
Before / after:
Before: “Which is better, the 95/100 bull put spread or the 100 call?”
After: A side-by-side table with max gain, max loss, breakevens, theta, vega. Three reasons to prefer A (defined risk, theta-positive, lower Vega). Three reasons to prefer B (higher upside, no upside cap, easier to adjust).
Risk tip: Greeks change daily. Re-run this comparison with current IV, not last week’s. For educational resources, the Options Industry Council and Cboe’s Options Institute are the free references (optionseducation.org and cboe.com/optionsinstitute).
Prompt 10 - Tax-Loss Harvesting Plan (Wash-Sale Aware)
Use case: Year-end tax planning / loss realization Why this works: Tax-loss harvesting is one of the few “free” returns in investing - but the wash-sale rule (IRC §1091) and the 30-day window can turn a win into an audit flag.
You are a tax-planning research assistant, not a tax advisor. I will paste a list of lots with ticker, cost basis, current price, and holding period. Output: (1) Lots with unrealized losses, ranked by dollar loss; (2) For each loss lot, three “replacement” candidates that are NOT substantially identical to the original holding, with a one-line rationale; (3) Wash-sale warnings - list any purchases of the same or substantially identical security inside the 30-day window before or after each sale; (4) Estimated short-term vs. long-term loss split; (5) Three questions I should bring to a CPA before executing. Do not give specific tax advice. Do not estimate my exact tax savings.
Before / after:
Before: “Sell the losers at year-end. Buy them back in 30 days.”
After: A ranked list of loss lots, three non-substantially-identical replacements per lot (e.g., sell an S&P 500 ETF loss and buy a total-market ETF), explicit wash-sale flag if any purchases fall in the 30-day window, short-term/long-term split, three CPA questions.
Risk tip: The wash-sale rule (Internal Revenue Code §1091) and the IRS Publication 550 are the only authoritative sources. AI is helpful for organization, not for IRS interpretation. Confirm everything with a CPA or enrolled agent.
A Word on What ChatGPT Is Doing Under the Hood
You don’t need a PhD to use these prompts, but a short explainer helps.
A “large language model” (LLM) is a deep-learning system trained on huge amounts of text to predict the next token in a sequence. The Bloomberg team published the first major finance-specific LLM - BloombergGPT, a 50-billion parameter model trained on 363 billion tokens of financial text - in a December 2023 update to their March 2023 paper (arXiv:2303.17564). It outperformed general-purpose LLMs on financial tasks without sacrificing general performance.
Why does that matter for you? Because it tells you the technology works on finance text. But the consumer tool you use - ChatGPT, Claude, Gemini - is general-purpose. It will get the shape of a financial task right (the structure of a 10-K, the categories of an earnings call) more often than not. It will get the specific numbers wrong more often than you want. That’s why every prompt above ends with “verify against a primary source.”
The other thing to know: LLMs hallucinate confidently. A hallucination is a statement that is fluent and authoritative and false. There is no way to eliminate this. There is only mitigation: tight prompts, primary-source verification, and a clear rule that the model advises but never decides.
The Regulatory and Ethical Line in 2026
If you ignore everything else in this article, remember these five things.
1. FINRA Rule 3110 (Supervision) applies to AI-generated content the same way it applies to human-generated content. FINRA Notice 24-09 (June 27, 2024) makes this explicit: “FINRA’s rules - which are intended to be technology neutral - and the securities laws more generally, continue to apply when member firms use Gen AI or similar technologies in the course of their businesses.” If you’re a registered representative, your firm’s supervisory system must cover your use of ChatGPT.
2. FINRA Rule 2210 (Communications with the Public) covers AI-created marketing. The May 10, 2024 update to FINRA’s Advertising Regulation FAQ explicitly addresses AI-created communications and chatbot supervision. If you share AI-generated research or analysis publicly, treat it as a communication and apply the same standards.
3. The SEC, NASAA, and FINRA joint “Artificial Intelligence (AI) and Investment Fraud” alert (January 25, 2024) is the canonical investor warning. It tells retail investors: “Be cautious about using AI-generated information to make investment decisions or to attempt to predict changes in the stock market’s direction or in the price of a security.” It also warns of deepfake audio and video used in “grandchild in distress” scams. Read it before you paste a transcript into anything.
