How AI Traders Actually Make Money (It’s Not One Chat)

📖 6 min read

The “Give Me a Stock Pick” Trader Is Gambling

Every day, thousands of people open ChatGPT and type “What stocks should I buy?” or “Is Bitcoin going up?” They get a balanced, hedged non-answer. Then they either ignore it or blindly follow the one ticker that was mentioned.

Meanwhile, the traders actually using AI to generate edge are running structured pipelines – not asking for picks. They’re using layered prompts to do in 20 minutes what used to take a research team 3 hours: scan for signals, build a thesis, stress-test it, size the position, and document the trade.

One prompt gets you a horoscope. Five layers get you a trading process.

Why “AI Stock Picks” Are Worthless

When you ask for a recommendation, the AI has no edge. It’s summarizing what’s already priced in. It has no real-time data. It doesn’t know your risk tolerance, portfolio, or time horizon. It’s giving you consensus, which is by definition not alpha.

But AI is exceptionally good at:

• Processing information faster than you can read it
• Finding patterns across multiple data points
• Stress-testing your thesis against contrary evidence
• Removing emotional bias from position sizing
• Maintaining consistency in your process

The edge isn’t in the answer. It’s in the process.

The 5-Layer Trading Pipeline

Layer 1: Signal Scan

Turn noise into ranked signals worth investigating.

The prompt:

Here's what I'm seeing in [market/sector] today:
[paste news headlines, price moves, on-chain data, earnings, or whatever raw inputs you track]

Scan for:
1. Unusual divergences (price vs. volume, sentiment vs. action, sector rotation signals)
2. Catalysts within 1-4 weeks (earnings, unlocks, regulatory deadlines, macro events)
3. Positioning extremes (crowded trades, max pain levels, funding rates)
4. Information asymmetry opportunities (something the market hasn't priced yet)

Rank the top 5 signals by:
- Confidence (how reliable is this signal historically?)
- Timing (how urgent - hours, days, weeks?)
- Magnitude (if correct, how big is the move?)

Do NOT recommend trades. Just surface what's worth investigating further.

Why this works: You’re not asking for alpha – you’re asking for triage. The AI helps you focus your limited attention on the 5 things that matter most out of 50 inputs. The explicit “do not recommend trades” keeps it analytical, not predictive.

Layer 2: Thesis Construction

Pick the top signal and build a proper investment thesis around it.

The prompt:

I'm investigating this signal:
[paste the signal you chose from Layer 1]

Build me a structured thesis:

1. WHAT: What exactly is the opportunity? (Be specific - direction, instrument, timeframe)
2. WHY NOW: What's the catalyst or timing element that creates urgency?
3. HOW IT PLAYS OUT: What's the sequence of events if this thesis is correct?
4. HISTORICAL ANALOG: When has something similar happened before? What was the outcome?
5. CONSENSUS VIEW: What does the market currently believe? Why might they be wrong?
6. KILL CONDITION: What single piece of evidence would invalidate this thesis entirely?

Be rigorous. If the thesis is weak, say so. I'd rather kill a bad idea now than lose money proving it wrong.

Why this works: The “kill condition” is the most important part. It turns a vague hunch into a testable hypothesis. If you can’t define what kills the trade, you don’t have a trade – you have a gamble.

Layer 3: Risk Model

Before touching position sizing, understand what can go wrong.

The prompt:

Here's my thesis:
[paste Layer 2 output]

Now destroy it. Play devil's advocate:

1. What are the 3 most likely ways this trade loses money?
2. What's the worst-case scenario (tail risk)? How bad can it get?
3. What correlated risks exist? (If this goes wrong, what else in my portfolio gets hit?)
4. What's the market's implied probability vs. my estimated probability?
5. What's the liquidity profile? Can I exit cleanly if wrong?

Then give me:
- Maximum acceptable loss (as % of position)
- Recommended stop-loss level with reasoning
- Time stop: when should I exit regardless of price if the thesis hasn't played out?

Assume I have [X] total portfolio value and am willing to risk [Y]% on this trade.

Why this works: Forcing the AI to attack your thesis before you enter removes confirmation bias. The time stop is crucial – most losing trades die from indecision, not from stop-losses. If the catalyst hasn’t fired in your window, the thesis is dead.

Layer 4: Position Sizing

Math, not feelings.

The prompt:

Based on this risk model:
[paste Layer 3 output]

Calculate my position size using these parameters:
- Portfolio value: [X]
- Maximum risk per trade: [Y]%
- Entry price: [current or limit price]
- Stop-loss level: [from Layer 3]
- Current portfolio correlation to this trade: [low/medium/high]

Give me:
1. Exact position size (units and dollar value)
2. Risk/reward ratio at my target
3. Kelly criterion suggestion (and why I should probably use half-Kelly)
4. Scaling plan: should I enter all at once or scale in? At what levels?

Show your math. I want to verify the numbers.

Why this works: “Show your math” catches calculation errors. The AI is good at Kelly criterion and correlation adjustments – things most retail traders either skip or calculate wrong. Half-Kelly suggestion builds in safety margin.

Layer 5: Trade Journal Entry

Document before you enter, not after.

The prompt:

Create a trade journal entry for this position:

Thesis: [1-line summary from Layer 2]
Direction: [long/short]
Entry: [price and size from Layer 4]
Stop: [from Layer 3]
Target: [from thesis]
Kill condition: [from Layer 2]
Time stop: [from Layer 3]
Risk/reward: [from Layer 4]

Pre-mortem: Write 2 paragraphs in future tense describing what it looks like when this trade FAILS. What went wrong? What did I miss? What was I feeling that biased me?

Conviction rating: Based on everything above, rate this trade 1-10 on conviction and explain why. If below 6, flag it as "size down or skip."

Why this works: The pre-mortem is the secret weapon. Writing the failure scenario before entering makes you emotionally prepared for it. When the trade goes against you, you’ve already processed it. You execute your plan instead of panic-closing.

The Difference in Practice

One-prompt trader:

• “Is ETH going up?” – gets a vague answer
• Buys based on feeling
• No stop-loss, no thesis, no exit plan
• Holds through -30% because “it’ll come back”
• Sells at the bottom out of fear

Layer Method trader:

• Scans 30 signals, picks the strongest
• Builds a thesis with a specific kill condition
• Knows the exact dollar amount at risk before entry
• Has a pre-written plan for both winning and losing scenarios
• Executes mechanically because the process already handled the emotions

Important Caveat

AI doesn’t have real-time data unless you plug it into feeds. The pipeline works best when YOU bring the raw data (price action, news, on-chain metrics) and the AI helps you process and structure it. Don’t ask the AI what the price is. Tell it what the price is and ask what it means.

Copy This Workflow