4. OpenAI’s Usage Policies (effective October 29, 2025) prohibit the use of ChatGPT to “provide tailored advice that requires a license” without a licensed professional and to “automate high-stakes decisions in sensitive areas” - with “financial activities and credit” explicitly named - without human review. This is the vendor telling you, in writing, that the model is not a substitute for advice and not a substitute for human review.
5. In the UK, the FCA’s “AI in financial services” hub (live in 2026) tracks the FCA’s “AI Lab,” “AI Sprint” outcomes (which gathered 400 industry experts), and the Mills Review on the long-term impact of AI on retail financial services. The FCA reports 75% of UK financial firms have adopted some form of AI. The regulator’s posture is “safe and responsible adoption.” For retail traders in the UK, the FCA’s existing high-risk investments rules, Consumer Duty, and financial promotions regime (S21 FSMA) apply to AI-generated material the same way they apply to human-generated material.
The punchline: in 2026 the regulatory stance is “technology-neutral.” You do not get a pass because a chatbot wrote it. You get the same standards, the same supervision, and the same liability.
Use Cases in Depth
The 10 prompts cover the core retail workflow. Here is how they map to the eight use cases the brief asked for.
Trade journaling
Prompts 1, 6, and 7 together form the journaling cycle. Prompt 1 (pre-trade) is the entry. Prompt 6 (post-trade) is the exit. Prompt 7 (emotional audit) is the meta-layer that catches patterns across entries and exits. Most retail traders skip one of these three and wonder why they can’t improve. The model will not fix that for you, but it will force the structure.
Post-trade review
Prompt 6 is the workhorse. The trick is to keep the prompt’s output machine-readable (process tag, pattern tag, R-multiple) so you can run counts and statistics across hundreds of trades. Without structure, journal entries are just stories.
Earnings-call summary
Prompt 4 is built for this. The “do not infer numbers” line is the most important clause. The hallucination risk for specific dollar figures in earnings transcripts is the highest I’ve seen in any use case.
Sector screening checklist
Prompt 5 is the screen builder. Run the output through a real screener (Finviz, TradingView, your broker’s). The model will help you design the filter; it should not be the filter.
Risk sizing
Prompt 2 is the calculator. Pair it with a spreadsheet you control. Verify the math on a sample before you trust it with real money.
Scenario planning
Prompt 8 is the stress-test. Re-run quarterly. The base case is always wrong; the question is by how much and in which direction.
Options strategy comparison
Prompt 9 is the side-by-side. Use a real options chain (your broker, the OCC, or Cboe) as the source of strikes, premiums, and Greeks. The prompt structures the comparison; the chain supplies the numbers.
Tax-loss harvesting planning
Prompt 10 is the organizer. The decision is yours and your CPA’s. The model can rank losses, suggest non-substantially-identical replacements, and flag wash-sale windows - but it cannot give you tax advice and OpenAI’s policies forbid it from trying.
Prompt Engineering Patterns Worth Keeping
A handful of meta-prompts make every prompt above work better.
Constrain the role. “You are a trade journal reviewer, not an investment advisor.” This single line cuts the most common failure: the model offering unsolicited buy/sell calls.
Constrain the output shape. Asking for a table, a checklist, or a numbered list reduces hedging and filler.
Forbid inference. “If a figure is unclear, write ‘unclear in source’ and stop.” This catches hallucination.
Cite the source. “Cite the page number or section” is a useful guardrail for long documents.
Flag for verification. “List three things I should verify against a primary source” turns the model into a checklist engine instead of an oracle.
Require the uncertainty. “List three things I am uncertain about” beats “are you sure?” every time, because it forces the model to enumerate rather than to dismiss.
These patterns are universal. They work in ChatGPT, Claude, Gemini, and open-source models. They are not magic. They are discipline.
Common Failure Modes
Even with great prompts, traders break the system in predictable ways.
Overtrust. The biggest risk is treating the model’s structure as evidence. A clean checklist is not a winning trade. The prompt makes you think clearly; the trade still has to work.
Hallucination. Numbers, quotes, and ticker-specific news are the highest-risk categories. Always verify against SEC EDGAR, the company’s investor relations site, or a primary press release.
Privacy. Don’t paste non-public information, account numbers, or anything covered by an NDA. FINRA’s prompt-injection guidance (March 6, 2026) makes the risks explicit: a malicious instruction hidden in a document can manipulate the model’s behavior in ways you can’t see.
Confirmation bias. Prompts don’t fix your priors. They expose them. The bull/bear prompt only works if you let it.
Over-automation. OpenAI’s own policies forbid “automation of high-stakes decisions” in finance without human review. Don’t build a system where the chatbot calls the trades.
Shelfware. Prompts you don’t use don’t help. Pick three, run them for a month, then add more.
The Limits of What AI Can Tell You About Markets
Three honest caveats.
Markets are not summarizable. The LLM is great at summarizing a 200-page 10-K into five bullet points. It is not great at telling you whether the stock is going to go up. Nobody is. Anyone who tells you otherwise is selling something.
Most alpha decays. If a public chatbot can articulate a strategy in plain English, that strategy is probably crowded. The 10 prompts here are deliberately on the process side for that reason.
The right answer to most prompts is “I don’t know.” If a prompt consistently produces confident answers about market direction, the prompt is too loose. Tighten the constraints until the model says “unclear in source” more often. That’s the signal it’s behaving like a good analyst.
FAQ
Is using ChatGPT for trade ideas legal? Yes, with two caveats. First, the output is your responsibility, not OpenAI’s. Second, if you are a registered representative, FINRA Rule 2210 and your firm’s supervisory system apply to anything you publish. The SEC, FINRA, and NASAA jointly warned in their January 25, 2024 investor alert that retail investors should “be cautious about using AI-generated information to make investment decisions or to attempt to predict changes in the stock market’s direction or in the price of a security.”
Can ChatGPT give me personalized financial advice? No - and OpenAI’s Usage Policies (effective October 29, 2025) explicitly prohibit it: “We don’t allow the provision of tailored advice that requires a license, such as legal or medical advice, without appropriate involvement by a licensed professional.” If you want personalized advice, talk to a registered investment professional and verify them via FINRA BrokerCheck or the SEC’s Investment Adviser Public Disclosure.
Will these prompts make me money? They will not. They will reduce process errors. Whether your strategy has an edge is a separate question that no prompt can answer. Most retail strategies don’t.
What’s the difference between ChatGPT and a Bloomberg terminal? A terminal gives you real-time market data, financial statements, and analytics. ChatGPT gives you a language interface that can summarize, structure, and reason over text. The two are complementary, not substitutes. If you need a price, a chart, or a real-time quote, use your broker or terminal - not a chatbot.
What about deepfakes and AI fraud? Real and growing. The January 25, 2024 SEC/FINRA/NASAA alert details scams involving AI-generated audio of family members, fake celebrity endorsements, and synthetic news events. FINRA’s June 2026 investor insight on “Deepfakes and Vishing” reinforces the same message. If a “relative in distress” calls asking for a wire transfer, call the relative on a number you already have.
Should I pay for ChatGPT Plus or use the free tier? For these prompts, either works. The Plus tier gives priority access during peak hours and access to the most recent models, which marginally improves output quality. It does not change the hallucination risk.
What about Claude, Gemini, or open-source models? The prompts transfer with minor rewording. Claude tends to be more cautious on financial advice. Gemini has real-time data hooks. Open-source models (Llama, Mistral, Qwen) can be self-hosted if privacy is a concern, though they require technical setup and are less capable on finance-specific tasks out of the box.
Can I use ChatGPT to backtest a strategy? Not directly. ChatGPT can’t run a backtest. It can help you design a backtest, document its assumptions, and critique its results. The execution still belongs in a real system (Python, your broker’s backtester, a platform like QuantConnect).
Sources
Primary regulators and standards-setters first, then major financial press, then research and academic references. All URLs verified accessible as of mid-2026.
- FINRA, “Regulatory Notice 24-09: FINRA Reminds Members of Regulatory Obligations When Using Generative Artificial Intelligence and Large Language Models,” June 27, 2024. https://www.finra.org/rules-guidance/notices/24-09
- FINRA, “Artificial Intelligence (AI)” Key Topic page, updated 2026 (includes March 6, 2026 prompt-injection guidance and January 27, 2026 AI-agents blog). https://www.finra.org/rules-guidance/key-topics/artificial-intelligence
- FINRA, “Understanding Generative AI and Prompt Injection Fundamentals,” March 6, 2026. https://www.finra.org/rules-guidance/key-topics/artificial-intelligence/understanding-generative-ai-and-prompt-injection-fundamentals
- FINRA, “How FINRA Member Firms Use GenAI,” July 14, 2025. https://www.finra.org/sites/default/files/2025-07/How-Member-Firms-Use-GenAI.pdf
- FINRA, “2026 FINRA Annual Regulatory Oversight Report - GenAI section,” December 9, 2025. https://www.finra.org/rules-guidance/guidance/reports/2026-finra-annual-regulatory-oversight-report/gen-ai
- FINRA, SEC Office of Investor Education and Advocacy, NASAA, “Artificial Intelligence (AI) and Investment Fraud,” January 25, 2024. https://www.finra.org/investors/insights/artificial-intelligence-and-investment-fraud
- FINRA, “Deepfakes and Vishing: What You Need to Know to Stay Protected,” June 2026 investor insight. https://www.finra.org/investors/insights/deepfakes-vishing
- FINRA, “Frequently Asked Questions About Advertising Regulation” (covers AI-created communications and chatbot supervision, updated May 10, 2024). https://www.finra.org/rules-guidance/guidance/faqs/advertising-regulation
- FINRA Foundation, “The Machines Are Coming (with Personal Finance Information). Do We Trust Them?” 2024 FINRA Foundation Study. https://www.finrafoundation.org/sites/finrafoundation/files/the-machines-are-coming.pdf
- SEC, “Investor.gov - Artificial Intelligence and Investing” educational resources. https://www.investor.gov/
- OpenAI, “Usage Policies,” effective October 29, 2025. https://openai.com/policies/usage-policies/
- FCA, “AI: artificial intelligence in financial services” hub, 2026 (includes AI Lab, AI Sprint summary, Mills Review). https://www.fca.org.uk/firms/ai-financial-services
- FCA, “AI Sprint Summary” (400 industry experts and stakeholders). https://www.fca.org.uk/publications/corporate-documents/ai-sprint-summary
- FCA, “FS25-5: AI Live Testing - Feedback Statement,” May 2025 (referenced from FCA AI hub). https://www.fca.org.uk/publications/feedback-statements/fs25-5-ai-live-testing
- ESMA, “Digital Finance and Innovation” workstream (covering DORA, MiCA, and DLT Pilot Regime relevant to AI/algorithmic trading infrastructure). https://www.esma.europa.eu/esmas-activities/digital-finance-and-innovation
- Wu, Irsoy, Lu, Dabravolski, Dredze, Gehrmann, Kambadur, Rosenberg, Mann (Bloomberg), “BloombergGPT: A Large Language Model for Finance,” arXiv:2303.17564, v3 December 21, 2023. https://arxiv.org/abs/2303.17564
- Cboe Exchange, “U.S. Options Daily Market Statistics,” 2026 (used as a live reference for put/call ratios, volume, and open interest; values change daily - check the current page before relying on any specific number). https://www.cboe.com/us/options/market_statistics/daily/
- Cboe, “The Options Institute” (free educational resources for options strategy and Greeks). https://www.cboe.com/optionsinstitute/
- Internal Revenue Service, “Publication 550: Investment Income and Expenses” (covers wash-sale rule under §1091). https://www.irs.gov/forms-pubs/about-publication-550
- NIST, “AI Risk Management Framework (AI RMF 1.0),” January 2023 (referenced by FINRA Notice 24-09 as a useful industry resource). https://nvlpubs.nist.gov/nistpubs/ai/NIST.AI.100-1.pdf
- FINRA, “Investor Insights: Protecting Your Investment Accounts From GenAI Fraud” (account-takeover and new-account fraud). https://www.finra.org/investors/insights/gen-ai-fraud-new-accounts-and-takeovers
- FINRA Rule 3110 (Supervision) and FINRA Rule 2210 (Communications with the Public), FINRA Manual. https://www.finra.org/finramanual
- SEC Rule 206(4)-1 (Marketing Rule), Investment Advisers Act of 1940. https://www.sec.gov/
- FCA, “Our Strategy 2025 to 2030” (including AI as a strategic priority). https://www.fca.org.uk/publication/corporate/our-strategy-2025-30.pdf